Detailed Description
For the purpose of making the objects, technical solutions and advantages of the embodiments of the present invention more apparent, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings, and it is apparent that the described embodiments are some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
The embodiment of the invention provides a method and a device for predicting faults of an energy storage liquid cooling system, which are used for solving the problems.
For the sake of understanding the present embodiment, first, the method for predicting the failure of the energy storage liquid cooling system disclosed in the present embodiment of the present invention is described in detail, and fig. 1 shows a flowchart corresponding to the embodiment of the present invention, and referring to fig. 1, the method includes the following steps:
Step S102, operation monitoring parameters corresponding to a preset energy storage liquid cooling system are obtained.
The operation monitoring parameters comprise monitoring parameters corresponding to preset key equipment of the energy storage liquid cooling system. The system mainly comprises a plurality of types of sensors which are arranged on key components, wherein the sensors mainly comprise a temperature sensor, a pressure sensor, a flow sensor, a voltage/current sensor and the like, the sensors are collected in real time through the multi-type sensors which are arranged on the key components, the key components comprise a cooling liquid pipeline, a battery module, a circulating pump and the like, data are uploaded to a local server or a cloud database through an industrial bus (such as a CAN bus or a Modbus protocol) or a wireless transmission module (such as LoRa or 5G), sampling frequency is dynamically adjusted according to the running state of equipment, the sampling frequency is 1Hz under normal working conditions, the sampling frequency is increased to 10Hz under abnormal working conditions to ensure fault characteristic capture, and original data are stored in a time sequence form and comprise fields such as equipment ID, time stamp, sensor type, numerical value and the like to form an initial monitoring data set.
Furthermore, the primary filtering and outlier rejection can be performed on the original signal in the data acquisition stage, for example, instantaneous interference noise is eliminated through sliding average filtering, and obvious outlier data is filtered through threshold detection, so that interference of subsequent processing is reduced. Furthermore, the data storage can adopt a time-division filing strategy, and the files are divided according to the unit of hours/day, so that the subsequent segmentation processing and the historical data backtracking are facilitated.
Step S104, based on the local noise level and time scale change of the operation monitoring parameters, data division is carried out on the operation monitoring parameters to obtain segmented data corresponding to a plurality of time periods respectively.
Step S106, determining the denoising degree corresponding to the segment data according to the noise intensity corresponding to the segment data.
Step S108, denoising the segmented data based on denoising strength, and determining operation monitoring parameters based on parameters to be measured corresponding to each time scale.
Wherein the system exhibits different dynamic behavior during different phases of operation (start-up, steady operation, load change, failure precursor). The fault types also have time scale differences, such as abrupt faults (e.g. pipeline breakage), short time scale and high noise, and slow faults (e.g. cooling efficiency reduction), long time scale and weak characteristics. In addition, data acquired by the sensor in different time periods is influenced by environmental interference, equipment states, sampling frequency and other factors, so that noise level fluctuation is large. Moreover, in normal operating conditions, the early failure signal is very weak, often submerged in high frequency noise or background fluctuations.
In order to solve the defect of the prior art that the energy is extracted from the fault characteristics, the embodiment of the invention segments the data based on the local noise difference of the data, so that the noise intensity, trend change and time scale in the same segment are more consistent. Based on the self-adaptive data partitioning mechanism, the subsequent processing can be respectively processed according to the data blocks with similar homogeneity, and the denoising and feature extraction accuracy is improved. Further, the denoising intensity is automatically adjusted according to the noise intensity corresponding to each piece of data, and personalized denoising operation is performed according to the specific denoising intensity, so that noise can be effectively restrained, loss of useful fault characteristics can be avoided, and accurate extraction of the characteristics under each time scale is ensured. In conclusion, the finally output data are clean parameters to be measured under multiple time scales, high-quality input data can be provided for the neural network, and the interpretability of the characteristics and the generalization capability of the model are improved.
Step S110, inputting the parameters to be tested into a pre-trained neural network model, and carrying out classification prediction on the parameters to be tested through the neural network model to determine a fault type prediction result corresponding to the parameters to be tested.
And step S112, determining a fault state corresponding to the energy storage liquid cooling system according to the fault type prediction result.
Further, the processed data is input into a preset neural network model (such as LSTM, GRU, transformer, CNN-LSTM hybrid model and the like) for reasoning, the model can output a prediction type, and the prediction type can be different fault grades or fault types. Such as the type of failure (e.g., coolant leakage, pump block), severity level, and location of occurrence. The prediction category may be determined based on a sample label corresponding to the training sample set. Wherein the sensor data may be associated with the fault event by time stamp alignment based on the device log and the fault maintenance record to determine the data tag. For example, when the system records a fault alarm (such as over-temperature protection trigger) or a manual maintenance report, monitoring data is traced back, and a data label of a corresponding time period is marked according to a fault occurrence time window (such as 30 minutes before the fault to 5 minutes after the fault). In one embodiment, the labeled data can be divided into a training set, a verification set and a test set according to the ratio of 7:2:1, wherein the training set is used for model parameter learning, the verification set is used for super-parameter tuning, the test set is used for final performance evaluation, and the generalization capability of the model is ensured.
Further, the model can also output fault class probability distribution, and when the highest probability exceeds a set threshold (such as 0.9), a hierarchical alarm is triggered, and the fault position and the suggested treatment measure are synchronously pushed to the operation and maintenance platform. The historical prediction result and the equipment state data can be stored in a knowledge base, so that the knowledge base can be used for updating the model periodic increment.
In summary, according to the energy storage liquid cooling system fault prediction method provided by the embodiment of the invention, the self-adaptive segmentation processing of the original data is realized by dynamically analyzing the local noise level and time scale change of the monitoring data. Features in different failure modes, especially failures with significant time scale differences (e.g., abrupt and slow-onset failures), can be better captured. In addition, the embodiment of the invention dynamically adjusts the denoising strategy according to the characteristics of each piece of segmented data, so that not only can the noise be effectively removed, but also the useful information can be reserved to the maximum extent, and the data quality of the subsequent model training can be improved. The embodiment of the invention ensures that weak but key fault characteristics can be accurately extracted even in a high noise environment by accurately controlling the denoising process, and is important for improving the sensitivity and the specificity of a fault diagnosis system.
Further, on the basis of the foregoing embodiment, the embodiment of the present invention further provides another method for predicting a fault of an energy storage liquid cooling system, and fig. 2 shows a flowchart corresponding to the embodiment of the present invention, and referring to fig. 2, the method includes the following steps:
Step S202, operation monitoring parameters corresponding to a preset energy storage liquid cooling system are obtained.
Step S204, based on the local noise level and time scale change of the operation monitoring parameters, data division is carried out on the operation monitoring parameters to obtain segmented data corresponding to a plurality of time scales respectively.
The operational monitoring data tends to exhibit non-stationary characteristics over time. For example, the plant may behave differently during different phases of operation (e.g., start-up, steady operation, shut-down). In addition, changes in external environmental conditions (e.g., temperature, humidity, etc.) can also lead to data non-stationarity. The noise levels for different time periods may vary significantly. Some periods may have higher background noise while other periods are relatively clean. Equipment failure or other sudden events can lead to sudden increases in local noise, affecting the accuracy of the global analysis. Aiming at the data characteristic, the embodiment of the invention provides a self-adaptive wavelet denoising and dynamic time segmentation strategy, and the data is divided by using a self-adaptive segmentation method through calculating local noise characteristics. And the dynamic adjustment of the denoising threshold value can be carried out based on the time scale change (such as short-term fluctuation and long-term trend) of the data, so that the method is flexibly suitable for the noise intensity and time scale difference under different fault modes, and the information loss possibly caused by the denoising of the fixed threshold value is avoided, thereby improving the retention and diagnosis accuracy of fault characteristics.
In specific implementation, the operation monitoring parameters are firstly subjected to segmentation processing through the following steps, so that the data in each segment not only have similar noise level, but also have similar time scale, and the characteristic distribution of different fault modes is adapted. The method comprises the steps of determining a data noise difference in a current time period according to a data noise level corresponding to each time point of a preset time period of operation monitoring parameters, determining a time segmentation point corresponding to the operation monitoring parameters based on the data noise difference, and carrying out data division on the operation monitoring parameters based on the time segmentation point and a preset time segmentation length penalty term to obtain segmented data corresponding to a plurality of time scales respectively.
In one embodiment, the time periods may be dynamically divided by K-means clustering, and in combination with the above steps, the clustering objective function may be expressed as:
in the formula,Representing a minimized objective function; Is a point in time within time period Tk; Is the original firstMonitoring data by the energy storage liquid cooling system at any time; For controlling the segment length penalty, e.g. set to 0.3. The embodiment of the invention can adopt a clustering algorithm to cluster the data to obtain the cluster to which each time point belongs. Based on the clustering result, the positions of the adjacent time points belonging to different clusters are potential time segment points. The length penalty term can be introduced in the clustering process to prevent the segmentation from being too frequent, and further, the data segmentation is performed based on the time segmentation point and the length constraint of the length penalty term, so that a plurality of segment data are obtained. The segmented data are obtained by segmentation according to noise differences corresponding to time points, and the time scale corresponding to each segmented data is the same, such as short-term time scale segmented data corresponding to a transient fault mode, medium-term time scale segmented data corresponding to a sudden fault, and long-term time scale segmented data corresponding to long-term trend and structural change. Compared with the traditional fixed threshold denoising process adopted under the traditional sliding window segmentation, the embodiment of the invention is based on adaptive segmentation and wavelet denoising process of non-fixed threshold, and can adapt to different fault modes, so as to better capture details and improve the retention and diagnosis accuracy of fault characteristics.
It should be noted that, the time scale difference of the failure modes of the energy storage liquid cooling system is obvious, and the embodiment of the invention can be implemented by the methodThe values with different sizes are set, so that unstable noise estimation caused by too short segments or covering time-varying characteristics caused by too long segments are prevented, and the non-stationary time sequence characteristics are further adapted.
Step S206, determining the denoising degree corresponding to the segment data according to the noise intensity corresponding to the segment data.
The noise intensity in the monitoring data of the energy storage liquid cooling system is different according to different fault modes, such as different noise distribution of temperature abrupt change and pressure gradual change faults. According to the embodiment of the invention, the denoising strength is dynamically adjusted according to the noise strength of the local data segment (namely the segmented data). The calculation mode is expressed as follows:
The method and the device combine the extremely poor and variance of the data in the segment, reduce the suppression coefficient when the noise intensity is high (variance is large), retain more potential fault characteristics, enhance the suppression when the noise intensity is low (variance is small), avoid overcomplification, and further solve the problem of overcompression of weak fault characteristics by fixed-threshold wavelet denoising.
In the above-mentioned formula(s),For the dynamic noise suppression coefficient, Tk is the kth time period; Represent the firstMonitoring a maximum value of the data over a period of time; representing a minimum value of the monitored data during the kth time period; The data volatility of the k-th time period is characterized by the variance of the data in the k-th time period.
Step S208, denoising the segmented data based on denoising strength, and determining the operation monitoring parameters based on the parameters to be measured corresponding to each time scale.
The embodiment of the invention carries out self-adaptive segmentation on the data according to the noise difference of the data in advance, so that each segment of segmented data has similar noise level and time scale respectively. Further, denoising each piece of sectional data according to denoising forces corresponding to different pieces of sectional data, and further obtaining parameters to be measured after denoising operation monitoring parameters. The data denoising method and device provided by the embodiment of the invention not only can distinguish different fault modes, but also can effectively reduce the loss of effective information corresponding to each fault mode. Specifically, the step of denoising the segmented data based on the denoising strength refers to the following formula:
in the formula,Monitoring data for the denoised energy storage liquid cooling system at the t time; Is the original firstMonitoring data by the energy storage liquid cooling system at any time. K is the number of segments, and the time scale difference of different fault modes is adapted by dynamically adjusting the number of segments. Wherein, theIn the embodiment of the invention, the characteristics of the local signals of the segmented data are extracted by using the wavelet odd function to denoise the data. The method can adapt to non-stationary signals. Wherein the function is asIs used as a center of the water tank,For the window width, a local signal is extracted. t is a time variable; Is the center point of the kth segment; the width of the k-th section window; the k-th segment data mean value is used for measuring the center position of each segment of data.
To verify the effectiveness of the present technique, embodiments of the present invention also employ a dual axis timing diagram versus performance of different time-slicing strategies. Taking monitoring data of an energy storage liquid cooling system as temperature data as an example, fig. 3 shows a schematic diagram of comparison results corresponding to adaptive segmentation according to an embodiment of the present invention. The method comprises the steps of simulating state data of abrupt change and gradual change of a temperature signal, further verifying effectiveness of a segmentation strategy, wherein signal amplitude of an ordinate is absolute temperature, unit is DEG C, and represents amplification of temperature compared with reference temperature, and the reference temperature is ambient temperature. The main graph shows a composite fault signal comprising high-frequency oscillation and a gradual change trend, pulse noise is superimposed at the same time, a background gray curve is noisy observed data, a black dotted line is an ideal real signal, the vertical parting line is used for differentiating the segmentation results of different methods, the test result shows that the green dotted line represents the self-adaptive segmentation of the technology, the self-adaptive segmentation of the technology is accurately divided near a fault mode conversion point (signal mutation starting position), a red solid line is an equidistant fixed segmentation, obvious offset occurs at the gradual change fault starting stage, an orange dotted line is a sliding window segmentation, a large number of redundant segments can be generated although the mutation is captured, the gradual change fault starting region is mainly shown in the local amplification graph, the segmentation line of the technology can be accurately aligned with the signal trend conversion point, and the comparison method has hysteresis or lead phenomena, so that the dynamic clustering strategy can be self-adaptively matched with the time-varying characteristic of the fault mode. It should be noted that the monitoring data of the energy storage liquid cooling system also includes other various parameters, such as pressure data, flow data, liquid level data, etc. Different types of monitoring data, the signal amplitude units in fig. 3 are different. When the monitoring data is pressure data, the unit can be Pa (pascal), kPa (kilopascal) or MPa (megapascal), and when the monitoring data is flow data, the unit can be L/min (liter/min) or m3/h (cubic meter/hour). Further, the remaining types of monitoring data may be represented by corresponding international standard units, which are not described herein.
Step S210, inputting parameters to be tested into a pre-trained neural network model, carrying out feature capture by utilizing a plurality of convolution cores of different scales on the parameters to be tested through a time sequence feature dynamic fusion layer of the neural network model, and determining local target features corresponding to the parameters to be tested.
The embodiment of the invention predicts data by classifying through a neural network model, wherein the neural network adopted by the embodiment of the invention adopts a layered time sequence feature extraction architecture and comprises an input layer, a time sequence feature dynamic fusion layer, an adaptive regression unit and a multi-scale residual error connection module. In one embodiment, an input layer receives time sequence data after denoising segmentation, the dimension is [ time step size x sensor channel number ], a time sequence characteristic dynamic fusion layer consists of parallel dynamic convolution kernels and attention mechanisms, local characteristics are extracted through 3 groups of convolution kernels with different scales and are fused with time attention weighting, an adaptive regression unit comprises 4 expert sub-networks, each sub-network consists of 3 layers of full-connection layers, an activation path is dynamically selected through a gating mechanism, and a multi-scale residual error connection module is bridged between hidden layers and fuses pooling characteristics of the first 3 layers to enhance gradient propagation. The output layer generates fault class probability by adopting a Softmax function, and takes the class with the largest fault class probability as the final classification class. It should be noted that the neural network architecture can be elastically expanded, for example, increasing the number of layers of the fully connected neural network, increasing the common convolution layer, increasing the pooling operation, etc., which all belong to the protection scope of the present technology.
When the method is specifically implemented, firstly, after the input layer receives data, the time sequence feature dynamic fusion layer determines local target features corresponding to parameters to be tested. Because the monitoring data of the energy storage liquid cooling system has multivariable coupling and non-stable time sequence characteristics, the conventional technology directly stacks the time sequence data or extracts statistical characteristics, and cannot capture the time-varying coupling relation of fault characteristics. According to the invention, the time sequence feature extraction layer fused by the dynamic convolution kernel and the attention mechanism is constructed, and the time sequence feature extraction layer combined by the dynamic convolution kernel and the attention mechanism is constructed, so that the time-varying coupling relation of fault features is effectively captured, and in the process, the dynamic convolution kernel weights the features at each moment through the time attention mechanism, so that the characterization capability of the fault features is improved.
When in specific implementation, the method comprises the following steps:
1) And carrying out nonlinear mapping on the characteristic parameters of the parameters to be detected at each moment through a preset multi-layer perceptron, and determining query vectors respectively corresponding to the parameters to be detected at each moment.
2) And determining the convolution kernel attention weight corresponding to the characteristic parameters of the parameters to be measured at each moment according to the attention mechanisms of the query vector and the key vector of the preset dynamic convolution kernel.
It should be noted that there is a time-varying coupling relationship between the energy storage liquid cooling system and multiple variables (such as temperature, pressure and flow), such as temperature fluctuation caused by abrupt change of the cooling liquid flow. The embodiment of the invention is characterized by inquiring the vector Q (t) (by the current momentGenerated through nonlinear mapping of multi-layer perceptron) and key vectorsThe attention mechanism of the (2) dynamically adjusts the attention weight of the convolution check features at different time points so as to capture the transient coupling relation of the fault features. Wherein, defineFor the query vector at the t-th moment, the obtaining mode is expressed as follows:; Is a multi-layer perceptron, such as a 3-layer perceptron network.
Further, defineFor the ith dynamic convolution kernel at the t moment, dynamically generating convolution kernel weight according to the characteristic of the current moment, and expressing the calculation mode as follows。
Wherein, theKi is the ith key vector, represents the key vector of the dynamic convolution kernel and is a training parameter; Transpose of Ki; For the dimension of the query vector and the key vector, the default is set to 64, and the calculation complexity of the attention weight is controlled; And (3) representing the adjustment parameters of the convolution kernel as training parameters for the ith value vector.
3) According to the convolution kernel weight and the time attention weight corresponding to the characteristic parameter, capturing the time sequence characteristic of the parameter to be detected, and determining the local target characteristic corresponding to the parameter to be detected.
Determining local target features refers to the following formula:
in the formula,And the fusion characteristic at the t time is expressed as a result of dynamic fusion of the time sequence characteristic.For the number of dynamic convolution kernels, we mean that multiple convolution kernels are used in the model to capture different types of timing features, e.g., set to 3; representing a convolution operation.Is the ith dynamic convolution kernel at time t.
Wherein, theAnd (3) for the time attention weight at the t moment, carrying out weighted fusion on the characteristics in the local time window, wherein the calculation mode is expressed as follows:
A local time index within the time sliding window; Is the local time window length; Represent the firstThe convolution kernel is atAttention coefficient of moment; Representing the de-noised firstMonitoring data of time. It should be noted that the fault feature may have local burstiness, such as a flow dip at the moment of stopping the cooling pump, and the embodiment of the present invention extracts and determines the time attention weight based on the local context information of the parameter to be measured by using the time sliding window. By combining the formulas, the embodiment of the invention extracts local context through the sliding window, and combines the Sigmoid function to adaptively allocate weight, so as to strengthen the characteristic contribution of key time points, further enhance the sensitivity to the mutant faults, and simultaneously keep the long-term trend of the slow-change faults; the function is activated for Sigmoid.
Furthermore, the embodiment of the invention also utilizes a time-frequency domain joint analysis method to analyze the resolving power of different methods on the non-stationary signals. Fig. 4 shows an effect schematic diagram corresponding to the dynamic convolution kernel according to the embodiment of the present invention. Referring to fig. 4, experimental focusing verifies the capturing effect of the dynamic convolution kernel and the attention mechanism on the time-varying characteristics through triple verification of time domain waveforms, power spectrum densities and time-frequency spectrograms. In fig. 4, the upper time domain curve shows that the blue signal of the technology eliminates low-frequency fluctuation noise in the gray original signal while retaining the attenuation oscillation characteristics of the original signal, and the lower left frequency spectrogram shows that the technical result keeps clear spectral peaks at 2Hz and 5 Hz. Wherein, the abscissa represents frequency, the ordinate represents Power spectral density (English is called Power SPECTRAL DENSITY, PSD for short), and the unit is decibel per hertz (dB/Hz) and represents decibel Power value in each Hz bandwidth. Whereas the spectral line of the traditional method has the frequency diffusion phenomenon. For the time-frequency diagram on the lower right side, the embodiment of the invention adopts color intensity to represent signal energy distribution, and the map corresponding to the technology completely reserves the short-time pulse characteristics of high-frequency transient components in the time dimension, thereby proving the capability of the dynamic convolution kernel to adjust the weight according to the local characteristics and overcoming the problem of time-frequency resolution loss caused by the fixed convolution kernel in the traditional method.
Step S212, performing adaptive activation path selection on parameters to be detected by using a preset expert selection matrix through an adaptive regression unit of the neural network model, and determining high-dimensional feature representation corresponding to the parameters to be detected.
In the fault diagnosis task of the energy storage liquid cooling system, the time sequence regression characteristics of different fault modes are obviously different, for example, the time sequence regression characteristics of different fault modes may be shown as gradual change and abrupt change. The conventional full connection layer cannot adapt to regression characteristic differences of different fault modes, and is easy to cause the lack of fitting of rare fault modes. The invention builds the adaptive regression unit, can adaptively select proper regression parameters according to different fault states, and can automatically adjust parameters in the regression process under various fault modes by an expert selection mechanism, thereby adapting to different types of fault characteristics and improving the generalization capability of fault diagnosis.
When in specific implementation, the method comprises the following steps:
1) And mapping the parameters to be detected into expert probability distribution by using a preset expert selection matrix.
2) And determining expert units corresponding to the characteristic parameters of the parameters to be measured at each moment respectively based on the expert probability distribution.
The regression characteristics of different fault modes are obviously different, such as the difference between a gradual change condition and a sudden change condition is larger. The embodiment of the invention uses the expert selection matrix E to select the characteristicsMapping to expert probability distribution to automatically switch model to regression parameters corresponding to fault modeAndAnd further, the problem of the lack of fitting of the traditional full-connection layer to rare fault modes is solved, and the diagnosis capability of the model to long-tail categories (low-frequency faults) is improved. Further, the most adapted expert unit is dynamically selected according to the input features. Wherein each element of the expert selection matrix represents a degree of influence or association of a particular feature on a particular failure mode. Wherein the dimension of the expert selection matrix of the embodiment of the invention is,The number of possible failure modes is indicated for the total number of experts. Wherein each row corresponds to a different feature, each column corresponds to a different failure mode, and each element reflects the relevance or contribution of the ith feature to the jth failure mode. These values may be determined in a variety of ways including, but not limited to, expert opinion, experimental data, historical fault data, and the like.
In an embodiment of the present invention, the expert unit may be determined by:
where s is the expert unit index of the expert selection mechanism,The index value corresponding to the maximum probability is represented.As a Softmax function.The probability distribution of expert selection is characterized, wherein each element corresponds to the probability of selecting the s-th expert. E is an expert selection matrix, which characterizes the linear transformation matrix from the feature space to the expert selection, as a training parameter.
3) And carrying out nonlinear transformation on the parameters to be detected based on the regression parameters corresponding to the expert units, and determining the high-dimensional characteristic representation corresponding to the parameters to be detected.
In specific implementations, reference is made to the following formula:
wherein F (t) is the output characteristic of the time sequence characteristic dynamic fusion layer.Is the firstLayer numberThe expert units output, characterize the nonlinear transformation result aiming at the specific fault mode; Is an output feature of layer 1; activating a function for a ReLU; the weight matrix of the s-th expert unit of the first layer is a training parameter; the bias term for the s-th expert unit of the first layer is a training parameter. s is the expert unit index of the expert selection mechanism, determined by the steps described above.
It should be noted that the adaptive regression unit functions to make parametersAndAnd (3) self-adaptive adjustment in the model training process, automatically adjusting a regression process according to different fault modes, and avoiding the under fitting of rare fault modes.
Step S214, based on a preset step convolution operation, the time scale features respectively corresponding to the local target feature and the high-dimensional feature representation are displayed and extracted.
It should be noted that the failure of the energy storage liquid cooling system may involve multiple time scale dynamics, such as a decrease in long-term cooling efficiency and short-term flow fluctuations. The embodiment of the invention also uses preset step length, such asExplicitly extracting short-term (m=1), medium-term (m=2), long-term (m=3) features, thereby enhancing the modeling ability of the model for complex failure modes. In one embodiment, a definition is made ofRepresenting step size asIs a one-dimensional convolution of (a) and (b).
Step S216, based on the contribution degree of the to-be-measured parameter to the time scale feature dynamic fusion layer or the adaptive regression unit, performing adaptive fusion on the time scale feature corresponding to the local target feature and the time scale feature corresponding to the high-dimensional feature representation, and determining the target fusion feature corresponding to the to-be-measured parameter.
Because the monitoring data of the energy storage liquid cooling system has dynamic characteristics of multiple time scales and long-term dependence, the gradient disappearance problem is serious in deep neural network training. Conventional residual connections ignore different time scale feature contribution differences. In order to solve the problem, the invention adopts layering multi-scale residual error and combines the multi-layer characteristics so as to effectively solve the problem of common gradient disappearance in the deep neural network. According to the embodiment of the invention, by using residual terms with different scales, the network can learn the characteristics under different time scales, so that the robustness and generalization capability of the model are improved.
Specifically, residual connection is performed on the time scale features based on residual terms so as to determine corresponding target fusion features. The process is expressed as:
in the formula,Is the firstLayer residual terms enhance cross-layer propagation of gradients.The maximum number of temporal scales for the residual connection, e.g., set to 3, represents the characteristics of the 3 layers before fusion.Is an output feature of the first-m layer.Representing a one-dimensional convolution with a step size of 2m.Residual weights for the mth scale of the first layer.
In specific implementation, the residual weight is dynamically adjusted according to the relevance of the current layer and the historical layer characteristics, and the calculation mode is expressed as follows: And measuring the contribution degree of the mth scale feature to the current layer. Wherein the weighting may be adaptive through a gating mechanism. It should be noted that the features of different time scales in the deep network have different contribution degrees to fault diagnosis, such as long-term dependence is more important to temperature control failure. The self-adaptive distribution of weights can be realized by taking the Sigmoid function as a gating mechanism, so that the dilution of important features by fixed residual connection is avoided, the gradient vanishing problem is further relieved, and meanwhile, the discriminant of the multi-scale fault features is reserved.
Wherein, theFor the weight matrix of the m-th scale, the spliced features are obtainedMapping to scalar weightsAnd controlling the fusion proportion of the features with different scales. Wherein, theFor the average pooling operation of step size 2m, the time dimension is compressed to 1/2m and the long-term dependence features are extracted.The characterization splices the pooled features with the current layer features to provide multi-scale context information.Is an output feature of the first layer.
And step S218, carrying out fault class probability prediction on the target fusion characteristics, and determining a fault class prediction result corresponding to the parameter to be detected.
By combining the steps, the embodiment of the invention generates the fault class probability by adopting a Softmax function through the output layer of the neural network model, and takes the class with the largest fault class probability as the final classification class.
Step S220, determining a fault state corresponding to the energy storage liquid cooling system according to the fault type prediction result.
In summary, the embodiment of the invention provides a self-adaptive wavelet denoising and dynamic time segmentation strategy, and the denoising threshold value is dynamically adjusted by calculating local noise characteristics and time scale changes, and the data is divided by using a self-adaptive segmentation method. Compared with the traditional fixed threshold wavelet denoising and sliding window segmentation, the method can flexibly adapt to noise intensity and time scale differences under different fault modes, avoids information loss possibly caused by fixed threshold denoising, and accordingly improves the retention and diagnosis accuracy of fault characteristics.
Furthermore, in order to cope with the multivariable coupling and nonstationary characteristics of the monitoring data of the energy storage liquid cooling system, the invention adopts a time sequence characteristic dynamic fusion layer, the layer combines a dynamic convolution kernel and a time attention mechanism, the time-varying fault characteristics can be effectively captured, and the self-adaptive regression unit adopts an expert selection mechanism, and can adaptively adjust parameters in the regression process according to different fault modes, so that the recognition capability of the model on various fault modes is improved.
Furthermore, the embodiment of the invention also designs a training method of the neural network model. The training steps of the embodiment of the invention are as follows:
1) A pre-constructed training sample set is obtained.
2) And inputting the training sample set into a preset neural network model in batches, and calculating the classification error, the regression error and the total loss corresponding to the expert diversity constraint item corresponding to the training sample set through a preset mixed loss function.
The invention adopts a mixed loss function, solves the problems of unbalanced fault category and accumulated regression error by combining classification loss, regression error and expert unit parameter constraint, and improves the accuracy of fault diagnosis.
When in specific implementation, the method comprises the following steps:
and a-determining a weighted regression loss weight according to the number of samples corresponding to different sample categories of a preset training sample set.
In one embodiment, regression loss of high-frequency faults can be reduced through an exponential decay function, excessive fitting of a plurality of models is avoided, model bias caused by unbalanced categories is relieved, and diagnosis precision of rare faults is improved.
In one embodiment, the weighted regression loss may be calculated by the following formula:
With reference to the formula, the embodiment of the invention dynamically adjusts the weight of the regression loss according to the number of samples, thereby relieving the problem of unbalanced category. Wherein the denominator partCharacterization by category sample numberThe weight of the regression loss is adjusted so that,Is an exponential function with a natural constant as a base; is a class balancing factor, e.g., set to 0.1.Is the total number of fault categories.Is the firstClass prediction values; Is the firstA true-like value; is the L2 norm.Is the number of samples of the kth class.The regression loss weight coefficient is, for example, set to 0.3.
And b, respectively carrying out parameter space differentiation constraint on the parameter vectors of a plurality of preset expert units according to a preset parameter vector constraint algorithm, and determining expert diversity constraint loss corresponding to the training sample set.
The expert unit is used for learning the preset target fault mode characteristics in the training sample set based on the corresponding parameter space. The embodiment of the invention also designs expert diversity constraint loss for encouraging different expert units to learn the differential characteristics and preventing a plurality of expert units from learning similar fault mode characteristics. As different experts are concerned with temperature characteristics and ignore pressure characteristics. In one embodiment, the expert unit may be encouraged to concentrate on the differential features by minimizing cosine similarity between the parameter vectors, thereby improving the resolution of the model to the composite fault. Specifically, it is characterized in that:。
Wherein, theFor the expert diversity weight coefficient, the parameter orthogonality of different expert units is constrained, for example, set to 0.2.Is the s and the sThe included angle of the parameter vector of each expert unit,The space difference of the parameters of different expert units is characterized, and the representation avoids the collapse of the expert mode by minimizing the cosine similarity among the expert parameters.
And c, calculating a mixing loss function corresponding to the training sample set according to expert diversity constraint loss, weighted regression loss weight and preset classification loss.
In one embodiment, cross entropy loss may be employed as the classification loss. In combination with the above steps, the following formula can be referred to for the calculation of the mixing loss function:
in the formula,For the mixing loss function, the overall training objective is characterized.Is a classification loss; For the classification loss weight coefficient, for example, set to 0.2.For the regression loss weight coefficient, for example, set to 0.3; for the expert diversity weight coefficient, the parameter orthogonality of different expert units is constrained, e.g., set to 0.2.
3) And updating parameters of the neural network model according to the total loss.
The parameter updating corresponding to the embodiment of the invention refers to the following formula:
in the formula,And characterizing the state of the parameters after incremental learning for the updated model parameters.The state of the parameter before updating is represented as the current model parameter.For learning rate, for example, set to 0.001.Gradient of the model parameters for the mixing loss function.Is an elastic weight coefficient, for example, set to 0.3; And characterizing the parameter state before the last update for the historical model parameters, and using the parameter state for elastic weight constraint. With reference to the formula, the embodiment of the invention performs parameter update amplitude constraint on the neural network model based on the parameter state corresponding to the model parameter by acquiring the model parameter corresponding to each iteration of the neural network model. Further, extracting total loss corresponding to a pre-stored memory sample from a pre-set memory bank, and updating parameters of the neural network model according to the total loss corresponding to the memory sample. The memory samples are samples in the training sample set.
It should be noted that the number of the substrates,Is used to limit the update amplitude of parameters, keep history knowledge, and destroy the existing fault diagnosis capability due to the data of the new fault mode by using the elastic weight coefficientThe parameter updating direction is constrained, catastrophic forgetting is prevented, and further rapid adaptation of the model to a new fault mode is realized, and meanwhile, stable diagnosis performance of the original fault is maintained.
Further, the above formula also includesRepresenting the gradient of the mixing loss function to the model parameters calculated based on the ith sample stored in the memory.For memory banks, store representative samples;Is the ith sample stored in the memory bank.For the memory bank gradient weight coefficient, the contribution strength of the memory sample to parameter updating is represented, for example, set to 0.5.
Further, failure modes of the energy storage and liquid cooling system may evolve over time, such as equipment aging, resulting in failure feature shifts. The prior art, when faced with a new failure mode, generally requires full retraining of the model, resulting in wasted computational resources and time delays. The embodiment of the invention also adopts an incremental learning method based on the double memory banks, so that full retraining is not needed when a new fault mode occurs, calculation and time expenditure is reduced, and the model can be continuously adapted to the new fault mode and stable performance is maintained through updating the double memory banks.
In specific implementation, the embodiment of the invention determines the new sample corresponding to the training sample set through online learning. Further, new samples may be measured by a kernel functionAnd memory bankOnly the difference samples are reserved, so that the diversity of the coverage data distribution of the memory bank is ensured, the redundancy of the memory bank samples in the incremental learning process is avoided, and the generalization capability of the model to the edge cases is improved. Further, based on the corresponding difference samples, the memory bank is updated with data. Wherein, the update strategy of the memory bank refers to the following formula:
in the formula,Performing assignment operation; representing new fault data received in the online learning process as a new sample at the current moment; for the similarity threshold, for example, setting to 0.8, ensuring that the memory bank covers data distribution boundary samples; For a kernel function based on feature space similarity, a similarity measure between samples is characterized, and the calculation mode is expressed as。For the kernel bandwidth of the kernel function based on feature space similarity, the similarity decay rate is controlled, e.g., set to 0.1.
4) And in the parameter updating process of the neural network model, carrying out probability random discarding on the characteristic elements of the training sample set based on a preset time sensitive discarding strategy.
In order to make the network structure slightly different at each iteration, the forced model learns more robust features, and the prior art is realized by randomly discarding part of neurons. However, the conventional Dropout easily breaks time continuity in the monitoring data of the energy storage liquid cooling system, which easily results in that the model cannot effectively capture the time sequence dependency relationship. The invention adopts the time sequence sensitive Dropout based on the feature importance, and the generalization capability of the model is improved by dynamically adjusting the probability of the Dropout while maintaining the time sequence continuity. In specific implementation, the probability of random discard in the embodiment of the invention is calculated based on the feature amplitude and diversity corresponding to the feature elements. Specifically, reference is made to the following formula:
in the formula,For a timing sensitive Dropout function, residual information is retained to maintain timing continuity when features are discarded.And the multi-scale residual error term at the t moment of the first layer is obtained by multi-scale residual error connection calculation.Is the characteristic vector of the first layer at the t moment; For element-by-element multiplication; for mask vectors, the per-probability is characterizedThe features are randomly discarded.
The calculation mode of the mask vector is expressed as follows:
Wherein, theIndicating compliance with a certain distribution; In the case of a bernoulli distribution,Characterizing each feature element to probabilityAnd (5) reserving. This distribution can be seen as a distribution from bernoulli (B (1,) Samples are taken, wherein the probability of success (i.e., neurons are retained) is (1-) The probability of failure (i.e. neurons discarded) is). It can be ensured that the network does not depend excessively on any single neuron, facilitating more efficient feature learning.
Wherein, the abovePreserving probability for the features, dynamically adjusting discarding probability according to the feature amplitude and diversity, and expressing the calculation mode asCharacterization features preserve the negative correlation of probability with magnitude and diversity, suppressing overfitting of high activation or single features.Is characteristic entropy, and the calculation mode is expressed asThe diversity of the feature distribution is characterized.The first layer is the t time featureDimension; as a logarithmic function, the default base is 10; Is the dimension of the feature vector at time t of the first layer.Characterizing the control intensity of the characteristic amplitude for the first control intensity; for the second control intensity, the control intensity of the feature diversity is characterized.
It should be noted that, the high activation feature (such as a sudden increase in pressure value) in the monitoring data of the energy storage liquid cooling system may include critical fault information, but a single feature (such as only abnormal temperature) may easily cause overfitting. Embodiments of the present invention by combining amplitude termsAnd entropy termAnd the over fitting tendency of the high-activation or low-diversity features is restrained, and the robustness of the model to noise interference is improved on the premise of maintaining the time sequence continuity.
4) And selecting the current optimal weight parameters until the neural network model meets the preset training requirement, and generating a target neural network model.
In combination with the steps, a small-batch gradient descent method can be adopted to update parameters by an Adam optimizer. The initial learning rate can be set to 0.001 and dynamically adjusted by adopting a cosine annealing strategy. In one embodiment, 256 time window samples may be input per batch, and the total loss of classification errors, regression errors, and expert diversity constraints is calculated by a mixed loss function, back-propagating the updated network weights. In the training process, the accuracy and the F1 fraction are evaluated on the verification set after each iteration, and if the verification loss is not reduced in 10 continuous rounds, the training is stopped in advance, so that the overfitting is prevented. And the final model selects weight parameters with optimal performance of the verification set, and freezes the network for online deployment.
In summary, the embodiment of the invention solves the problem of bias of classification loss to unbalanced data in the traditional method by adopting the mixed loss function, particularly for rare fault class processing, improves the precision of fault diagnosis by optimizing classification and regression errors, and enhances the generalization capability of the model in a multi-fault mode by expert diversity constraint.
In addition, the traditional Dropout method possibly damages the time sequence of the data when processing the time sequence data, and the invention provides a time sequence sensitive Dropout strategy, which can effectively inhibit overfitting while maintaining the time sequence continuity. The invention also adopts the incremental learning method based on the double memory banks, does not need full retraining when a new fault mode occurs, adapts to the new fault mode through an incremental updating mode, solves the problem that the traditional model needs retraining when facing the new mode, obviously reduces the calculation and time expenditure, simultaneously avoids the catastrophic forgetting phenomenon, and maintains the long-term stability and accuracy of the model.
Further, the embodiment of the invention also utilizes the line graph with the confidence interval to evaluate the stability of the online incremental learning strategy. Fig. 5 shows a schematic diagram of an effect corresponding to online incremental learning according to an embodiment of the present invention. According to the scene that the new fault mode continuously appears in the continuous operation process of the equipment, the embodiment of the invention compares the traditional full-quantity updating method with the double-memory-bank strategy of the technology. Referring to fig. 5, the horizontal axis represents the online learning round, the vertical axis diagnosis accuracy reflects the comprehensive diagnosis capability of the model to the new and old fault modes, the confidence interval of the green curve corresponding to the technology is 100% in 3 groups of experiments, which is higher than that of the comparison method and is higher than the threshold confidence interval (95% confidence interval) commonly used in the field for evaluating whether the confidence reaches an excellent level, which indicates that the sample diversity maintaining mechanism of the memory effectively balances the new knowledge absorption and the old knowledge retention, while the red curve of the traditional method obviously declines after the fifth round, the confidence interval continuously expands, and the performance fluctuation caused by catastrophic forgetting is shown, which indicates that the elastic weight constraint of the technology plays a role in improving the stability of the system.
Further, on the basis of the foregoing embodiment, fig. 6 shows a schematic structural diagram corresponding to the embodiment, and referring to fig. 6, the apparatus includes a data acquisition module 100 configured to acquire an operation monitoring parameter corresponding to a preset energy storage liquid cooling system, the operation monitoring parameter includes a monitoring parameter corresponding to a preset key device of the energy storage liquid cooling system, a data processing module 200 configured to divide data of the operation monitoring parameter based on a local noise level and a time scale change of the operation monitoring parameter to obtain segmented data corresponding to a plurality of time periods respectively, a calculation module 300 configured to determine a denoising degree corresponding to the segmented data according to a noise strength corresponding to the segmented data, a denoising module 400 configured to denoise the segmented data based on the denoising degree, determine a parameter to be measured corresponding to each time scale respectively, an execution module 500 configured to input the parameter to be measured into a neural network model trained in advance, perform classification prediction on the parameter to be measured through the neural network model, and determine a fault type prediction result corresponding to the parameter to be measured, and an output module 600 configured to determine a fault state corresponding to the energy storage liquid cooling system according to the fault type prediction result.
The energy storage and liquid cooling system fault prediction device provided by the embodiment of the invention has the same technical characteristics as the method embodiment, so that the same technical problems can be solved, and the same technical effects can be achieved.
Further, the execution module 500 is further configured to perform feature capturing by using a time sequence feature dynamic fusion layer of a neural network model and checking parameters to be measured by using a plurality of convolution check of different scales, determine local target features corresponding to the parameters to be measured, perform adaptive activation path selection on the parameters to be measured by using a preset expert selection matrix by using an adaptive regression unit of the neural network model, determine high-dimensional feature representations corresponding to the parameters to be measured, display and extract time scale features corresponding to the local target features and the high-dimensional feature representations respectively based on a preset step length convolution operation, perform adaptive fusion on the time scale features corresponding to the local target features and the time scale features corresponding to the high-dimensional feature representations based on contribution of the parameters to be measured to the time sequence feature dynamic fusion layer or the adaptive regression unit, determine target fusion features corresponding to the parameters to be measured, perform fault category probability prediction on the target fusion features, and determine fault category prediction results corresponding to the parameters to be measured.
The execution module 500 is further configured to perform nonlinear mapping on the feature parameter of each moment of the parameter to be measured through a preset multi-layer perceptron, determine an inquiry vector corresponding to the parameter to be measured at each moment, determine a convolution kernel attention weight corresponding to the feature parameter of each moment of the parameter to be measured according to the inquiry vector and an attention mechanism of a key vector of a preset dynamic convolution kernel, capture a time sequence feature of the parameter to be measured according to the convolution kernel weight and a time attention weight corresponding to the feature parameter, and determine a local target feature corresponding to the parameter to be measured, wherein the time attention weight is determined by extracting local context information of the parameter to be measured by using a time sliding window.
The execution module 500 is further configured to map the parameter to be measured to an expert probability distribution by using a preset expert selection matrix, determine expert units corresponding to the feature parameters of the parameter to be measured at each moment based on the expert probability distribution, perform nonlinear transformation on the parameter to be measured based on the regression parameters corresponding to the expert units, and determine a high-dimensional feature representation corresponding to the parameter to be measured.
The data processing module 200 is further configured to determine a data noise difference in the current time period according to a data noise level corresponding to each time point of a preset time period of the operation monitoring parameter, determine a time segmentation point corresponding to the operation monitoring parameter based on the data noise difference, and divide the operation monitoring parameter into segments corresponding to a plurality of time scales based on the time segmentation point and a preset time segmentation length penalty term.
The execution module 500 is further configured to obtain a pre-constructed training sample set, input the training sample set to a preset neural network model in batches, calculate a classification error, a regression error and a total loss corresponding to expert diversity constraint terms corresponding to the training sample set through a preset mixed loss function, update parameters of the neural network model according to the total loss, and perform probability random discarding on feature elements of the training sample set based on a preset time sensitive discarding strategy in a parameter updating process of the neural network model, calculate a probability of random discarding based on feature amplitude and diversity corresponding to the feature elements until the neural network model meets a preset training requirement, and select a current optimal weight parameter to generate a target neural network model.
The execution module 500 is further configured to determine a weighted regression loss weight according to the number of samples corresponding to different sample types of a preset training sample set, respectively perform parameter space differentiation constraint on parameter vectors of a plurality of preset expert units according to a preset parameter vector constraint algorithm, determine expert diversity constraint loss corresponding to the training sample set, wherein the expert units are configured to learn preset target failure mode features in the training sample set based on the corresponding parameter spaces, and calculate a mixing loss function corresponding to the training sample set according to the expert diversity constraint loss, the weighted regression loss weight and the preset classification loss.
The execution module 500 is further configured to obtain a model parameter corresponding to each iteration of the neural network model, perform parameter update amplitude constraint on the neural network model based on a parameter state corresponding to the model parameter, extract a total loss corresponding to a pre-stored memory sample from a pre-set memory bank, and perform parameter update on the neural network model according to the total loss corresponding to the memory sample, where the memory sample is a sample in a training sample set.
The execution module 500 is further configured to determine a new sample corresponding to the training sample set through online learning, measure similarity between the new sample and the training sample set, determine a difference sample between the new sample and the training sample set, and update data of the memory based on the difference sample.
The embodiment of the invention also provides an electronic device, which comprises a memory, a processor and a computer program stored on the memory and capable of running on the processor, wherein the processor executes the computer program to realize the steps of the method shown in any one of the figures 1 to 2. Embodiments of the present invention also provide a computer readable storage medium having a computer program stored thereon, which when executed by a processor performs the steps of the method shown in any of the above figures 1 to 2. The embodiment of the present invention further provides a schematic structural diagram of an electronic device, as shown in fig. 7, where the electronic device includes a processor 71 and a memory 70, where the memory 70 stores computer executable instructions that can be executed by the processor 71, and the processor 71 executes the computer executable instructions to implement the method shown in any of the foregoing fig. 1 to 2. In the embodiment shown in fig. 7, the electronic device further comprises a bus 72 and a communication interface 73, wherein the processor 71, the communication interface 73 and the memory 70 are connected by the bus 72.
The memory 70 may include a high-speed random access memory (RAM, random Access Memory), and may further include a non-volatile memory (non-volatile memory), such as at least one magnetic disk memory. The communication connection between the system network element and the at least one other network element is achieved via at least one communication interface 73 (which may be wired or wireless), which may use the internet, a wide area network, a local network, a metropolitan area network, etc. The Bus 72 may be an ISA (Industry Standard Architecture ) Bus, a PCI (PERIPHERAL COMPONENT INTERCONNECT, peripheral component interconnect standard) Bus, or EISA (Extended Industry Standard Architecture ) Bus, etc., or an AMBA (Advanced Microcontroller Bus Architecture, standard for on-chip buses) Bus, where AMBA defines three buses, including an APB (ADVANCED PERIPHERAL Bus) Bus, an AHB (ADVANCED HIGH-performance Bus) Bus, and a AXI (Advanced eXtensible Interface) Bus. The bus 72 may be classified as an address bus, a data bus, a control bus, or the like. For ease of illustration, only one bi-directional arrow is shown in FIG. 7, but not only one bus or type of bus.
The processor 71 may be an integrated circuit chip with signal processing capabilities. In implementation, the steps of the above method may be performed by integrated logic circuitry in hardware or instructions in software in the processor 71. The Processor 71 may be a general-purpose Processor, including a central processing unit (Central Processing Unit, CPU), a network Processor (Network Processor, NP), a digital signal Processor (DIGITAL SIGNAL Processor, DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable gate array (FPGA) or other Programmable logic device, discrete gate or transistor logic device, or discrete hardware components. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like. The steps of the method disclosed in connection with the embodiments of the present application may be embodied directly in the execution of a hardware decoding processor, or in the execution of a combination of hardware and software modules in a decoding processor. The software modules may be located in a random access memory, flash memory, read only memory, programmable read only memory, or electrically erasable programmable memory, registers, etc. as well known in the art. The storage medium is located in a memory and the processor 71 reads the information in the memory and in combination with its hardware performs the method as shown in any of the foregoing figures 1 to 2.
The computer program product of the method and the device for predicting the failure of the energy storage liquid cooling system provided by the embodiment of the invention comprises a computer readable storage medium storing program codes, wherein the instructions included in the program codes can be used for executing the method described in the method embodiment, and specific implementation can be referred to the method embodiment and will not be repeated here. It will be clear to those skilled in the art that, for convenience and brevity of description, reference may be made to the corresponding process in the foregoing method embodiment for the specific working process of the above-described system, which is not described herein again. In addition, in the description of embodiments of the present invention, unless explicitly stated and limited otherwise, the terms "mounted," "connected," and "connected" are to be construed broadly, and may be, for example, fixedly connected, detachably connected, or integrally connected, mechanically connected, electrically connected, directly connected, indirectly connected via an intermediate medium, or in communication between two elements. The specific meaning of the above terms in the present invention will be understood by those skilled in the art in specific cases. The functions, if implemented in the form of software functional units and sold or used as a stand-alone product, may be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention may be embodied essentially or in a part contributing to the prior art or in a part of the technical solution, in the form of a software product stored in a storage medium, comprising several instructions for causing a computer device (which may be a personal computer, a server, a network device, etc.) to perform all or part of the steps of the method according to the embodiments of the present invention. The storage medium includes a U disk, a removable hard disk, a Read-Only Memory (ROM), a random access Memory (RAM, random Access Memory), a magnetic disk, an optical disk, or other various media capable of storing program codes.
In the description of the present invention, it should be noted that the directions or positional relationships indicated by the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc. are based on the directions or positional relationships shown in the drawings, are merely for convenience of describing the present invention and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a specific orientation, be configured and operated in a specific orientation, and thus should not be construed as limiting the present invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and are not to be construed as indicating or implying relative importance.
It should be noted that the foregoing embodiments are merely illustrative embodiments of the present invention, and not restrictive, and the scope of the invention is not limited to the foregoing embodiments, but it should be understood by those skilled in the art that any modification, variation or substitution of some technical features described in the foregoing embodiments may be easily made within the scope of the present invention without departing from the spirit and scope of the technical solutions of the embodiments. Therefore, the protection scope of the invention is subject to the protection scope of the claims.