Disclosure of Invention
The embodiment of the invention provides an intelligent operation and maintenance and predictive maintenance method and system for a multi-layer structure of a rail transit platform door, which can solve the problems in the prior art.
In a first aspect of an embodiment of the present invention,
The intelligent operation and maintenance and predictive maintenance method for the multi-layer structure of the rail transit platform door is provided, and comprises the following steps:
a multi-layer data acquisition system of the rail transit platform door is established, the operation parameters of the platform door are acquired through a vibration sensor, a temperature sensor, a displacement sensor and a current sensor which are arranged on a platform door bottom mechanism, a driving mechanism and a door body structure, the operation parameters comprise vibration frequency, driving current, component temperature and door opening and closing displacement, and the operation parameters are transmitted to an edge computing unit through a field bus to perform data preprocessing, so that a platform door health state feature vector is generated;
Inputting the platform door health state feature vector into a deep neural network comprising a convolution layer, a pooling layer and a full-connection layer, training the deep neural network by combining historical maintenance data through time sequence correlation analysis and multi-scale feature extraction on the feature vector, and constructing a platform door fault prediction model based on deep learning;
Based on the output result of the platform door fault prediction model based on deep learning, a decision tree algorithm is adopted to dynamically evaluate the maintenance priority of the platform door part, an intelligent maintenance scheme comprising maintenance time, maintenance items and maintenance resources is generated, a maintenance instruction is pushed to on-site maintenance personnel through a mobile terminal, new data acquired in the maintenance process are fed back to the deep learning model for on-line optimization updating, and self-adaptive adjustment of the platform door maintenance strategy is realized.
Transmitting the operation parameters to an edge computing unit through a field bus for data preprocessing, and generating the platform door health state characteristic vector comprises the following steps:
Collecting platform door operation parameters through a vibration sensor arranged on a platform door bottom mechanism, a current sensor arranged on a platform door driving mechanism, a temperature sensor arranged on a platform door body structure and a displacement sensor, wherein the sampling frequency of the vibration sensor is 1000Hz, the sampling frequency of the temperature sensor is 1Hz, the sampling frequency of the displacement sensor is 100Hz, and the sampling frequency of the current sensor is 500Hz;
The edge computing unit adopts a dual-core ARM processor, performs 3-layer decomposition on the operation parameters by using a db4 wavelet basis function, filters high-frequency noise by a soft threshold method, identifies and eliminates abnormal values based on a3 sigma criterion, and uniformly aligns the operation parameters to a time axis of 100Hz by adopting a piecewise linear interpolation method;
Extracting speed curve characteristics, acceleration characteristics and displacement characteristics in the process of opening and closing a platform door as working condition characteristics, calculating an envelope spectrum of a vibration signal and a harmonic ratio of a current signal through a sliding time window as performance degradation characteristics, establishing environment influence characteristics by combining correlation analysis of a temperature change trend and vibration response, combining the working condition characteristics, the performance degradation characteristics and the environment influence characteristics to form 145-dimensional characteristic vectors, performing dimension reduction processing on the 145-dimensional characteristic vectors by adopting a principal component analysis method, and selecting principal components with accumulated contribution rate reaching 95% to generate the platform door health state characteristic vectors.
The platform door fault prediction model based on deep learning comprises a time sequence characteristic branch, a space characteristic branch and an attention characteristic branch;
The time sequence feature branch comprises three one-dimensional convolution layers connected in series, the convolution kernel sizes of the three one-dimensional convolution layers are 3, 5 and 7 respectively, the output of each one-dimensional convolution layer is connected with a largest pooling layer, the pooling window size of the largest pooling layer is 2, and the output of the time sequence feature branch is used for representing the time sequence feature of the health state feature vector;
The space characteristic branch comprises four two-dimensional convolution layers connected in series, the four two-dimensional convolution layers adopt 4 multiplied by 4 convolution kernels, the channel numbers of the four two-dimensional convolution layers are 32, 64, 128 and 256 in sequence, and the output of the space characteristic branch is used for representing the space characteristic of the health state characteristic vector;
The attention characteristic branch comprises a self-attention calculating unit, the self-attention calculating unit calculates association weights among the dimensions of the health state characteristic vector, the association weights are normalized through a softmax function to obtain normalized weights, and the output of the attention characteristic branch is used for representing the attention characteristic of the health state characteristic vector;
And carrying out feature fusion on the time sequence feature, the space feature and the attention feature to obtain a fusion feature, wherein the fusion feature is used for predicting the failure of the platform door.
The construction of the platform door fault prediction model based on deep learning comprises the following steps:
Acquiring a platform door health state feature vector, performing recursion graph analysis on the platform door health state feature vector, mapping the platform door health state feature vector to a high-dimensional phase space to obtain feature track points by setting a phase space reconstruction parameter with an embedding dimension of 5 and a time delay of 2, and calculating Euclidean distances between the feature track points to construct a recursion matrix;
Decomposing the platform door health state feature vector into a plurality of integral mode functions by a local mode decomposition method, and calculating instantaneous frequency and instantaneous amplitude to the integral mode functions to obtain microscale features;
The platform door health state feature vector is decomposed into a trend item, a period item and a residual item through a variation mode decomposition algorithm to obtain a macro-scale feature, and the micro-scale feature, the mesoscale feature and the macro-scale feature are input into a deep neural network for training to obtain a platform door fault prediction model.
Based on the output result of the platform door fault prediction model based on deep learning, dynamically evaluating the maintenance priority of the platform door part by adopting a decision tree algorithm, and generating an intelligent maintenance scheme comprising maintenance time, maintenance project and maintenance resource comprises the following steps:
Constructing a maintenance priority evaluation index based on a residual life predicted value and a fault type probability distribution output by a deep learning model, mapping the residual life predicted value in a segmented mode according to a time interval to obtain a health state index, calculating to obtain a fault risk index based on the product of the fault type probability distribution and a fault type hazard degree weight matrix, and calculating to obtain an operation influence index by a hierarchical analysis method in combination with the flow rate of a station door guest, the position of a station and the state of standby resources;
Inputting the health state index, the fault risk index and the operation influence index into a CART decision tree, wherein the minimum sample division number of the CART decision tree is set to be 50, the minimum purity threshold of the node is set to be 0.85, and the pruning parameter is determined to be 0.02 through cross verification to obtain a maintenance priority evaluation result of the platform door component;
The method comprises the steps of establishing an optimized objective function of time cost and risk cost based on a maintenance priority evaluation result, obtaining a maintenance time window through calculation of a genetic algorithm, selecting a maintenance scheme from a maintenance process library based on fault type probability distribution, determining maintenance item combinations through association rule mining historical maintenance records, and carrying out simulation verification on the maintenance time window and the maintenance item combinations of the maintenance scheme by adopting a discrete event simulation method to obtain an intelligent platform door maintenance scheme.
Establishing an optimized objective function of time cost and risk cost based on the maintenance priority evaluation result, and calculating to obtain a maintenance time window through a genetic algorithm comprises the following steps:
Constructing a time cost function and a risk cost function based on a maintenance priority evaluation result, wherein the time cost function comprises a weighted sum of maintenance operation time cost, operation delay cost and equipment disabling cost, and the risk cost function comprises a product of failure occurrence probability, failure loss cost and an exponential decay function of maintenance time margin;
Chromosome coding is carried out on maintenance starting time points by adopting a real number coding mode, population scale is set to be one hundred, evolution algebra is set to be two hundred generations, individual selection probability is calculated by a roulette selection method based on grades, and the individual selection probability is calculated by the ratio of the current individual ranking to the total ranking number;
And reserving five percent of individuals with the highest fitness in each generation of population to directly enter the next generation, and periodically adopting a simulated annealing algorithm to perform local search optimization on the five percent of individuals with the highest fitness to obtain an optimal maintenance time window of the platform door.
In a second aspect of an embodiment of the present invention,
Providing a rail transit platform door multi-layer architecture intelligent operation maintenance and predictive maintenance system, comprising:
The system comprises a first unit, a second unit, a third unit, a fourth unit, a fifth unit, a sixth unit and a seventh unit, wherein the first unit is used for establishing a multi-layer data acquisition system of the rail transit platform door, acquiring platform door operation parameters including vibration frequency, driving current, component temperature and door opening and closing displacement through a vibration sensor, a temperature sensor, a displacement sensor and a current sensor which are arranged on a platform door bottom mechanism, a driving mechanism and a door body structure, transmitting the operation parameters to an edge computing unit through a field bus for data preprocessing, and generating a platform door health state feature vector;
the second unit is used for inputting the platform door health state feature vector into a deep neural network comprising a convolution layer, a pooling layer and a full-connection layer, training the deep neural network by combining historical maintenance data through time sequence correlation analysis and multi-scale feature extraction on the feature vector, and constructing a platform door fault prediction model based on deep learning;
And the third unit is used for dynamically evaluating the maintenance priority of the platform door part by adopting a decision tree algorithm based on the output result of the platform door fault prediction model based on deep learning, generating an intelligent maintenance scheme comprising maintenance time, maintenance items and maintenance resources, pushing maintenance instructions to on-site maintenance personnel through the mobile terminal, feeding back new data acquired in the maintenance process to the deep learning model for on-line optimization updating, and realizing the self-adaptive adjustment of the platform door maintenance strategy.
In a third aspect of an embodiment of the present invention,
There is provided an electronic device including:
A processor;
a memory for storing processor-executable instructions;
wherein the processor is configured to invoke the instructions stored in the memory to perform the method described previously.
In a fourth aspect of an embodiment of the present invention,
There is provided a computer readable storage medium having stored thereon computer program instructions which, when executed by a processor, implement the method as described above.
The beneficial effects of the application are as follows:
1. Multi-layer data acquisition and pretreatment:
by arranging a plurality of sensors at each key part of the platform door, a comprehensive data acquisition system is established, and the running state of the platform door can be monitored in real time. The edge computing unit is combined to perform data preprocessing, so that a health state feature vector is generated, high-quality input data is provided for subsequent fault prediction, and the accuracy and the instantaneity of the fault prediction are improved.
2. Deep learning fault prediction model:
and carrying out time sequence correlation analysis and multi-scale feature extraction on the platform door health state feature vector by adopting a deep neural network comprising a convolution layer, a pooling layer and a full-connection layer. Training is carried out by combining historical maintenance data, and a fault prediction model based on deep learning is constructed. The method can effectively capture the complex mode of the running state of the platform door, improves the accuracy and predictability of fault prediction, and provides reliable decision basis for preventive maintenance.
3. Intelligent maintenance scheme generation and adaptive optimization:
Based on the output of the fault prediction model, the maintenance priority is dynamically evaluated by using a decision tree algorithm, and an intelligent maintenance scheme is generated. And the maintenance instructions are pushed to maintenance personnel through the mobile terminal, so that the maintenance efficiency is improved. Meanwhile, new data acquired in the maintenance process are fed back to the deep learning model, so that online optimization updating of the model is realized, and the maintenance strategy can be adjusted in a self-adaptive mode. The closed-loop optimization mechanism ensures continuous improvement of maintenance strategies and improves the overall reliability and operation efficiency of the platform door system.
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 embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention, and it is apparent that the described embodiments are only some embodiments of the present invention, 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 technical scheme of the invention is described in detail below by specific examples. The following embodiments may be combined with each other, and some embodiments may not be repeated for the same or similar concepts or processes.
Fig. 1 is a flow chart of a method for intelligent operation, maintenance and predictive maintenance of a multi-layer structure of a rail transit platform door according to an embodiment of the invention, as shown in fig. 1, the method includes:
S101, a multi-layer data acquisition system of a rail transit platform door is established, platform door operation parameters are acquired through vibration sensors, temperature sensors, displacement sensors and current sensors which are arranged on a platform door bottom mechanism, a driving mechanism and a door body structure, the operation parameters comprise vibration frequency, driving current, component temperature and door opening and closing displacement, and the operation parameters are transmitted to an edge computing unit through a field bus to perform data preprocessing, so that a platform door health state feature vector is generated;
s102, inputting the platform door health state feature vector into a deep neural network comprising a convolution layer, a pooling layer and a full-connection layer, training the deep neural network by combining historical maintenance data through time sequence correlation analysis and multi-scale feature extraction on the feature vector, and constructing a platform door fault prediction model based on deep learning;
S103, dynamically evaluating the maintenance priority of the platform door part by adopting a decision tree algorithm based on the output result of the platform door fault prediction model based on deep learning, generating an intelligent maintenance scheme comprising maintenance time, maintenance items and maintenance resources, pushing maintenance instructions to on-site maintenance personnel by a mobile terminal, feeding back new data acquired in the maintenance process to the deep learning model for on-line optimization updating, and realizing self-adaptive adjustment of the platform door maintenance strategy.
The specific implementation mode of the rail transit platform door multi-layer structure intelligent operation maintenance and predictive maintenance method is as follows:
Firstly, a multi-layer data acquisition system of the rail transit platform door is established. Vibration sensor, temperature sensor, displacement sensor and current sensor are respectively installed on the platform door bottom mechanism, driving mechanism and door body structure. The vibration sensor collects vibration frequency data of all parts of the platform door, and the range is 0-1000Hz. The temperature sensor collects the working temperature of key components such as a driving motor, a controller and the like, and the range is-20 ℃ to 80 ℃. The displacement sensor collects displacement data in the process of opening and closing the platform door, and the precision is 0.1mm. The current sensor collects the working current of the driving motor, and the range is 0-20A. The sensors transmit the acquired data to an edge computing unit on site in real time through an RS485 bus.
The edge computing unit performs preprocessing on the received original data, including denoising, filtering, normalization and other operations. And extracting key characteristics such as statistics of mean value, variance, peak value and the like of vibration frequency, maximum value, average value and the like of temperature, maximum value of displacement, change rate and the like, effective value of current, harmonic content and the like. These features are combined to form a health status feature vector for the platform door, and a piece of feature vector data is generated every 5 minutes.
Next, the generated health status feature vector is input into a pre-trained deep neural network. The network contains 3 convolutional layers, 2 pooling layers, and 2 fully-connected layers. The convolution layer uses a convolution check input feature of 3x3 to carry out convolution operation and extract local features. The pooling layer adopts maximum pooling of 2x2, and reduces characteristic dimension. The fully connected layer maps features to fault type spaces. The input to the network is a 72x1 eigenvector (24 hours 144 samples) and the output is a probability distribution of 9 common fault types for the platform gate.
In the network training, 50 ten thousand platform door operation data accumulated in 3 years and corresponding maintenance records are used as a training set. The random gradient descent method is adopted to optimize network parameters, the initial value of learning rate is 0.01, and 50 epochs are descended by 10% each time training is performed. After training 200 epochs, the failure prediction accuracy on the test set reached 92%.
Based on the trained deep learning model, fault prediction is carried out on the real-time input platform door health state feature vector, and the probability of various faults in the future 24 hours is obtained. And inputting the prediction result into a decision tree algorithm, and quantitatively evaluating the maintenance priority of each component by taking factors such as fault probability, fault influence degree, maintenance difficulty and the like into consideration. The evaluation results were scores of 1-10, with 10 indicating the highest priority.
And automatically generating an intelligent maintenance scheme by the system according to the maintenance priority output by the decision tree. The scheme includes specific maintenance time schedule, maintenance project list and required maintenance resources. For example, for a drive motor failure of priority 9, the system is scheduled for preventive maintenance within 24 hours, maintenance items including replacement of bearings, cleaning of coils, etc., and required resources include 2 servicemen, 1 set of special tools, etc.
The generated maintenance scheme is pushed to field maintenance personnel through the mobile APP. The APP interface visually displays the priority, specific content, required tools and other information of the maintenance task. Maintenance personnel can feed back information such as fault phenomenon, maintenance operation and the like in the maintenance process through the APP. These newly acquired data will be used to further optimize the deep learning model, such as adding new fault samples, adjusting model parameters, etc., thereby continually improving the prediction accuracy of the model.
The beneficial effects of the method are mainly as follows:
1. The intelligent monitoring and predictive maintenance of the whole life cycle of the platform door are realized, and the reliability and the operation efficiency of the equipment are greatly improved. The key parameters are monitored in real time through the multiple sensors, and the fault prediction is carried out by combining a deep learning algorithm, so that potential faults can be found in time, and serious consequences caused by the operation of equipment with diseases are avoided.
2. The maintenance cost and the manpower requirement of the platform door are remarkably reduced. The traditional periodic maintenance mode is changed to predictive maintenance based on equipment status, reducing unnecessary maintenance work. The pushing of the intelligent maintenance scheme greatly improves the maintenance efficiency and reduces the dependence on high-skill maintenance personnel.
3. A set of sustainable optimized intelligent operation and maintenance system is constructed. By continuously collecting operation data and maintenance feedback, the deep learning model is continuously optimized, so that the fault prediction capability is continuously improved. The self-adaptive adjustment of the system ensures that the maintenance strategy always maintains the optimal state, and provides powerful support for long-term safe and efficient operation of rail transit.
In an alternative embodiment, the operating parameters are transmitted to the edge computing unit via the fieldbus for data preprocessing, and generating the platform door health status feature vector includes:
Collecting platform door operation parameters through a vibration sensor arranged on a platform door bottom mechanism, a current sensor arranged on a platform door driving mechanism, a temperature sensor arranged on a platform door body structure and a displacement sensor, wherein the sampling frequency of the vibration sensor is 1000Hz, the sampling frequency of the temperature sensor is 1Hz, the sampling frequency of the displacement sensor is 100Hz, and the sampling frequency of the current sensor is 500Hz;
The edge computing unit adopts a dual-core ARM processor, performs 3-layer decomposition on the operation parameters by using a db4 wavelet basis function, filters high-frequency noise by a soft threshold method, identifies and eliminates abnormal values based on a3 sigma criterion, and uniformly aligns the operation parameters to a time axis of 100Hz by adopting a piecewise linear interpolation method;
Extracting speed curve characteristics, acceleration characteristics and displacement characteristics in the process of opening and closing a platform door as working condition characteristics, calculating an envelope spectrum of a vibration signal and a harmonic ratio of a current signal through a sliding time window as performance degradation characteristics, establishing environment influence characteristics by combining correlation analysis of a temperature change trend and vibration response, combining the working condition characteristics, the performance degradation characteristics and the environment influence characteristics to form 145-dimensional characteristic vectors, performing dimension reduction processing on the 145-dimensional characteristic vectors by adopting a principal component analysis method, and selecting principal components with accumulated contribution rate reaching 95% to generate the platform door health state characteristic vectors.
The specific implementation mode is as follows:
First, various sensors are installed in the platform door system to collect operating parameters. The vibration sensor is arranged on the platform door bottom mechanism, the sampling frequency is 1000Hz, and the vibration sensor is used for monitoring the vibration condition of the door body during operation. A current sensor is arranged on the platform door driving mechanism, the sampling frequency is 500Hz, and the current sensor is used for monitoring the current change of the driving motor. The platform door body structure is provided with a temperature sensor, the sampling frequency is 1Hz, and the temperature sensor is used for monitoring the temperature change of the door body. Meanwhile, a displacement sensor is arranged on the door body, the sampling frequency is 100Hz, and the displacement sensor is used for monitoring the position change in the door body opening and closing process.
The sensor is divided into three groups, wherein the mechanical group comprises a vibration sensor, the driving group comprises a current sensor, and the door body group comprises a temperature sensor and a displacement sensor. Each group is provided with an independent CAN controller and an independent CAN transceiver, and the anti-interference capability is improved by adopting a differential signal transmission mode. Using time division multiplexing, the operating parameters acquired by each sensor are transmitted over the Fieldbus to the edge calculation unit at a transmission rate of 250 kbps.
The edge computing unit adopts a dual-core ARM processor and has stronger real-time data processing capability. First, the received operating parameters are preprocessed. And 3 layers of wavelet decomposition is carried out on the operation parameters by using the db4 wavelet basis function, and different frequency components are separated. And then thresholding the high frequency component by adopting a soft thresholding method to filter out high frequency noise. Outliers are identified based on the 3σ criterion, and data points that exceed the mean by plus or minus 3 standard deviations are marked as outliers and removed.
Next, the operating parameters of the different sampling frequencies are aligned uniformly to the time axis of 100Hz using a piecewise linear interpolation method. For example, for a 1000Hz sampled vibration signal, 1 point is taken for every 10 data points, and for a 1Hz sampled temperature signal, 99 data points are linearly interpolated between the adjacent two sample points. Thus, different parameters can be aligned in time, and subsequent analysis is facilitated.
And then extracting time domain and frequency domain characteristics from the preprocessed operation parameters. The time domain features comprise a mean value used for representing the overall level of the signal, a standard deviation used for representing the fluctuation degree of the signal, a peak value factor used for representing the impact property of the signal and a margin factor used for representing the stability of the signal. The frequency domain features comprise main frequency components, namely frequency components with the largest amplitude in the signal spectrum, frequency band energy distribution, frequency spectrum division into a plurality of frequency bands, and calculation of energy duty ratio of each frequency band. And meanwhile, calculating correlation coefficients among different operation parameters, and analyzing correlations among the parameters.
And extracting characteristic parameters in the process of opening and closing the platform door as working condition characteristics. speed curve characteristics include maximum speed, average speed, etc., acceleration characteristics include maximum acceleration, acceleration time, etc., and displacement characteristics include door opening time, door closing time, etc. By setting a sliding time window, such as 5 seconds, an envelope spectrum is calculated on the vibration signals in the window, and characteristic frequencies and amplitudes in the envelope spectrum are extracted. The ratio of each subharmonic to the fundamental wave is calculated for the current signal as a harmonic ratio feature. These characteristics may reflect performance degradation of the device.
And combining the correlation analysis of the temperature change trend and the vibration response to establish the environmental impact characteristics. For example, a correlation coefficient between the temperature change rate and the root mean square value of vibration is calculated to reflect the degree of influence of temperature on vibration. The operating mode features, performance degradation features, and environmental impact features are combined to form 145-dimensional feature vectors.
And finally, reducing the dimension of the 145-dimensional feature vector by adopting a principal component analysis method. And calculating a covariance matrix of the eigenvectors, and solving the eigenvalues and the eigenvectors. And sorting the characteristic values from large to small, and selecting the first few main components with the accumulated contribution rate reaching 95%. For example, when the cumulative contribution rate of the first 10 principal components reaches 95%, the 145-dimensional feature vector is reduced to 10 dimensions, and the platform door health state feature vector is generated.
The method has the beneficial effects that:
1. through collaborative monitoring and grouping transmission of a plurality of sensors, the platform door operation parameters are comprehensively acquired, and the data integrity and reliability are improved.
2. The edge computing unit is adopted to conduct data preprocessing and feature extraction, so that data transmission quantity is reduced, and instantaneity and expandability of the system are improved.
3. By combining the time domain frequency domain characteristics, the working condition characteristics, the performance degradation characteristics and the environmental impact characteristics, a multidimensional health state characteristic vector is constructed, the health state of the platform door is comprehensively reflected, and an effective basis is provided for subsequent fault diagnosis and predictive maintenance.
In an alternative embodiment, the deep-learned platform door fault prediction model includes a temporal feature branch, a spatial feature branch, and an attention feature branch;
The time sequence feature branch comprises three one-dimensional convolution layers connected in series, the convolution kernel sizes of the three one-dimensional convolution layers are 3, 5 and 7 respectively, the output of each one-dimensional convolution layer is connected with a largest pooling layer, the pooling window size of the largest pooling layer is 2, and the output of the time sequence feature branch is used for representing the time sequence feature of the health state feature vector;
The space characteristic branch comprises four two-dimensional convolution layers connected in series, the four two-dimensional convolution layers adopt 4 multiplied by 4 convolution kernels, the channel numbers of the four two-dimensional convolution layers are 32, 64, 128 and 256 in sequence, and the output of the space characteristic branch is used for representing the space characteristic of the health state characteristic vector;
The attention characteristic branch comprises a self-attention calculating unit, the self-attention calculating unit calculates association weights among the dimensions of the health state characteristic vector, the association weights are normalized through a softmax function to obtain normalized weights, and the output of the attention characteristic branch is used for representing the attention characteristic of the health state characteristic vector;
And carrying out feature fusion on the time sequence feature, the space feature and the attention feature to obtain a fusion feature, wherein the fusion feature is used for predicting the failure of the platform door.
The deep learning-based platform door fault prediction model comprises a time sequence characteristic branch, a space characteristic branch and a attention characteristic branch. First, a characteristic vector of the health state of the platform door is obtained, wherein the characteristic vector comprises data acquired by a plurality of sensors in a certain time period.
The time sequence feature branch is used for extracting time sequence features of the health state feature vector. The branch comprises three one-dimensional convolution layers connected in series, and the convolution kernel sizes are 3, 5 and 7 respectively. The output of each one-dimensional convolution layer is connected with a largest pooling layer, and the pooling window size is 2. In particular, the first one-dimensional convolution layer performs a convolution operation using 3 one-dimensional convolution check input features of size 3, and then downsampled by the max pooling layer. The second one-dimensional convolution layer convolves with 5 one-dimensional convolution kernels of size 5, and then undergoes maximum pooling. The third one-dimensional convolution layer convolves with 7 one-dimensional convolution kernels of size 7, and finally undergoes maximum pooling. Such a multi-scale convolution kernel design may capture characteristic patterns for different time spans.
The spatial feature branch is used for extracting the spatial features of the health state feature vector. The branch comprises four two-dimensional convolution layers connected in series, 4×4 convolution kernels are adopted, and the channel numbers are 32, 64, 128 and 256 in sequence. In particular, the first two-dimensional convolution layer convolves the input signature using 32 4 x4 convolution kernels to output a 32-channel signature. The second layer uses 64 4 x4 convolution kernels, outputting a 64-channel signature. The third layer uses 128 4 x4 convolution kernels, outputting a 128-channel signature. The fourth layer uses 256 4 x4 convolution kernels, which ultimately output a 256-channel signature. This layer-by-layer increasing channel number design can progressively extract higher-level spatial features.
The attention feature branches are used to calculate the associations between the dimensions of the health state feature vectors. The branch comprises a self-attention calculating unit which first calculates the association weights between the dimensions of the feature vectors. In particular implementations, feature vectors may be converted into a query matrix, a key matrix, and a value matrix, and the attention score calculated by matrix multiplication. And then normalizing the associated weights by using a softmax function to obtain normalized attention weights. And finally multiplying the attention weight by a value matrix to obtain the attention characteristic.
And carrying out feature fusion on the time sequence features, the space features and the attention features to obtain fusion features. In specific implementation, three features can be flattened into a one-dimensional vector, then concatates are spliced together, and then feature fusion is carried out through a full-connection layer, so that final fusion feature representation is obtained.
And predicting the failure of the platform door based on the fusion characteristics. The fusion features may be input into a classifier (e.g., a softmax classifier) that predicts the type of failure and probability of failure of the platform door. The fusion characteristics may also be input into a regression model to predict the remaining useful life of the platform door.
In a specific application, sensor data for the normal operation of the platform door and various fault conditions can be collected to construct a training data set. The deep learning model is trained by using the marked data, so that the deep learning model can learn the characteristic modes under different states. During actual prediction, the health state feature vector of the current platform door is input, and the model can output a fault prediction result.
The technical scheme has the following beneficial effects:
1. The multi-branch feature extraction fully mines time sequence, space and attention information of the health state features, and feature representation capability is improved.
2. The multi-scale convolution kernel design of the time sequence characteristic branches can capture modes of different time spans, and modeling capacity of time sequence dependence is enhanced.
3. The attention mechanism can adaptively pay attention to important features, so that the interpretability and the prediction accuracy of the model are improved.
In an alternative embodiment, the deep learning based platform door fault prediction model comprises:
Acquiring a platform door health state feature vector, performing recursion graph analysis on the platform door health state feature vector, mapping the platform door health state feature vector to a high-dimensional phase space to obtain feature track points by setting a phase space reconstruction parameter with an embedding dimension of 5 and a time delay of 2, and calculating Euclidean distances between the feature track points to construct a recursion matrix;
Decomposing the platform door health state feature vector into a plurality of integral mode functions by a local mode decomposition method, and calculating instantaneous frequency and instantaneous amplitude to the integral mode functions to obtain microscale features;
The platform door health state feature vector is decomposed into a trend item, a period item and a residual item through a variation mode decomposition algorithm to obtain a macro-scale feature, and the micro-scale feature, the mesoscale feature and the macro-scale feature are input into a deep neural network for training to obtain a platform door fault prediction model.
The specific implementation mode is as follows:
Firstly, a platform door health state characteristic vector is obtained, wherein the characteristic vector comprises multidimensional data such as door body position, door body speed, door body acceleration, motor current, air pressure and the like. And carrying out recursive graph analysis on the feature vector, setting a phase space reconstruction parameter with an embedding dimension of 5 and a time delay of 2, and mapping the feature vector to a 5-dimensional phase space to obtain a feature track point. And calculating Euclidean distance between the characteristic track points, and constructing a recursion matrix. For example, for a feature vector [10, 2,0.5, 1.2, 80] at a certain moment, a locus point [10, 2,0.5, 1.2, 80; 2,0.5, 1.2, 80, 11; 0.5, 1.2, 80, 11, 2.1 ] is obtained after phase space reconstruction. And calculating Euclidean distance between the track points to obtain a recursion matrix.
The eigenvectors are then time-frequency decomposed using complex Morlet wavelets. 5 different scales are selected for decomposition, and 5 groups of coefficients are obtained. The cross-correlation coefficients between the 5 sets of coefficients are calculated to obtain a 5x5 cross-correlation matrix. The feature vector is then decomposed into 5 integral mode functions using a local mode decomposition method. And calculating the instantaneous frequency and the instantaneous amplitude value of each integral modal function to obtain microscale characteristics. For example, for a certain integral modal function, the instantaneous frequency sequence [10, 12, 11, 13.] Hz, the instantaneous amplitude sequence [0.5, 0.6,0.4, 0.7.].
And then constructing a time-frequency energy spectrum by combining the feature vector with the Hilbert transform through empirical mode decomposition. And calculating a marginal spectrum and an instantaneous frequency spectrum to obtain mesoscale characteristics. For example, a marginal spectrum [100, 80, 60, 40, 20] and an instantaneous frequency spectrum [15, 14, 13, 12,11] Hz are calculated. And then decomposing the feature vector into a trend term, 3 period terms and a residual term by using a variation modal decomposition algorithm to obtain the macro-scale feature. For example, the trend term is [10,11, 12, 13, ], the period term 1 is [1, 0, -1, 0, ], the period term 2 is [0.5, 0, -0.5,0, ], the period term 3 is [0.2, 0, -0.2, 0, ], the residual term is [0.1, -0.1, 0.2, -0.2, ].
And finally, taking the microscale features, mesoscale features and macroscale features as inputs, and sending the inputs into a 5-layer deep neural network for training. The network structure is that 64 nodes of an input layer and 32, 16 and 8 nodes of 3 hidden layers are respectively arranged, and 2 nodes of an output layer respectively represent the fault probability and the residual service life. The training was 500 rounds with an Adam optimizer with a learning rate of 0.001. Finally, a platform door fault prediction model is obtained, and the occurrence probability of faults and the residual service life can be predicted.
The method has the beneficial effects that:
1) By multi-scale feature extraction, the health state of the platform door system is comprehensively described, and the accuracy of fault prediction is improved.
2) By adopting methods such as a recursion diagram, time-frequency analysis and the like, nonlinear dynamic characteristics of the platform door system are effectively extracted, and generalization capability of the model is enhanced.
3) By combining with a deep learning method, the end-to-end fault prediction is realized, and the method has stronger adaptability and robustness compared with the traditional method.
In an alternative embodiment, based on the output result of the deep-learning platform door fault prediction model, dynamically evaluating the maintenance priority of the platform door component by adopting a decision tree algorithm, and generating an intelligent maintenance scheme comprising maintenance time, maintenance project and maintenance resource comprises:
Constructing a maintenance priority evaluation index based on a residual life predicted value and a fault type probability distribution output by a deep learning model, mapping the residual life predicted value in a segmented mode according to a time interval to obtain a health state index, calculating to obtain a fault risk index based on the product of the fault type probability distribution and a fault type hazard degree weight matrix, and calculating to obtain an operation influence index by a hierarchical analysis method in combination with the flow rate of a station door guest, the position of a station and the state of standby resources;
Inputting the health state index, the fault risk index and the operation influence index into a CART decision tree, wherein the minimum sample division number of the CART decision tree is set to be 50, the minimum purity threshold of the node is set to be 0.85, and the pruning parameter is determined to be 0.02 through cross verification to obtain a maintenance priority evaluation result of the platform door component;
The method comprises the steps of establishing an optimized objective function of time cost and risk cost based on a maintenance priority evaluation result, obtaining a maintenance time window through calculation of a genetic algorithm, selecting a maintenance scheme from a maintenance process library based on fault type probability distribution, determining maintenance item combinations through association rule mining historical maintenance records, and carrying out simulation verification on the maintenance time window and the maintenance item combinations of the maintenance scheme by adopting a discrete event simulation method to obtain an intelligent platform door maintenance scheme.
The invention provides a platform door intelligent maintenance scheme generation method based on deep learning and decision tree. The method comprises the steps of firstly predicting residual life and probability distribution of fault types of a platform door by using a deep learning model, then constructing maintenance priority evaluation indexes based on prediction results, evaluating maintenance priority of a platform door part by using a decision tree algorithm, and finally generating an intelligent maintenance scheme comprising maintenance time, projects and resources. The specific implementation steps are as follows:
first, a maintenance priority evaluation index is constructed based on the residual life prediction value and the fault type probability distribution output by the deep learning model. The method specifically comprises the following steps:
(1) And carrying out sectional mapping on the residual life predicted value according to the time interval to obtain the health state index. For example, the remaining life prediction value may be divided into 5 intervals: [0,30 days), [30 days, 90 days), [90 days, 180 days), [180 days, 365 days), [365 days, + -infinity), respectively correspond to health status indexes of 1,2,3, 4, 5.
(2) And calculating to obtain a fault risk index based on the product of the fault type probability distribution and the fault type hazard degree weight matrix. Firstly, a fault type hazard degree weight matrix is established, for example, weights of 0.3, 0.2, 0.1, 0.3 and 0.1 are respectively assigned to 5 common fault types (door jam, door failure, door leakage, door lock failure and emergency unlocking device failure) of the platform door. And multiplying the probability distribution of the fault type by the weight matrix to obtain a fault risk index.
(3) And calculating by a hierarchical analysis method according to the traffic of the platform door, the site position and the standby resource condition to obtain the operation influence index. Firstly, constructing a hierarchical structure model, taking the passenger flow volume, the site position and the standby resource condition as three primary indexes, and setting a plurality of secondary indexes. And then determining the weight of each index through expert scoring, and finally calculating to obtain the comprehensive operation influence index.
Next, the health status index, the fault risk index and the operation influence index are input into the CART decision tree, and maintenance priority evaluation is performed on the platform door component. The minimum sample division number of the CART decision tree is set to 50, the minimum purity threshold of the node is set to 0.85, and the pruning parameter is determined to be 0.02 through 10-fold cross validation. The decision tree output results in a maintenance priority of the platform door components, which can be classified into three classes of "high", "medium" and "low".
And then, based on the maintenance priority evaluation result, establishing an optimized objective function of the time cost and the risk cost, and calculating through a genetic algorithm to obtain a maintenance time window. The objective function may be expressed as a weighted sum of a time cost that takes into account the impact of maintenance time on operation and a risk cost that takes into account the loss of failure that may be caused by delayed maintenance. The population size of the genetic algorithm is set to be 100, the crossover probability is 0.8, the variation probability is 0.1, and the iteration number is 1000.
Next, a maintenance plan is selected from the repair process library based on the fault type probability distribution, and a maintenance project combination is determined by mining historical repair records through association rules. Firstly, selecting a corresponding maintenance scheme from a maintenance process library according to fault type probability distribution. And then carrying out association rule mining on the historical maintenance records by using an Apriori algorithm, setting a support degree threshold value to be 0.02, and setting a confidence degree threshold value to be 0.6 to obtain association relations among maintenance items, thereby determining reasonable maintenance item combinations.
And finally, performing simulation verification on a maintenance time window and maintenance item combination of the maintenance scheme by adopting a discrete event simulation method to obtain the intelligent maintenance scheme of the platform door. In the simulation process, the resource constraints of maintenance personnel, equipment, materials and the like are considered, and random events in the maintenance process, such as maintenance time fluctuation, maintenance quality fluctuation and the like, are simulated. And comparing the effects of different maintenance schemes through multiple simulation experiments, and selecting an optimal intelligent maintenance scheme.
The beneficial effects of the method are mainly as follows:
(1) The scientificity and the accuracy of maintenance decisions are improved. Residual life prediction and fault type probability distribution prediction are carried out on the platform door through a deep learning model, and maintenance priority evaluation is carried out on the platform door part by using a decision tree algorithm in combination with multidimensional evaluation indexes, so that the scientificity and accuracy of maintenance decisions are effectively improved.
(2) The optimal configuration of maintenance resources is realized. Based on the maintenance priority evaluation result, a maintenance time window is optimized through a genetic algorithm, and reasonable maintenance item combinations are determined through association rule mining, so that the optimal configuration of maintenance resources is realized, and the maintenance efficiency is improved.
(3) The reliability and adaptability of the maintenance scheme is enhanced. The discrete event simulation method is adopted to verify the maintenance scheme, so that the random factors in the actual maintenance process are considered, the reliability and the adaptability of the maintenance scheme are enhanced, and the overall operation stability of the platform door system is improved.
In an alternative embodiment, the maintenance priority evaluation result establishes an optimized objective function of time cost and risk cost, and calculating the maintenance time window through a genetic algorithm includes:
Constructing a time cost function and a risk cost function based on a maintenance priority evaluation result, wherein the time cost function comprises a weighted sum of maintenance operation time cost, operation delay cost and equipment disabling cost, and the risk cost function comprises a product of failure occurrence probability, failure loss cost and an exponential decay function of maintenance time margin;
Chromosome coding is carried out on maintenance starting time points by adopting a real number coding mode, population scale is set to be one hundred, evolution algebra is set to be two hundred generations, individual selection probability is calculated by a roulette selection method based on grades, and the individual selection probability is calculated by the ratio of the current individual ranking to the total ranking number;
And reserving five percent of individuals with the highest fitness in each generation of population to directly enter the next generation, and periodically adopting a simulated annealing algorithm to perform local search optimization on the five percent of individuals with the highest fitness to obtain an optimal maintenance time window of the platform door.
The specific implementation mode is as follows:
first, a time cost function and a risk cost function are constructed based on the maintenance priority evaluation result. The time cost function includes three components, maintenance operation time cost, operation delay cost, and equipment outage cost. The three costs are weighted and summed according to different weights to obtain a total time cost. The maintenance operation time cost mainly considers the labor cost and the material cost, the operation delay cost mainly considers the economic loss caused by train delay caused by maintenance, and the equipment shutdown cost considers the opportunity cost of equipment shutdown. The risk cost function is multiplied by three factors, namely the probability of occurrence of a fault, the cost of loss of fault, and an exponential decay function related to the maintenance time margin. The fault occurrence probability can be obtained through historical data analysis, the fault loss cost is estimated according to the severity of different fault types, and an exponential decay function of the maintenance time margin is used for representing the risk increase caused by delay maintenance.
Next, a genetic algorithm is used to optimize the maintenance time window. Firstly, chromosome coding is carried out, and a real number coding mode is adopted to code the maintenance starting time point. For example, if a certain equipment maintenance window is 0-24 hours, the chromosome may be encoded as a real number between 0 and 24. The population size is set to be 100, and the evolution algebra is set to be 200 generations. In the selection operation, a rank-based roulette selection method is employed. The method comprises the steps of firstly sequencing individuals in a population according to fitness values from large to small, and then calculating the selection probability of each individual. The probability of selection is equal to the ratio of the current individual rank to the total rank number. For example, if a rank name is 10 th, there are 100 total individuals, the selection probability is 10/100=0.1.
After the selection is completed, the selected individuals are subjected to cross operation. Here, adaptive arithmetic crossover operations are employed, with crossover probabilities decreasing linearly with increasing algebra. For example, the initial crossover probability may be set to 0.9 and the termination crossover probability to 0.6, then the crossover probability for the t-th generation may be expressed as 0.9-0.3t/200. In the crossing process, two parent individuals are randomly selected, a random number alpha between 0 and 1 is generated, and if alpha is smaller than the crossing probability, the crossing operation is carried out. The crossing mode is arithmetic crossing, namely the gene values of the two child individuals are respectively weighted average of the corresponding gene values of the two parent individuals.
After the crossing is completed, the individual is subjected to mutation operation. Here, gaussian mutation operation is adopted, and mutation probability and mutation amplitude are exponentially attenuated with the increment of evolution algebra. For example, the initial variation probability may be set to 0.1 and the termination variation probability to 0.001, and the variation probability of the t-th generation may be expressed as 0.1X0.99. The variation amplitude is also exponentially decayed, the initial value can be set to 1, and the termination value can be set to 0.1. In the mutation process, each gene position is judged, and if the random number is smaller than the mutation probability, mutation is carried out. The variation mode is to superimpose a Gaussian random number with the mean value of 0 and the standard deviation of the current variation amplitude on the basis of the original gene value.
In order to improve the convergence performance of the algorithm, an elite retention strategy is adopted, namely 5% of individuals with the highest fitness in each generation of population are retained to directly enter the next generation. Meanwhile, in order to avoid sinking into local optimum, the elite individuals are periodically subjected to local search optimization by adopting a simulated annealing algorithm. Specifically, simulated annealing optimization is carried out on elite individuals every 10 generations. In the simulated annealing process, a neighborhood solution is generated by taking the current solution as a center, if the neighborhood solution is better than the current solution, the neighborhood solution is accepted, otherwise, a worse solution is accepted with a certain probability. The initial temperature was set to 100, the final temperature was set to 1, and the cooling coefficient was 0.95.
By the steps, the optimal maintenance time window of the platform door can be obtained. For example, suppose a platform door system contains 10 key components, each of which has a maintenance time window of 0-24 hours. After optimization, the optimal maintenance time window may be [2.5, 8.3, 12.7, 15.9, 18.2, 20.1, 21.8, 22.9, 23.5, 23.8]. This means that the optimal maintenance start times for the 10 key components are 2:30 a.m., 8:18 a.m., 12:42 a.m., 3:54 a.m., 6:12 a.m., 8:06 a.m., 9:48 a.m., 10:54 a.m., 11:30 a.m., and 11:48 a.m., respectively.
The beneficial effects of the method are mainly as follows:
1. By constructing an optimized objective function of time cost and risk cost, maintenance cost and potential risk are comprehensively considered, good balance between cost and safety can be achieved, and scientificity and rationality of maintenance decision are improved.
2. The genetic algorithm is adopted to carry out optimization solution, and the adaptive cross, gaussian variation and other improvement strategies are combined, so that the local optimum can be effectively avoided, the global searching capacity and the convergence speed of the algorithm are improved, and a better maintenance time window scheme is obtained.
3. Elite retention strategies and simulated annealing local search are introduced, so that the development capability of optimal solutions is enhanced while population diversity is ensured, the quality of an optimization result is further improved, and the obtained maintenance time window scheme is more reliable and efficient.
Fig. 2 is a schematic structural diagram of an intelligent operation and maintenance and predictive maintenance system for a rail transit platform door multi-layer architecture according to an embodiment of the present invention, as shown in fig. 2, the system includes:
The system comprises a first unit, a second unit, a third unit, a fourth unit, a fifth unit, a sixth unit and a seventh unit, wherein the first unit is used for establishing a multi-layer data acquisition system of the rail transit platform door, acquiring platform door operation parameters including vibration frequency, driving current, component temperature and door opening and closing displacement through a vibration sensor, a temperature sensor, a displacement sensor and a current sensor which are arranged on a platform door bottom mechanism, a driving mechanism and a door body structure, transmitting the operation parameters to an edge computing unit through a field bus for data preprocessing, and generating a platform door health state feature vector;
the second unit is used for inputting the platform door health state feature vector into a deep neural network comprising a convolution layer, a pooling layer and a full-connection layer, training the deep neural network by combining historical maintenance data through time sequence correlation analysis and multi-scale feature extraction on the feature vector, and constructing a platform door fault prediction model based on deep learning;
And the third unit is used for dynamically evaluating the maintenance priority of the platform door part by adopting a decision tree algorithm based on the output result of the platform door fault prediction model based on deep learning, generating an intelligent maintenance scheme comprising maintenance time, maintenance items and maintenance resources, pushing maintenance instructions to on-site maintenance personnel through the mobile terminal, feeding back new data acquired in the maintenance process to the deep learning model for on-line optimization updating, and realizing the self-adaptive adjustment of the platform door maintenance strategy.
In a third aspect of an embodiment of the present invention,
There is provided an electronic device including:
A processor;
a memory for storing processor-executable instructions;
wherein the processor is configured to invoke the instructions stored in the memory to perform the method described previously.
In a fourth aspect of an embodiment of the present invention,
There is provided a computer readable storage medium having stored thereon computer program instructions which, when executed by a processor, implement the method as described above.
The present invention may be a method, apparatus, system, and/or computer program product. The computer program product may include a computer readable storage medium having computer readable program instructions embodied thereon for performing various aspects of the present invention.
It should be noted that the above embodiments are merely for illustrating the technical solution of the present invention and not for limiting the same, and although the present invention has been described in detail with reference to the above embodiments, it should be understood by those skilled in the art that the technical solution described in the above embodiments may be modified or some or all of the technical features may be equivalently replaced, and these modifications or substitutions do not make the essence of the corresponding technical solution deviate from the scope of the technical solution of the embodiments of the present invention.