Disclosure of Invention
The disclosure provides a training method, a device, a medium and a prediction method for a bearing residual life prediction model, so as to solve the problem that the residual life of a bearing cannot be accurately predicted.
According to a first aspect of the disclosure, a method for training a residual life prediction model of a bearing is provided, the method comprises the steps of obtaining a bearing vibration signal, conducting adaptive frequency processing on the bearing vibration signal to obtain fusion characteristics, conducting exponential smoothing processing on the fusion characteristics to obtain exponential smoothing output, conducting Hash wavelet characteristic enhancement processing on the exponential smoothing output to obtain Hash wavelet characteristic enhancement output, conducting division on the Hash wavelet characteristic enhancement output to obtain a training set, training an initial bearing residual life prediction model by the aid of the obtained training set to obtain a target bearing life prediction model, wherein the initial bearing residual life prediction model comprises an encoder and a decoder, the encoder carries out convolution operation on the bearing vibration signal by means of a one-dimensional convolution layer, the output of the one-dimensional convolution layer serves as the input of an adaptive frequency processing module, the output of the adaptive frequency processing module serves as the input of a cycle time sequence convolution module, the output of the cycle time sequence convolution module serves as the input of a bidirectional gating circulation unit, and the decoder carries out attention processing on the output of the bidirectional gating circulation unit, and the attention processing module serves as the input of the unidirectional gating module, and the adaptive frequency processing module serves as the input of the adaptive frequency processing module.
In some embodiments, the adaptive frequency processing of the bearing vibration signal comprises performing fast Fourier transform on the bearing vibration signal to obtain frequency domains, calculating energy of each frequency domain, performing normalization processing to obtain normalized frequency domain energy, generating an adaptive frequency domain mask based on the normalized frequency domain energy and an adaptive threshold parameter, multiplying the frequency domain with corresponding complex weights to obtain a weighted frequency domain, multiplying the frequency domain with the adaptive frequency domain mask to obtain a masking frequency domain, performing inverse fast Fourier transform on the weighted frequency domain and the masking frequency domain to obtain a time domain, extracting features of the time domain through a plurality of convolution layers, and fusing the features to obtain a fused feature.
In some embodiments, the performing exponential smoothing on the bearing vibration signal includes determining an exponential smoothing weight and a smoothing initial weight, performing fourier convolution on the bearing vibration signal according to the exponential smoothing weight to obtain a fourier convolution output, and calculating according to the fourier convolution output, the smoothing initial weight, and the bearing vibration signal to obtain an exponential smoothing output.
In some embodiments, the Haer wavelet feature enhancement processing is performed on the bearing vibration signal, and the Haer wavelet feature enhancement processing comprises the steps of performing dimension expansion on the bearing vibration signal to obtain a change result, performing two-dimensional convolution and Haer wavelet transformation on the change result to obtain a transformation result, performing two-dimensional convolution and transposition on the transformation result, performing Haer wavelet transformation to obtain an original feature, removing a third dimension of the original feature to obtain a conversion feature, and performing feature fusion on a second dimension of the conversion feature to obtain a Haer wavelet feature enhancement output.
In some embodiments, the output of the adaptive frequency processing module is used as input of a cyclic timing convolution module, wherein the cyclic timing convolution module processes input data, and the cyclic timing convolution module comprises the steps of normalizing the input data to obtain an affine transformation result, performing fast Fourier transform on the affine transformation result to obtain a cyclic convolution result, performing first-layer linear transformation and activation on the cyclic convolution result to obtain an activation result, performing second-layer linear transformation on the activation result to obtain a second-layer linear transformation result, and performing de-normalization on the second-layer linear transformation result to obtain a de-normalization result.
In some embodiments, the normalizing the input data to obtain affine transformation results includes: Obtaining the average value of the input dataWherein, the method comprises the steps of, wherein,The sequence of input data is represented as such,The length of the data sequence is indicated,Representing sequence numbers, according to the formula: Obtaining standard deviationWherein, the method comprises the steps of, wherein,Representing a constant, according to the formula: obtaining a normalization resultAnd according to the formula: Obtaining affine transformation resultsWherein, the method comprises the steps of, wherein,AndAll represent a parameter that can be learned.
In some embodiments, the performing the fast fourier transform on the affine transformation result to obtain the cyclic convolution result includes: Obtaining a cyclic convolution resultWherein, the method comprises the steps of, wherein,The fast fourier transform is represented by a set of coefficients,Representing the inverse fast fourier transform of the signal,The convolution kernel is represented as a function of the convolution kernel,Representing element multiplication; the first layer linear transformation and activation are carried out on the cyclic convolution result to obtain an activation result, which comprises the following steps: Obtaining the activation resultWherein, the method comprises the steps of, wherein,Representing the weight matrix of the fully connected layer,The term of the bias is indicated,And the second layer linear transformation is carried out on the activation result to obtain a second layer linear transformation result, which comprises the following steps of: Obtaining a second layer linear transformation resultWherein, the method comprises the steps of, wherein,Representing the weight matrix of the fully connected layer,And (3) representing the bias term, and performing denormalization on the second-layer linear transformation result to obtain a denormalization result, wherein the denormalization result comprises the following steps of: obtaining a denormalized result。
According to a second aspect of the disclosure, a device for training a residual life prediction model of a bearing is provided, which comprises an acquisition module for acquiring a bearing vibration signal, an adaptive frequency processing module for performing adaptive frequency processing on the bearing vibration signal to obtain a fusion characteristic, an exponential smoothing processing module for performing exponential smoothing processing on the fusion characteristic to obtain an exponential smoothing output, a Haer wavelet characteristic enhancement processing module for performing Haer wavelet characteristic enhancement processing on the exponential smoothing output to obtain a Haer wavelet characteristic enhancement output, a division module for dividing the Haer wavelet characteristic enhancement output to obtain a training set, and a training module for training the residual life prediction model of the initial bearing by using the acquired training set to obtain a target bearing life prediction model, wherein the initial bearing residual life prediction model comprises an encoder and a decoder, the encoder performs a processing step on the bearing vibration signal by using a one-dimensional convolution layer, the output of the one-dimensional convolution layer is used as an input of the adaptive frequency processing module, the output of the adaptive frequency processing module is used as an input of the cycle time sequence convolution module, the output of the cycle time sequence convolution processing module is used as an input of the cycle time sequence convolution module, and the output of the cycle time sequence convolution processing module is used as an input of a one-way control unit, and the output of the cycle time sequence unit is used as a one-way control unit.
According to a third aspect of the present disclosure, there is provided a method for predicting the residual life of a bearing, the method comprising acquiring a real-time bearing vibration signal, inputting the real-time bearing vibration signal into a residual life prediction model of the bearing trained by the residual life prediction model training method of the bearing as described above, and predicting the residual life of the bearing.
According to a fourth aspect of the present disclosure, there is provided a computer readable storage medium having stored thereon computer program instructions which, when executed by a processor, implement a bearing remaining life prediction model training method as described above or a bearing remaining life prediction method as described above.
By adopting the technical scheme, the method has the beneficial technical effects that the method for processing the self-adaptive frequency is adopted, the characteristic representation is enhanced by utilizing Fourier analysis, the long-term and short-term interaction of the bearing vibration information is captured, and meanwhile, the noise is reduced through the self-adaptive threshold.
The method adopts an exponential smoothing processing method, combines fast Fourier transformation and exponential weighted smoothing, can effectively smooth input signals, reduces the influence of short-term stirring, and simultaneously reserves long-term trend.
The method for enhancing the characteristics of the Harr wavelet is adopted, the Harr wavelet is utilized to decompose the vibration signal into sub-bands with different frequencies, the low-frequency trend and the high-frequency fault information are respectively extracted, the capturing capacity of key characteristics of the bearing vibration signal is enhanced after the characteristic resolution is reduced, and the accuracy and the robustness of model prediction are improved.
The method adopts cyclic time sequence convolution, reduces the interference of external environment on data distribution through dynamic mean value removal normalization, and improves the robustness and generalization capability of the model.
The method for self-adapting frequency is added in the stages of the encoder and the decoder, the capacity of decoding the complex time sequence of the model is improved, and the characteristic mining of the vibration signal is deeper and stronger.
As above, when the historical bearing vibration signal is preprocessed, the adaptive frequency processing, the exponential smoothing processing and the Harr wavelet characteristic enhancement processing are sequentially carried out, the characteristic representation of the bearing vibration signal is enhanced, long-term and short-term interaction of the vibration signal can be captured, noise is removed, meanwhile, the influence of short-term stirring is reduced through the smoothing processing, finally, the bearing vibration signal is decomposed into different frequency sub-bands by utilizing the Harr wavelet characteristic enhancement, low-frequency trend and high-frequency fault information are respectively extracted, the capturing capacity of key characteristics of the bearing vibration signal is enhanced after the characteristic resolution is reduced, and the obtained bearing vibration signal is focused on the key characteristics, so that the prediction precision of the model can be improved by utilizing the training models.
In the convolutional neural network, the encoder combines a one-dimensional convolutional layer, an adaptive frequency processing module, a cyclic timing convolutional module and a bidirectional gating cyclic unit, and can weight and filter the frequency domain representation of the bearing vibration signal and effectively focus on important information in a specific frequency range. By calculating the frequency domain energy of the bearing vibration signal and normalizing, the model can automatically adjust which frequency components have larger contribution, thereby realizing dynamic frequency domain enhancement. The decoder combines the attention mechanism and the unidirectional gating unit to calculate the attention weight between the current concealment state and the encoder output to obtain the context vector. Before the input of the unidirectional gating unit, the adaptive frequency processing module is used for fusing and compressing the attention output and the embedded features, so that redundancy is reduced, the focusing capability of the model on important features is enhanced, then the input and the context vector are spliced, and the output of the next time step is generated through the unidirectional gating unit. The improvement of the encoder and the decoder enables the model to have more accurate prediction capability, and provides support for maintenance, fault early warning and performance optimization of the bearing.
Detailed Description
Various exemplary embodiments of the present disclosure will now be described in detail with reference to the accompanying drawings. It should be noted that the relative arrangement of the components and steps, numerical expressions and numerical values set forth in these embodiments do not limit the scope of the present disclosure unless it is specifically stated otherwise.
Meanwhile, it should be understood that the sizes of the respective parts shown in the drawings are not drawn in actual scale for convenience of description.
The following description of at least one example embodiment is merely illustrative in nature and is in no way intended to limit the disclosure, its application, or uses.
Techniques, methods, and apparatus known to one of ordinary skill in the relevant art may not be discussed in detail, but are intended to be part of the specification where appropriate.
In all examples shown and discussed herein, any particular value is to be understood as being merely exemplary and not limiting. Thus, other examples of the exemplary embodiments may have different values.
It should be noted that like reference numerals and letters refer to like items in the following figures, and thus once an item is defined in one figure, no further discussion thereof is necessary in subsequent figures.
At present, the bearing is a vital component in various mechanical equipment and is widely applied to the fields of motors, automobiles, aerospace, industrial automation and the like. With the long-term operation of mechanical equipment, bearings can experience different degrees of wear and damage, directly affecting the stability and operating efficiency of the equipment. Therefore, the accurate prediction of the remaining service life of the bearing has important significance for maintenance, fault early warning and performance optimization of equipment, but at present, no way for accurately predicting the remaining service life of the bearing exists.
In view of this, the present disclosure proposes a method, apparatus, medium and method for training a bearing residual life prediction model, which employs a method of adaptive frequency processing, uses fourier analysis to enhance feature representation, and captures long-term and short-term interactions of bearing vibration information while mitigating noise through adaptive thresholds. The method adopts an exponential smoothing processing method, combines fast Fourier transformation and exponential weighted smoothing, can effectively smooth input signals, reduces the influence of short-term stirring, and simultaneously reserves long-term trend. The method for enhancing the characteristics of the Harr wavelet is adopted, the Harr wavelet is utilized to decompose the vibration signal into sub-bands with different frequencies, the low-frequency trend and the high-frequency fault information are respectively extracted, the capturing capacity of key characteristics of the bearing vibration signal is enhanced after the characteristic resolution is reduced, and the accuracy and the robustness of model prediction are improved. The method adopts cyclic time sequence convolution, reduces the interference of external environment on data distribution through dynamic mean value removal normalization, and improves the robustness and generalization capability of the model. The method for self-adapting frequency is added in the stages of the encoder and the decoder, the capacity of decoding the complex time sequence of the model is improved, and the characteristic mining of the vibration signal is deeper and stronger. The residual service life of the bearing can be predicted more accurately through the technology disclosed by the invention, and support is provided for maintenance, fault early warning and performance optimization of equipment.
It should be noted that convolutional neural networks are a type of feedforward neural network that includes convolutional calculation and has a deep structure. The method extracts local features of input data through convolution operation, forms complex feature representation through multi-layer convolution and pooling operation, and finally performs tasks such as classification or regression through a full connection layer.
The convolution neural network has the advantages that local characteristics are captured, vibration modes in different frequency ranges can be extracted through setting proper convolution kernel sizes and step sizes, so that the local structure of signals can be better understood, parameters are shared, the parameter number of the network can be greatly reduced, translational invariance can be achieved through convolution operation, translational change in the signals can be better processed, simple local characteristics are extracted by a low-level convolution layer, more abstract and complex characteristics are extracted by a high-level convolution layer, the structure of the vibration signals can be better represented, and CNN can be better suitable for different types of vibration signals and has better generalization capability.
FIG. 1 is a flow chart illustrating a method of training a bearing residual life prediction model according to some embodiments of the present disclosure. As shown in fig. 1, the method for training the residual life prediction model of the bearing includes steps S110 to S160.
In step S110, a bearing vibration signal is acquired.
In step S120, adaptive frequency processing is performed on the bearing vibration signal to obtain a fusion characteristic.
In some embodiments, the adaptive frequency processing of the bearing vibration signal comprises performing fast Fourier transform on the bearing vibration signal to obtain frequency domains, calculating energy of each frequency domain, performing normalization processing to obtain normalized frequency domain energy, generating an adaptive frequency domain mask based on the normalized frequency domain energy and an adaptive threshold parameter, multiplying the frequency domain with corresponding complex weights to obtain a weighted frequency domain, multiplying the frequency domain with the adaptive frequency domain mask to obtain a masking frequency domain, performing inverse fast Fourier transform on the weighted frequency domain and the masking frequency domain to obtain a time domain, extracting features of the time domain through a plurality of convolution layers, and fusing the features to obtain a fused feature.
As shown in FIG. 3, the adaptive frequency processing is performed by an adaptive frequency module (AE) in charge of first processing the input signalPerforming fast Fourier transform to convert it from time domain to frequency domain to obtain. Namely: In the followingIn order to input a signal to the device,Is a fast fourier transform function and,Is the output after the fast fourier transform.
Computing the energy of each frequency domain component after the fast fourier transformAnd carrying out normalization treatment to obtainTo balance the energy differences of the different frequency components. Namely:, In the followingIs the output after the fast fourier transform,Representing the energy of the signal in the frequency domain, calculating the modular long square of each frequency component,Is the calculated frequency domain energy.Is to an energy matrixThe median is found along the dimension of each batch, resulting in a median energy for each batch.A very small constant, avoiding zero errors, initially。Is normalized frequency domain energy, representing the relative energy of each frequency component.
Based on normalized energy and adaptive threshold parametersGenerating an adaptive frequency domain maskTo selectively retain important frequency components and discard non-important frequency components. Namely: In the followingIs the normalized frequency domain energy and,Is an adaptive threshold parameter, is a learnable parameter,Is a binary mask that is generated and,Is a function for judging positive and negative.
Subsequently, the signal and the corresponding complex weightMultiplying to obtain weighted frequency domain signalAnd in the adaptive maskingIs used for generating a frequency domain signal after maskingThe frequency domain signal is further adjusted. Namely:, In the followingIs the output after the fast fourier transform,Is a matrix of weights that are to be used,Is a binary mask that is generated and,Is the weighted frequency domain signal and,Is a masked frequency domain signal.
Finally, the processed frequency domain signal is converted back to the time domain through inverse fast Fourier transform to obtain an output signal. Namely: In the followingIs the weighted frequency domain signal and,Is the frequency domain signal after masking,Is an inverse fast fourier function of the data,Is a time domain signal after inverse fast fourier transform.
On the basis, the time domain features are further extracted through a convolution layer, and the final feature representation is obtained through fusion by combining the outputs of different convolution pathsThe whole signal processing and feature extraction process is completed. Namely:,,, In the followingThe time domain signal after the inverse fast fourier transform represents a convolution operation,,AndRepresenting convolution kernels of different sizes,Is a one-dimensional convolution function,Is the result of the convolution kernel 1,Is through an activation functionThe latter signal, increases its nonlinearity,Is the result of the convolution kernel 2,Is through an activation functionThe latter signal, increases its nonlinearity,AndIs the signal after the feature fusion,Is a feature after the third layer convolution processing.
The frequency domain weighting is carried out on the bearing vibration signal by utilizing the self-adaptive frequency masking method, so that high-frequency noise components can be effectively removed, and low-frequency information which is critical to the judgment of the health state of the bearing can be reserved.
In step S130, the fusion feature is subjected to an exponential smoothing process, so as to obtain an exponential smoothing output.
In some embodiments, the performing exponential smoothing on the fusion feature includes determining an exponential smoothing weight and a smoothing initial weight, performing fourier convolution on the bearing vibration signal according to the exponential smoothing weight to obtain a fourier convolution output, and calculating according to the fourier convolution output, the smoothing initial weight, and the bearing vibration signal to obtain an exponential smoothing output.
As shown in FIG. 4, the exponential smoothing process is processed by an exponential Smoothing Module (SM) by first calculating an exponential weighting coefficient for each time stepThe weights control the degree of smoothness of the input signal over the time step. Coefficient of exponential decayThe relative importance of the new information and the history information is determined. I.e.,In the formula,Is a smoothing factor which is used to smooth the image,Between 0 and 1, for controlling the degree of smoothing,Is in the time stepIs used to determine the smoothing weight of the (c),Is the initial weight of the exponential smoothing,Is the total number of time steps.
Then, the input signal is subjected to Fourier convolution, and the convolution kernel is the calculated exponential smoothing weight. Namely: In which, in the process,Is an input signal which is provided with a signal,Is in the time stepIs used to determine the smoothing weight of the (c),Is the output of the convolution fourier transform,The convolution fourier function formula is as follows:
,, In which, in the process,Is a fast fourier transform of the data obtained,Is an input signal which is provided with a signal,Is the convolution kernel of the incoming signal,The signal is converted into the frequency domain,The convolution kernel is converted into the frequency domain, followed by a dot product in the frequency domain,Is the convolution and complex conjugate of the representation in the frequency domain,Is the result of the convolution in the frequency domain,Is an inverse fast Fourier function, converts it into the time domain, and employs a rolling operationEnsuring alignment of convolution results to obtain Fourier convolution results。
After convolution, the initial weight is combined for initialization to control the smoothed initial value, the output of the exponential smoothing module is obtained,Is the output of the exponential smoothing module.
And the vibration signal is smoothed by adopting an exponential smoothing algorithm, so that the influence of short-term fluctuation on life prediction is reduced, and the stability of data is improved.
In step S140, the exponential smoothing output is subjected to hal wavelet characteristic enhancement processing, so as to obtain a hal wavelet characteristic enhancement output.
In some embodiments, the Haer wavelet feature enhancement processing is performed on the exponential smoothing output, and the Haer wavelet feature enhancement processing comprises the steps of performing dimension expansion on a bearing vibration signal to obtain a change result, performing two-dimensional convolution and Haer wavelet transformation on the change result to obtain a transformation result, performing two-dimensional convolution and transposition on the transformation result, performing Haer wavelet transformation to obtain an original feature, removing a third dimension of the original feature to obtain a conversion feature, and performing feature fusion on a second dimension of the conversion feature to obtain the Haer wavelet feature enhancement output.
As shown in FIG. 5, the Harr wavelet feature enhancement processing is performed by a Harr wavelet feature enhancement module (HE), which expands the dimension of the input feature to provide more feature information channels, and matches the input state requirement of the module, and then performs a Harr filter on the input feature through forward Harr wavelet transformation to obtain a resolution-reduced result. Namely:, In which, in the process,Representing an increased input featureIs defined by a first dimension of the first dimension,Representing the specified dimension of the object,The results after the change are shown as such,A two-dimensional convolution operation is represented,Representing a haar filter, dividing the input into 4 parts, a similar part, a horizontal part, a vertical part, a diagonal part,Is the result of the haar wavelet transform reduced resolution.
And then, carrying out inverse haar wavelet transformation to restore the characteristics, re-adjusting the shape of the output tensor according to the sequence of the haar wavelet transformation, then, restoring the image from the transformed low-resolution space to the original characteristic space by inverse convolution operation, finally, removing the newly added third dimension, restoring the dimension of the characteristics by characteristic fusion, and obtaining the output of the haar wavelet characteristic enhancement module. Namely:, In which, in the process,Is the result of the haar transform reduced resolution,Representing the operation of two-dimensional convolution and transposition,Representing the haar filter and the filter is applied,Is the original feature recovered by the inverse transform,Is to removeA function of the third dimension is provided,Representing the specified dimension of the object,Is a function of the feature fusion,Expressed in the removal of the third dimensionFeature fusion is performed in a second dimension of the (c) to recover the feature of the reduced dimension,Is the output of the hal wavelet feature enhancement module.
The frequency components related to fault development and life prediction in the bearing vibration signals are separated and focused by adopting haar wavelet transformation, so that the accuracy and the robustness of model prediction are improved.
In step S150, the hal wavelet feature enhanced output is divided to obtain a training set.
In step S160, training an initial bearing residual life prediction model by using the obtained training set to obtain a target bearing life prediction model, wherein the initial bearing residual life prediction model comprises an encoder and a decoder, the encoder comprises a step of processing a bearing vibration signal by using a one-dimensional convolution layer, wherein the output of the one-dimensional convolution layer is used as the input of an adaptive frequency processing module, the output of the adaptive frequency processing module is used as the input of a cyclic timing convolution module, the output of the cyclic timing convolution module is used as the input of a bidirectional gating circulation unit, and the decoder comprises a step of processing the output of the bidirectional gating circulation unit by using the output of the attention mechanism processing module as the input of the adaptive frequency processing module, and the output of the adaptive frequency processing module is used as the input of a unidirectional gating unit and the output of the unidirectional gating unit is input to a full-connection layer.
In some embodiments, a cyclic time sequence convolution module for reducing interference of external environment on data distribution is introduced into an initial bearing residual life prediction model, wherein the cyclic time sequence convolution module processes input data, and comprises the steps of normalizing the input data to obtain an affine transformation result, carrying out fast Fourier transformation on the affine transformation result to obtain a cyclic convolution result, carrying out first-layer linear transformation and activation on the cyclic convolution result to obtain an activation result, carrying out second-layer linear transformation on the activation result to obtain a second-layer linear transformation result, and carrying out de-normalization on the second-layer linear transformation result to obtain a de-normalization result.
As shown in fig. 6, in some embodiments, the normalizing the input data to obtain affine transformation results includes: Obtaining the average value of the input dataWherein, the method comprises the steps of, wherein,The sequence of input data is represented as such,The length of the data sequence is indicated,Representing sequence numbers, according to the formula: Obtaining standard deviationWherein, the method comprises the steps of, wherein,Representing a constant, according to the formula: obtaining a normalization resultAnd according to the formula: Obtaining affine transformation resultsWherein, the method comprises the steps of, wherein,AndAll represent a parameter that can be learned.
Performing fast fourier transform on the affine transformation result to obtain a cyclic convolution result, including:
According to the formula:
, Obtaining a cyclic convolution resultWherein, the method comprises the steps of, wherein,The fast fourier transform is represented by a set of coefficients,Representing the inverse fast fourier transform of the signal,The convolution kernel is represented as a function of the convolution kernel,Representing element multiplication;
The first layer linear transformation and activation are carried out on the cyclic convolution result to obtain an activation result, which comprises the following steps: Obtaining the activation resultWherein, the method comprises the steps of, wherein,Representing the weight matrix of the fully connected layer,The term of the bias is indicated,And the second layer linear transformation is carried out on the activation result to obtain a second layer linear transformation result, which comprises the following steps of: Obtaining a second layer linear transformation resultWherein, the method comprises the steps of, wherein,Representing the weight matrix of the fully connected layer,And (3) representing the bias term, and performing denormalization on the second-layer linear transformation result to obtain a denormalization result, wherein the denormalization result comprises the following steps of: obtaining a denormalized result。
Specifically, firstly, the dimension of an input feature is expanded, more feature information channels are provided, the input type state requirements of a module are matched, and then a haar filter is carried out on the input feature through forward haar wavelet transformation to obtain a resolution-reducing result. Namely:, In the followingRepresenting an increased input featureIs defined by a first dimension of the first dimension,Representing the specified dimension of the object,Is the result of the change after the change,A two-dimensional convolution operation is represented,Representing a haar filter, dividing the input into 4 parts, a similar part, a horizontal part, a vertical part, a diagonal part,Is the result of the haar wavelet transform reduced resolution.
And then, carrying out inverse haar wavelet transformation to restore the characteristics, re-adjusting the shape of the output tensor according to the sequence of the haar wavelet transformation, then, restoring the image from the transformed low-resolution space to the original characteristic space by inverse convolution operation, finally, removing the newly added third dimension, restoring the dimension of the characteristics by characteristic fusion, and obtaining the output of the haar wavelet characteristic enhancement module. Namely:, In the followingIs the result of the haar transform reduced resolution,Representing the operation of two-dimensional convolution and transposition,Representing the haar filter and the filter is applied,Is the original feature recovered by the inverse transform,Is to removeA function of the third dimension is provided,Representing the specified dimension of the object,Is a function of the feature fusion,Expressed in the removal of the third dimensionFeature fusion is performed in a second dimension of the (c) to recover the feature of the reduced dimension,Is the output of the hal wavelet feature enhancement module.
In a cyclic timing convolution module (RC), input data is normalized first to ensure proper scale for training stability and efficiency. Normalized by subtracting the meanAnd divided by standard deviationTo normalize the data so that the data is distributed over the same range. Namely:,,, In which, in the process,Is an input sequence of the sequence of inputs,Is the length of the sequence and,The number of the sequence is indicated,Indicating a constant, preventing zero-divide errors,Is the average value of the input data and,Is the standard deviation of the two-dimensional image,Is the result of the normalization and,Is the result of the affine transformation,AndIs a learnable parameter.
Normalizing the transformed resultThe circular convolution is realized through the fast Fourier transform, and then the feature mapping is carried out through the full connection layer, so that the feature is enhanced, and the expression capacity of the model is improved. Namely:,, In the followingIs a fast fourier transform of the data obtained,Is an inverse fast fourier transform of the data,Is the result of the affine transformation,Is a convolution kernel which is a convolution kernel,Is an element multiplication and is a method of element multiplication,Is the result of the cyclic convolution.,Is a weight matrix of the full connection layer,,Is a bias term.Is the result of the first layer linear transformation and activation,Is the function of the activation and,Is the result of the second layer linear transformation.
And finally, the result after the linear transformation is denormalized, and the data is restored to the original scale.In which, in the process,As a result of the linear variation of the second layer,AndIs a parameter that can be learned and is,Is the average value of the input data and,Is the standard deviation of the two-dimensional image,Is the result of the de-normalization and recovery of the data scale.
As shown in fig. 2, in some embodiments, the prediction method includes:
firstly, extracting vibration signal characteristics, and using an adaptive frequency module and an exponential smoothing module for preliminary vibration signal processing to obtain outputAnd calculating the frequency domain characteristics of the bearing vibration data such as the mean value, the root mean square, the kurtosis, the skewness, the peak-to-peak value, the variance, the peak factor, the pulse factor, the margin factor, the shape factor, the square root amplitude value, the waveform factor and the like. And then the characteristics are fused together to obtain the comprehensive vibration characteristics. Namely:, In the followingRepresenting the adaptive frequency module,Representing an exponential smoothing module,Representing a haar wavelet feature enhancement module,Is the result of applying the adaptive frequency module and the exponential smoothing module,,The average value is a computing function of the root mean square and other characteristics,Is a comprehensive vibration characteristic.
After feature extractionAnd an initial hidden stateThe characteristic representation is enhanced and long-term and short-term interaction is captured by inputting the characteristic representation into the encoder together, compressing the characteristic by convolution operation, and inputting the characteristic representation into the adaptive frequency module again to reduce noise. Then and initial hidden stateTogether input to a bi-directional gating unit to obtain an encoder outputAnd the current hidden state. Namely:
;
。
In the middle ofIs a one-dimensional convolution function,Is a characteristic of the integrated vibration, and is characterized by the integrated vibration,Representing the adaptive frequency module,Representing a cyclic timing convolution module,Is in an initial hidden state and is in a hidden state,Is a function of the feature fusion,Representing feature fusion in a second dimension of the adaptive frequency module output and the cyclic timing convolution module output,Is a two-way gate control unit,Is the output of the encoder and,Is the current hidden state.
The adaptive frequency module is again used for compressing and fusing the attention output at the stage of the decoderAnd context informationThis helps reduce redundancy and enhances the focusing ability of the model on important features. Namely:, In the followingIs the hidden state of the current time step,Is the hidden state of the previous step,Is thatThe function of the function is that,Is an exponential function of the number of times,Is the length of the time series,Is the output of the attention of the person,Is context information.
Output attentionAnd context informationFusion, inputting the fusion and the hidden state of the current time step into a gating unit to obtain outputWill thenAnd context informationThe linear layers are accessed together to obtain the decoder output for the current time step. Namely:, In which, in the process,Is a unidirectional gating unit, and the gating unit outputsHidden state,Is a linear function of the magnitude of the signal,Is the output of the decoder, i.e. the prediction result.
The XJTU bearing data set is provided by a unit and comprises the complete running to fault data of the rolling bearing running under 15 different working conditions, the sampling frequency used for collecting the data is 25.6kHz, the vibration signals of the bearing in the X direction and the Y direction are sampled, and 32768 samples are recorded every 60 seconds. The information of the data set and the task is shown in table one. For each, 3 tests were performed, namely bearing 1 and bearing 2 were set as training sets, the others were set as test sets.
Table-one XJTU bearing dataset task
As shown in table two, the evaluation index RMSE of the bearing dataset at XJTU was compared with that of the prior art.
Watch II
As shown in table three, the evaluation index MAE of the bearing dataset at XJTU was compared with that of the prior art.
Watch III
As shown in fig. 7 and 8, the residual life prediction framework of the bearing is verified on the XJTU bearing dataset, and the result shows that the method is superior to the existing method in root mean square error RMSE and mean absolute error MAE, and the effectiveness of the method is proved.
As above, the prediction results of the tasks of bearing 1_3 and bearing 2_3 of the proposed method at XJTU bearing dataset are as shown in fig. 7 and 8. The mission division of XUTU bearing data sets is shown in table one, and the method and some advanced bearing life prediction methods in recent years are shown in XJTU bearing data set pairs of RMSE and MAE in mission a and mission B, such as table two and table three.
The reader is noted that the input characteristic sequence is processed through the result convolutional neural network, the self-adaptive frequency domain module and the cyclic time sequence convolutional module, the frequency domain representation of the input signal is weighted and filtered, important information in a specific frequency range can be effectively focused, the frequency domain energy of the input signal is calculated and normalized, and the model can automatically adjust which frequency components have larger contribution to the signal, so that dynamic frequency domain enhancement is realized. The decoder combines the attention mechanism and the GRU to calculate the attention weight between the current concealment state and the encoder output to obtain the context vector. The attention output and embedded features can be fused and compressed using an adaptive frequency domain module prior to the GRU input, which helps reduce redundancy and enhance the focusing ability of the model on important features, and then concatenating the input with the context vector, generating the output for the next time step through the GRU layer.
Fig. 9 is a block diagram illustrating a bearing remaining life model training apparatus according to some embodiments of the present disclosure. As shown in fig. 9, the bearing remaining life model training apparatus 900 includes:
An acquisition module 910 configured to acquire a bearing vibration signal;
the adaptive frequency processing module 920 is configured to perform adaptive frequency processing on the bearing vibration signal to obtain a fusion characteristic;
an exponential smoothing processing module 930 configured to perform exponential smoothing processing on the bearing vibration signal to obtain an exponential smoothed output;
the hal wavelet characteristic enhancement processing module 940 is configured to perform hal wavelet characteristic enhancement processing on the bearing vibration signal to obtain a hal wavelet characteristic enhancement output;
the dividing module 950 is configured to divide the bearing vibration signal subjected to the adaptive frequency processing, the exponential smoothing processing, and the hal wavelet characteristic enhancement processing to obtain a training set and a test set;
The training module 960 is configured to train the initial bearing residual life prediction model by using the obtained training set to obtain a target bearing life prediction model, wherein the initial bearing residual life prediction model comprises an encoder and a decoder, the encoder comprises a step of processing a bearing vibration signal by using a one-dimensional convolution layer, the output of the one-dimensional convolution layer is used as the input of an adaptive frequency processing module, the output of the adaptive frequency processing module is used as the input of a cycle time sequence convolution module, the output of the cycle time sequence convolution module is used as the input of a bidirectional gating circulation unit, and the decoder comprises a step of processing the output of the bidirectional gating circulation unit by using the output of the attention mechanism processing module as the input of the adaptive frequency processing module, and the output of the adaptive frequency processing module is used as the input of a unidirectional gating unit, and the output of the unidirectional gating unit is input to the full connection layer.
In the device of the embodiment of the disclosure, a device for training a residual life model of a bearing is provided, the disclosure adopts a method for processing self-adaptive frequencies, utilizes Fourier analysis to enhance characteristic representation and captures long-term and short-term interaction of bearing vibration information, and simultaneously reduces noise through self-adaptive thresholds. The method adopts an exponential smoothing processing method, combines fast Fourier transformation and exponential weighted smoothing, can effectively smooth input signals, reduces the influence of short-term stirring, and simultaneously reserves long-term trend. The method for enhancing the characteristics of the Harr wavelet is adopted, the Harr wavelet is utilized to decompose the vibration signal into sub-bands with different frequencies, the low-frequency trend and the high-frequency fault information are respectively extracted, the capturing capacity of key characteristics of the bearing vibration signal is enhanced after the characteristic resolution is reduced, and the accuracy and the robustness of model prediction are improved. The method adopts cyclic time sequence convolution, reduces the interference of external environment on data distribution through dynamic mean value removal normalization, and improves the robustness and generalization capability of the model. The method for self-adapting frequency is added in the stages of the encoder and the decoder, the capacity of decoding the complex time sequence of the model is improved, and the characteristic mining of the vibration signal is deeper and stronger. The residual service life of the bearing can be predicted more accurately through the technology disclosed by the invention, and support is provided for maintenance, fault early warning and performance optimization of equipment.
FIG. 10 is a block diagram illustrating a bearing residual life model training device according to further embodiments of the present disclosure.
As shown in fig. 10, the bearing remaining life model training apparatus 1000 includes a memory 1010, and a processor 1020 coupled to the memory 1010. The memory 1010 is used to store instructions for performing a corresponding embodiment of the method for training a residual life model of a bearing. The processor 1020 is configured to perform the bearing remaining life model training method in any of the embodiments of the present disclosure based on instructions stored in the memory 1010.
FIG. 11 is a block diagram illustrating a computer system for implementing some embodiments of the present disclosure.
As shown in FIG. 11, computer system 1100 may be embodied in the form of a general purpose computing device. Computer system 1100 includes memory 1110, processor 1120, and bus 1130 that connects the different system components.
The memory 1110 may include, for example, system memory, nonvolatile storage media, and the like. The system memory stores, for example, an operating system, application programs, boot Loader (Boot Loader), and other programs. The system memory may include volatile storage media, such as Random Access Memory (RAM) and/or cache memory. The non-volatile storage medium stores, for example, instructions for performing a corresponding embodiment of at least one of the bearing remaining life model training methods. Non-volatile storage media include, but are not limited to, disk storage, optical storage, flash memory, and the like.
Processor 1120 may be implemented as discrete hardware components such as a general purpose processor, a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field Programmable Gate Array (FPGA) or other programmable logic device, discrete gates, or transistors. Accordingly, each module such as the acquisition module, the adaptive frequency processing module, the exponential smoothing processing module, the haar wavelet feature enhancement processing module, the division module, and the training module may be implemented by a Central Processing Unit (CPU) executing instructions of executing corresponding steps in a memory, or may be implemented by a dedicated circuit executing corresponding steps.
Bus 1130 may use any of a variety of bus architectures. For example, bus structures include, but are not limited to, an Industry Standard Architecture (ISA) bus, a Micro Channel Architecture (MCA) bus, and a Peripheral Component Interconnect (PCI) bus.
Computer system 1100 may also include input/output interfaces 1140, network interfaces 1150, storage interfaces 1160, and the like. These interfaces 1140, 1150, 1160, and memory 1110 and processor 1120 may be connected by bus 1130. The input/output interface 1140 may provide a connection interface for input/output devices such as a display, mouse, keyboard, etc. Network interface 1150 provides a connection interface for a variety of networking devices. The storage interface 1160 provides a connection interface for external storage devices such as a floppy disk, a USB flash disk, an SD card, and the like.
Various aspects of the present disclosure are described herein with reference to flowchart illustrations and/or block diagrams of methods, apparatus and computer program products according to embodiments of the disclosure. It will be understood that each block of the flowchart illustrations and/or block diagrams, and combinations of blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer readable program instructions.
These computer-readable program instructions may be provided to a processor of a general purpose computer, special purpose computer, or other programmable apparatus to produce a machine, such that the instructions, which execute via the processor, create means for implementing the functions specified in the flowchart and/or block diagram block or blocks.
These computer readable program instructions may also be stored in a computer readable memory that can direct a computer to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart and/or block diagram block or blocks.
The present disclosure may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects.
As shown in fig. 12, an embodiment of the present disclosure further provides a bearing remaining life prediction method, including steps S1210 to S1220.
In step S1210, a real-time bearing vibration signal is acquired;
In step S1220, the real-time bearing vibration signal is input to the bearing remaining life prediction model obtained by training and training using the above-mentioned bearing remaining life prediction model training method, to predict the bearing remaining life.
The present disclosure provides a method, apparatus, medium, and method for training a bearing residual life prediction model, which employs a method for adaptive frequency processing, utilizes fourier analysis to enhance feature representation, and captures long-term and short-term interactions of bearing vibration information while mitigating noise through an adaptive threshold. The method adopts an exponential smoothing processing method, combines fast Fourier transformation and exponential weighted smoothing, can effectively smooth input signals, reduces the influence of short-term stirring, and simultaneously reserves long-term trend. The method for enhancing the characteristics of the Harr wavelet is adopted, the Harr wavelet is utilized to decompose the vibration signal into sub-bands with different frequencies, the low-frequency trend and the high-frequency fault information are respectively extracted, the capturing capacity of key characteristics of the bearing vibration signal is enhanced after the characteristic resolution is reduced, and the accuracy and the robustness of model prediction are improved. The method adopts cyclic time sequence convolution, reduces the interference of external environment on data distribution through dynamic mean value removal normalization, and improves the robustness and generalization capability of the model. The method for self-adapting frequency is added in the stages of the encoder and the decoder, the capacity of decoding the complex time sequence of the model is improved, and the characteristic mining of the vibration signal is deeper and stronger. The residual service life of the bearing can be predicted more accurately through the technology disclosed by the invention, and support is provided for maintenance, fault early warning and performance optimization of equipment.
Thus far, the bearing remaining life prediction model training method, apparatus, medium, and prediction method according to the present disclosure have been described in detail. In order to avoid obscuring the concepts of the present disclosure, some details known in the art are not described. How to implement the solutions disclosed herein will be fully apparent to those skilled in the art from the above description.
Although specific embodiments of the disclosure have been described in detail by way of example, it will be appreciated by those skilled in the art that the above examples are for illustration only and are not intended to limit the scope of the disclosure. It will be appreciated by those skilled in the art that modifications may be made to the above embodiments without departing from the scope and spirit of the disclosure. The scope of the present disclosure is defined by the appended claims.