A solenoid valve fault diagnosis method and system based on domain adversarial transfer learning

Abstract

The present invention discloses a solenoid valve fault diagnosis method and system based on domain adversarial transfer learning. The method comprises: building a balanced air cylinder control system test bench, supplementing the locomotive's online operating data with an offline test data set, screening the offline test data as a source domain based on information divergence, mixing the source domain data with the target domain data from actual operation, and using this as the input for the solenoid valve fault diagnosis method; extracting pressure characteristic values ​​in stages based on the periodic characteristics of the brake pressure output curve, and obtaining fault diagnosis features of higher importance through evidence theory; using the training data and the corresponding fault type as input and output data, respectively, training the diagnostic model to obtain a fault classifier model; and inputting the characteristics of the solenoid valve pressure time series to be tested into the fault classifier to obtain the final solenoid valve fault diagnosis result. The present invention can effectively improve the accuracy of solenoid valve fault diagnosis.

Description

Technical Field

[0001] The present invention belongs to the technical field of solenoid valve fault diagnosis in train brake systems, and in particular relates to a solenoid valve fault diagnosis method and system based on domain adversarial transfer learning. Background Art

[0002] The brake system is a multi-module system with numerous working parts and complex operating modes. Failure in any one component can impact the operation of the entire system. Based on the structure and operating principles of distributed brake systems, brake system failures can be categorized into two main layers: signal layer and physical layer. Statistics from fault data recorded during actual train operation show that physical layer failures primarily include valve failures, sensor failures, and cylinder pipe failures, while signal layer failures primarily include processor failures and network transmission failures. Generally speaking, fault diagnosis of key components is of greater importance.

[0003] Fault diagnosis of key brake components requires analysis of the fault type, location, and even severity to inform maintenance decisions, restore normal operation, and improve brake safety. The complex coupling relationships among brake components make it difficult to construct accurate mechanism models, resulting in low accuracy in brake fault diagnosis based on these models. Due to the redundant backup capabilities between distributed brake modules, component fault data collected during locomotive operation is extremely scarce, impacting the effectiveness of data-driven diagnostic methods.

[0004] There are some existing fault diagnosis patents based on transfer learning, such as the Chinese patent publication number CN111860677A, published on October 30, 2020, and titled "A Rolling Bearing Fault Diagnosis Method Based on Transfer Learning Based on Partial Domain Adversarial." This invention proposes a strategy for weighting source domain samples before domain classification, which improves the sample domain adaptability and solves the problem of unsupervised label prediction for rolling bearings in the target domain. The Chinese patent publication number CN115144747A, published on October 4, 2022, is titled "A Multi-condition Motor Fault Diagnosis Method and System Based on Adversarial Transfer Learning." This invention uses a convolutional neural network as a feature extractor to extract features and adds a local domain discriminator to achieve multi-condition motor fault diagnosis. In fault diagnosis applications, existing solenoid valve fault diagnosis methods are overly dependent on actual fault operation data; in fault diagnosis methods, existing adversarial transfer learning fault diagnosis patents mainly focus on further selection of source domain and target domain samples, lack further processing of diagnostic features, and fail to extract features that are highly relevant to solenoid valve fault diagnosis, resulting in insufficient diagnostic accuracy. Summary of the Invention

[0005] To address the aforementioned technical issues in the prior art, the present invention provides a solenoid valve fault diagnosis method and system based on domain adversarial transfer learning. Leveraging the limited requirement for online fault data from transfer learning, the present invention uses the solenoid valve within the brake module as the primary diagnostic target. Using offline test data and collected locomotive pressure time series, the present invention trains a diagnostic model capable of accurately diagnosing brake module faults, thereby improving the accuracy of solenoid valve fault diagnosis.

[0006] The present invention solves the above technical problems with a technical solution: a solenoid valve fault diagnosis method based on domain adversarial transfer learning, comprising the following steps:

[0007] S1: Extract the actual operating solenoid valve pressure time series from the solenoid valves with different fault types in the real train balancing module;

[0008] S2: By adjusting parameters to simulate different fault types of the solenoid valve, the pressure time series of the solenoid valve is collected from the balanced air cylinder control system test bench.

[0009] S3: The train balancing module operation is divided into four stages: rapid charging, stable charging, rapid exhaust, and stable exhaust. The pressure time series of the solenoid valves related to relief and braking are extracted in each stage to obtain the fault diagnosis features of the solenoid valves.

[0010] S4: Extract multiple different time difference and pressure difference features in the rapid filling and exhaust stage and the stable filling and exhaust stage;

[0011] S5: Analyze feature importance based on evidence theory and resolve feature conflicts to obtain features with higher relevance to solenoid valve fault diagnosis.

[0012] S6: Integrate the offline equalization module test bench pressure time series and actual operation fault data to solve the problem of lack of fault labels in actual operation data. The extracted charging and exhaust solenoid valve operation frequency is used as a set of training data.

[0013] S7: Using the training data and the solenoid valve fault type as the input and output of the fault diagnosis model, respectively, the transfer learning model is trained to obtain the fault diagnosis result of the solenoid valve;

[0014] S8: Repeat S4 to S7, back-propagate the domain adversarial classifier parameters and the fault classifier parameters, and obtain the final fault classifier model;

[0015] S9: Obtain the time series of the total air cylinder, balancing air cylinder, and train pipe pressure of the real train balancing module where the solenoid valve to be tested is located. Use the domain adversarial training method to obtain common features in the source domain and target domain, and input them into the fault classifier in S8 to obtain the final solenoid valve fault diagnosis result.

[0016] In step S2 of the present invention, the solenoid valve fault diagnosis types include: normal solenoid valve, leakage fault of air charging solenoid valve, leakage fault of air exhaust solenoid valve, aging of spring, and rubber ring fault.

[0017] In step S4 of the present invention, the pressure characteristic values ​​extracted during the rapid filling and exhausting and stable filling and exhausting stages include time difference and pressure difference.

[0018] Time difference characteristics: Based on the results of the pressure curve stage division, the time spent on the pressure changes of the key components of the train control module, namely the balancing air cylinder, train pipe, pre-control air cylinder and brake cylinder, in the four stages of rapid exhaust stage, stable exhaust stage, rapid filling stage and stable filling stage is used as the time difference characteristics of brake fault diagnosis.

[0019] Pressure difference characteristics: Considering the tracking characteristics between the pressure values ​​of key components, the pressure tracking of different components can reflect the type of brake failure. Therefore, four sets of pressure difference data between components with tracking characteristics are collected: the difference between the equalizing air cylinder pressure target value and the equalizing air cylinder pressure, the difference between the equalizing air cylinder pressure and the train pipe pressure, the difference between the pre-control air cylinder pressure target value and the pre-control air cylinder pressure, and the difference between the pre-control air cylinder pressure and the brake cylinder pressure. Based on the results of the stage division, initial features are extracted from these four sets of pressure differences to construct a multidimensional brake state feature.

[0020] In step S5 of the present invention, the importance of features is analyzed using evidence theory to obtain data features that are sensitive to solenoid valve fault diagnosis.

[0021] A1: Set the feature debate framework Φ = {θ1,θ2,…θ N}, define the mapping function m to represent the acceptance degree of the selected features, and

[0022] m(φ)=0

[0023]

[0024] Where φ indicates that the view is not within the debate framework, m(A) represents the basic probability distribution value of argument A, which describes the degree of acceptance of view A, that is, the evidence. If m(A)>0, A is called a focal element.

[0025] A2: On the debate framework Φ, calculate the trust function Beli(A) and likelihood function P of the selected features Beli (A), and obtain the corresponding confidence interval [Beli(A),P Beli (A)], for argument A

[0026]

[0027]

[0028] In the formula represents the adversarial set of argument A;

[0029] A3: For the selected features, after fusing evidence from other arguments, the final fault diagnosis features are obtained

[0030]

[0031]

[0032] In the formula is a conflict factor, reflecting argument B A and C A conflicting evidence.

[0033] In step S7 of the present invention, the migration from the test bench data to the actual operation data is achieved through the domain adversarial transfer learning method, further comprising:

[0034] B1: Calculate the information divergence of different source domain datasets on the testbed with respect to the target domain, and select the source domain with the smallest information divergence as part of the input of the domain adversarial transfer learning model;

[0035] B2: The feature generator is used to generate key features of samples in the mixed sample set, and the fault classifier identifies faults based on the key features;

[0036] B3: The domain classifier classifies the data in the feature space and determines which domain the data comes from.

[0037] B4: The fault classifier diagnoses the input features and gives the final fault diagnosis result.

[0038] The feature generator and domain classifier designed in the present invention adopt a fully connected neural network, and the fault classifier designed adopts a convolutional neural network.

[0039] A solenoid valve fault diagnosis system based on domain adversarial transfer learning includes a data acquisition module, a balanced air cylinder control system test bench module, a feature extraction module, a domain adversarial module, and a fault diagnosis module;

[0040] The data acquisition module is used to collect actual operating pressure time series using solenoid valves with different fault types in a real train balancing module. During I / O data acquisition, pressure sensors collect pressure time series on the main air cylinder, balancing air cylinder, and train pipes. Data from each channel is synchronized in time, with a sampling rate of up to 250kSa / s, a resolution of 12 bits, an input and output buffer size of 16KB, and a USB 2.0 interface. The pressure time series are transmitted to an industrial computer via TCP / IP.

[0041] In addition, the data acquisition module is used for: When the balancing air cylinder control module of the distributed brake is in normal operation, the pressure data of each key component collected is time series data, and the time interval between adjacent sampling points is 50ms. By amplifying a single filling and exhaust process, a filling and exhaust pressure curve can be obtained. After setting the target pressure value to 600kPa, the pressure of the brake balancing air cylinder will rise immediately, and then the train pipe pressure will also rise. After the target pressure value is readjusted to 420kPa, the pressure of the train pipe and the balancing air cylinder both drop from the stable value of 600kPa to 420kPa.

[0042] The balancing air cylinder control system test bench module is used to: simulate different fault types of the solenoid valve through parameter adjustment, so as to collect and obtain a large number of pressure time series under different fault types. In order to obtain the data set required for component fault diagnosis through the balancing air cylinder control module test, the key pressure change process of the distributed brake balancing air cylinder module in the actual operation scenario can be simulated by continuously repeating the braking process and the relief process. The system component is the core of the accelerated aging test platform, which is used to simulate the function realization and data generation of the balancing air cylinder control system. In the system component, the air compressor provides the air source for the air pipeline and cylinder components. The database of the industrial computer is used to store the monitoring data in the hardware platform uploaded by the data acquisition board, and at the same time provides an operating human-computer interaction interface for issuing brake relief commands to the brake control module;

[0043] The feature extraction module is used to extract common features from the mixed real train equalization module pressure time series and the equalization air cylinder control test bench pressure time series;

[0044] The domain adversarial module is used to: identify which pressure time series set the input feature comes from;

[0045] The fault diagnosis module is used to: use the mixed real train balancing module pressure time series and the balanced air cylinder control test bench pressure time series as input, and the real train balancing module fault type as output to train a convolutional neural network model to obtain a solenoid valve fault diagnosis model;

[0046] The fault diagnosis module is also used to: obtain the pressure time series of the balancing air cylinder and the train pipe in the real train balancing module where the solenoid valve to be tested is located, mix the extracted data and the test bench pressure time series, input the extracted features into the solenoid valve fault diagnosis model, and obtain the fault type of the solenoid valve to be tested.

[0047] The pressure time series acquisition subsystem described in this invention is powered by a DC 110V ± 30% supply. Its external wired network interfaces include those for collecting train communication network data and brake process data. Its external wireless network interfaces include those for uploading log files via the station's local area network and real-time process data via the mobile network. It is used to run algorithms to estimate real-time model parameters, predict future hybrid energy storage system states using prediction algorithms, solve multi-objective optimization problems, and access parameter data from a database.

[0048] Beneficial effects of the present invention: The domain adversarial transfer learning proposed in the present invention uses offline source domain samples to expand the online running data set, and extracts common features through the adversarial analysis between the feature generator and the domain classifier. This adversarial transfer method achieves the migration from offline test data to the target data domain of online running data of key components by generating a set of source domain and target domain data sets that fuse the features of the source domain data and the target domain data. This solves the problem of difficulty in diagnosing component faults caused by insufficient fault sample data collected during the operation of the solenoid valve, and improves the accuracy of the solenoid valve fault classification model when online fault data is limited. Moreover, the fault diagnosis model based on the convolutional neural network uses the effective feature set derived through evidence theory as input, which not only ensures the accuracy of fault diagnosis, but also improves the efficiency of the solenoid valve fault diagnosis model. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 is a flow chart of the present invention;

[0050] Figure 2 This is a schematic diagram of the operation of the solenoid valve of a real train pressure control module provided by an embodiment of the present invention;

[0051] Figure 3 This is a flow chart of the pneumatic control of a train pipe provided by an embodiment of the present invention;

[0052] Figure 4 This is a schematic structural diagram of a test bench for a balanced air cylinder control system provided by an embodiment of the present invention;

[0053] Figure 5 This is a structural diagram of the domain adversarial transfer learning algorithm provided by an embodiment of the present invention;

[0054] Figure 6 It is a system module structure diagram of the present invention. DETAILED DESCRIPTION

[0055] The technical solution of the present invention is described clearly and completely below with reference to the accompanying drawings and specific embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0056] like Figure 1 As shown, the electromagnetic valve fault diagnosis method based on domain adversarial transfer learning of the present invention includes the following steps:

[0057] S1: Solenoid valves with different fault types from real train balancing modules, Figure 2 From the action status of the charging and exhaust solenoid valves, it can be known that the solenoid valve fault diagnosis types include: normal solenoid valve, charging solenoid valve leakage fault, exhaust solenoid valve leakage fault, spring aging, and rubber ring fault. Sampling is done every 50ms to extract the actual operating pressure time series. The pressure time series collected in real time by the pressure sensor can well reflect the operating status of the brake solenoid valve. By amplifying the charging and exhaust process, the charging and exhaust pressure curve can be obtained. After setting the target pressure value to 600kPa, the brake equalizing air cylinder pressure will rise immediately, and then the train pipe pressure will also rise. After the target pressure value is readjusted to 420kPa, the train pipe and the equalizing air cylinder pressures both drop from the stable value of 600kPa to 420kPa. The pneumatic control process is as follows. Figure 3 As shown;

[0058] S2: From the balanced air cylinder control system test bench, different fault types of the solenoid valve are simulated by parameter adjustment, and the offline test bench test pressure time series is collected to expand the actual operation pressure time series. In order to obtain the data set required for component fault diagnosis through the balanced air cylinder control module test, the key pressure change process of the distributed brake balanced air cylinder module in the actual operation scenario is simulated by continuously repeating the braking process and the relief process. Figure 4 As shown, the system components are the core of the accelerated aging test platform, simulating the functional implementation and data generation of the balanced air cylinder control system. Within the system components, an air compressor provides air to the air lines and cylinder components. The industrial computer's database stores monitoring data from the hardware platform uploaded by the data acquisition board and provides a human-computer interface for issuing brake release commands to the brake control module.

[0059] S3: The train balancing module operation is divided into four stages: rapid charging, stable charging, rapid exhaust, and stable exhaust. The relevant pressure time series of relief and braking are extracted in each stage to obtain the fault diagnosis characteristics of the solenoid valve;

[0060] S4: Extract multiple different time difference and pressure difference features in the rapid filling and exhaust stages and the stable filling and exhaust stages. For the sensor pressure change data of the four key components of the control module, namely the balancing air cylinder, train pipe, pre-control air cylinder and brake cylinder, it is necessary to construct brake fault diagnosis features that are different from abnormality detection to realize the extraction of original features of brake fault diagnosis. And sort these features by importance to filter out important features. According to the stage division of the braking and relief process of the brake mechanism, calculate the time difference and pressure difference signals of the distributed brake output pressure to obtain the time difference feature and pressure difference feature;

[0061] S5: After obtaining a large number of different types of pressure change features, there are related features, irrelevant features and redundant features among these different types of features. However, irrelevant features or redundant features will reduce the accuracy of the key component fault diagnosis model, so it is crucial to further select fault features and reduce feature dimensions. Based on the evidence theory to analyze the importance of features, through feature conflict resolution, we can obtain features with higher relevance to solenoid valve fault diagnosis; using evidence theory to analyze the importance of features, we can obtain data features that are sensitive to solenoid valve fault diagnosis. Specifically:

[0062] A1: Set the feature debate framework Φ = {θ1,θ2,…θ N}, define the mapping function m to represent the acceptance degree of the selected features, and

[0063] m(φ)=0

[0064]

[0065] Where φ indicates that the view is not within the debate framework, m(A) represents the basic probability distribution value of argument A, which describes the degree of acceptance of view A, that is, the evidence. If m(A)>0, A is called a focal element.

[0066] A2: On the debate framework Φ, calculate the trust function Beli(A) and likelihood function P of the selected features Beli (A), and obtain the corresponding confidence interval [Beli(A),P Beli (A)], for argument A

[0067]

[0068]

[0069] In the formula represents the adversarial set of argument A;

[0070] A3: For the selected features, after fusing evidence from other arguments, the final fault diagnosis features are obtained

[0071]

[0072]

[0073] In the formula is a conflict factor, reflecting argument B A and C A conflicting evidence.

[0074] S6: The offline sufficient equalization module test bench pressure time series and a small amount of actual operation fault data are integrated to solve the problem of lack of fault labels in actual operation data. The extracted charging and exhaust solenoid valve operation frequency is used as a set of training data.

[0075] S7: Use the training data and the solenoid valve fault type as the input and output of the fault diagnosis model, respectively, to train the transfer learning model. Figure 5 As shown, the domain adversarial transfer learning fault diagnosis scheme first initializes the parameters, and then the feature generator extracts the features of the source domain data and inputs the features into the fault classifier. The fault classifier is trained using the source domain sample features, and mainly diagnoses the component fault condition by relying on the recognition ability of the fault classifier. The features generated by the feature generator are also passed into the domain classifier. Because the source domain samples are labeled, when extracting features, not only the situation of the subsequent domain classifier should be considered, but also the labeled samples of the source domain should be used for supervised training to take into account the accuracy of classification. The continuous confrontation between the feature classifier and the domain classifier enables the feature generator to continuously update the common features. Through continuous iteration and back propagation, the parameters of the feature generator are updated to minimize the fault classification loss, thereby achieving accurate component fault diagnosis. Specifically:

[0076] B1: Calculate the information divergence of different source domain datasets on the testbed with respect to the target domain, select the source domain with the smallest information divergence as part of the input of the domain adversarial transfer learning model, and calculate the source domain information divergence using the following formula:

[0077]

[0078] Where H(P,Q) represents the cross entropy, which is used to measure the difference between the source domain probability distribution P(x) and the target domain probability distribution Q(x), and H(P) is the information entropy, which represents the average value of self-information.

[0079] B2: The feature generator is used to generate key features of samples in the mixed sample set, enabling the fault classifier to identify faults as closely as possible. This also makes it impossible for the domain classifier to distinguish which domain the features come from, thus ensuring the universality of the extracted features.

[0080] B3: The domain classifier classifies the data in the feature space and tries to identify the domain from which the data comes.

[0081] B4: The fault classifier diagnoses the input features and gives the final fault diagnosis result. The target loss function during training is calculated as follows:

[0082]

[0083] Where L y and L dare the cross entropy losses of the fault classifier and domain classifier, respectively, n s and n t are the number of samples in the source domain and target domain respectively, D s and D t are the source domain and target domain sample sets respectively, d i is the fault prediction result of the i-th sample, y i is the measured fault of the i-th sample. The corresponding parameters of the fault classifier and domain classifier are updated as follows:

[0084]

[0085]

[0086] In the formula and Updated parameters for feature generator, fault classifier, and domain classifier, respectively.

[0087] The present invention adopts a recursive feature elimination method to select effective feature pressure time series from the collected data, and fills the empty values ​​in the actual operation data by averaging the values ​​at the previous and next moments.

[0088] S8: Repeat S4 to S7, back-propagate the domain adversarial classifier parameters and the fault classifier parameters, and obtain the final fault classifier model;

[0089] S9: Obtain the time series of the total air cylinder, balancing air cylinder, and train pipe pressure of the real train balancing module where the solenoid valve to be tested is located. Use the domain adversarial training method to obtain common features in the source domain and target domain, and input them into the fault classifier in S8 to obtain the final solenoid valve fault diagnosis result.

[0090] Figure 5 This is an architecture diagram of the domain adversarial transfer learning fault diagnosis method of an embodiment of the present invention: it includes a feature extraction module, a domain adversarial module, and a fault diagnosis module. To implement fault diagnosis of the solenoid valve, the solenoid valve is first tested on a test bench to generate multiple source domain data, and the source domain closest to the current actual operation data is selected based on the information divergence. Then, the source domain data is input into the feature generator, and the feature generator maps the input into a high-level feature representation through a deep convolutional neural network. After that, the generated features are input into the fault classifier, which diagnoses the fault based on the extracted features and continuously adjusts the model parameters based on the fault labels of the test data. At the same time, the mixed samples of the test bench source domain data and the actual operation data of the brake are generated through the feature generator to generate features, and the model parameters are adjusted using gradient reversal based on the results of the domain classifier, so that the generated features have universality in the source domain and the target domain.

[0091] In this embodiment, the domain adversarial transfer learning method further includes: in order to avoid overfitting when the target supervision is very small, adversarial training is performed using a feature generator and a domain classifier. The input of the domain classifier is the extracted features, and the loss of each classifier is back-propagated to update the parameters. The gradient reversal layer is used to obtain the gradient from the domain classifier, but changes the sign before back-propagating to the feature generator. The information extracted by the feature generator is passed to the domain classifier, and then the domain classifier determines whether the incoming information is from the source domain or the target domain, and calculates the loss. The training goal of the domain classifier is to classify the input information into the correct domain category as much as possible, while the training goal of the feature generator is exactly the opposite. The purpose of the extracted features is to prevent the domain classifier from correctly determining which domain the information comes from, thus forming an adversarial relationship.

[0092] The structure diagram of the solenoid valve fault diagnosis system based on domain adversarial transfer learning provided by the present invention is as follows: Figure 6 As shown. The diagnostic system includes: a data acquisition module, a balanced air cylinder control system test bench module, a feature extraction module, a domain confrontation module, and a fault diagnosis module. The data acquisition module collects the voltage time series of the actual operation of the solenoid valve. At the same time, the balanced air cylinder control system test bench module generates a test voltage time series to make up for the fault data lacking in the data acquisition module. The feature extraction module extracts the time difference characteristics and pressure difference characteristics of the test voltage time series and the actual voltage time series respectively. On the one hand, the fault diagnosis module diagnoses the solenoid valve fault through the actual voltage time series to obtain the fault diagnosis result; on the other hand, the domain confrontation module identifies whether the feature comes from the balanced air cylinder control system test bench module or the data acquisition module, further calculates the domain loss and feeds it back to the feature extraction module. The feature extraction module further reselects features until the domain confrontation module cannot distinguish. Through the continuous confrontation between the domain confrontation module and the feature extraction module, the test voltage time series is selected to make up for the insufficient actual pressure time series, and finally the solenoid valve fault diagnosis is realized.

[0093] The detailed functions of each module of the solenoid valve fault diagnosis system are as follows:

[0094] The data acquisition module is used to collect actual operating pressure time series using solenoid valves with different fault types in a real train balancing module. During I / O data acquisition, pressure sensors collect pressure time series on the main air cylinder, balancing air cylinder, and train pipes. Data from each channel is synchronized in time, with a sampling rate of up to 250kSa / s, a resolution of 12 bits, an input and output buffer size of 16KB, and a USB 2.0 interface. The pressure time series are transmitted to an industrial computer via TCP / IP.

[0095] In addition, the data acquisition module is used for: When the balancing air cylinder control module of the distributed brake is in normal operation, the pressure data of each key component collected is time series data, and the time interval between adjacent sampling points is 50ms. By amplifying a single filling and exhaust process, a filling and exhaust pressure curve can be obtained. After setting the target pressure value to 600kPa, the pressure of the brake balancing air cylinder will rise immediately, and then the train pipe pressure will also rise. After the target pressure value is readjusted to 420kPa, the pressure of the train pipe and the balancing air cylinder both drop from the stable value of 600kPa to 420kPa.

[0096] The balancing air cylinder control system test bench module is used to: simulate different fault types of the solenoid valve through parameter adjustment, so as to collect and obtain a large number of pressure time series under different fault types. In order to obtain the data set required for component fault diagnosis through the balancing air cylinder control module test, the key pressure change process of the distributed brake balancing air cylinder module in the actual operation scenario can be simulated by continuously repeating the braking process and the relief process. The system component is the core of the accelerated aging test platform, which is used to simulate the function realization and data generation of the balancing air cylinder control system. In the system component, the air compressor provides the air source for the air pipeline and cylinder components. The database of the industrial computer is used to store the monitoring data in the hardware platform uploaded by the data acquisition board, and at the same time provides an operating human-computer interaction interface for issuing brake relief commands to the brake control module;

[0097] The feature extraction module is used to extract common features from the mixed real train equalization module pressure time series and the equalization air cylinder control test bench pressure time series;

[0098] The domain adversarial module is used to: identify which pressure time series set the input feature comes from;

[0099] The fault diagnosis module is used to: use the mixed real train balancing module pressure time series and the balanced air cylinder control test bench pressure time series as input, and the real train balancing module fault type as output to train a convolutional neural network model to obtain a solenoid valve fault diagnosis model;

[0100] The fault diagnosis module is also used to: obtain the pressure time series of the balancing air cylinder and the train pipe in the real train balancing module where the solenoid valve to be tested is located, mix the extracted data and the test bench pressure time series, input the extracted features into the solenoid valve fault diagnosis model, and obtain the fault type of the solenoid valve to be tested.

[0101] In summary, in this embodiment, a distributed brake fault diagnosis method based on domain adversarial transfer learning is used. First, the brake fault diagnosis features are initialized according to the evidence theory, and then the offline source domain samples and the actual operation samples are integrated to expand the brake fault diagnosis training samples. The cause analysis of the abnormal operation of the distributed brake is realized through the training of the fault diagnosis domain adversarial transfer learning classification model.

[0102] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A solenoid valve fault diagnosis method based on domain adversarial transfer learning, characterized in that: The steps include: S1: Extract the actual operating pressure time series of the solenoid valves from the solenoid valves with different fault types in the train balancing module; S2: From the balanced air cylinder control system test bench, different fault types of the solenoid valve are simulated by parameter adjustment, and the pressure time series of the solenoid valve tested on the offline test bench are collected; S3: The train balancing module operation is divided into four stages: rapid charging, stable charging, rapid exhaust, and stable exhaust. The pressure time series of relief and braking are extracted in each stage to obtain the fault diagnosis characteristics of the solenoid valve; S4: Extract multiple different time difference and pressure difference features in the rapid filling and exhaust stage and the stable filling and exhaust stage; S5: Analyze feature importance based on evidence theory and resolve feature conflicts to obtain features with higher relevance to solenoid valve fault diagnosis. S6: Fusion of the offline equalization module test bench pressure time series and the actual operating fault pressure time series to address the lack of fault labels in actual operating data. The extracted charging and exhaust solenoid valve operating frequencies are used as a set of training data. S7: Using the training data and the solenoid valve fault type as the input and output of the fault diagnosis model, respectively, the transfer learning model is trained to obtain the fault diagnosis result of the solenoid valve; S8: Repeat S4 to S7, back-propagate the domain adversarial classifier parameters and the fault classifier parameters, and obtain the final fault classifier model; S9: Obtain the time series of the total cylinder, balancing air cylinder, and train pipe pressure of the real train balancing module where the solenoid valve to be tested is located. Use the domain adversarial training method to obtain common features in the source domain and target domain, and input them into the fault classifier in S8 to obtain the final solenoid valve fault diagnosis result.

2. The solenoid valve fault diagnosis method based on domain adversarial transfer learning according to claim 1 is characterized in that: In step S2, the solenoid valve fault diagnosis types include: normal solenoid valve, air charging solenoid valve leakage fault, air exhaust solenoid valve leakage fault, spring aging, and rubber ring fault.

3. The solenoid valve fault diagnosis method based on domain adversarial transfer learning according to claim 1 is characterized in that: In step S4, the pressure characteristic values ​​of the rapid filling and exhausting and stable filling and exhausting stages are extracted, including the time difference and the pressure difference, which are specifically: Time difference characteristics: Based on the results of the pressure curve stage division, the time spent on the pressure changes of the key components of the train control module, the balancing air cylinder, the train pipe, the pre-control air cylinder, and the brake cylinder in the four stages of rapid exhaust, stable exhaust, rapid filling, and stable filling is used as the time difference characteristics for brake fault diagnosis; The pressure difference features are: the difference between the target value of the balancing air cylinder pressure and the balancing air cylinder pressure, the difference between the balancing air cylinder pressure and the train pipe pressure, the difference between the target value of the pre-control air cylinder pressure and the pre-control air cylinder pressure, and the difference between the pre-control air cylinder pressure and the brake cylinder pressure. Initial features are extracted from these four sets of pressure differences to construct multi-dimensional brake state features.

4. The solenoid valve fault diagnosis method based on domain adversarial transfer learning according to claim 1 is characterized in that: In step S5, the importance analysis of features is performed using evidence theory to obtain data features that are sensitive to solenoid valve fault diagnosis, further comprising: A1: Set the feature debate framework Φ = {θ1,θ2,…θ N }, define the mapping function m to represent the acceptance degree of the selected features, and m(φ)=0 Where φ indicates that the view is not within the debate framework, m(A) represents the basic probability distribution value of argument A, which describes the degree of acceptance of view A, that is, the evidence. If m(A)>0, A is called a focal element. A2: On the debate framework Φ, calculate the trust function Beli(A) and likelihood function P of the selected features Beli (A), and obtain the corresponding confidence interval [Beli(A),P Beli (A)], for argument A Where A represents the adversarial set of argument A; A3: For the selected features, after fusing evidence from other arguments, the final fault diagnosis features are obtained In the formula is a conflict factor, reflecting argument B A and C A conflicting evidence.

5. The solenoid valve fault diagnosis method based on domain adversarial transfer learning according to claim 1 is characterized in that: In step S7, the migration from the test bench data to the actual operation data is achieved through the domain adversarial transfer learning method; further comprising: B1: Calculate the information divergence of different source domain datasets on the testbed with respect to the target domain, and select the source domain with the smallest information divergence as part of the input of the domain adversarial transfer learning model; B2: The feature generator is used to generate key features of samples in the mixed sample set, and the fault classifier identifies faults based on the key features; B3: The domain classifier classifies the data in the feature space and determines which domain the data comes from. B4: The fault classifier diagnoses the input features and gives the final fault diagnosis result.

6. The solenoid valve fault diagnosis method based on domain adversarial transfer learning according to claim 1 is characterized in that: The recursive feature elimination method is used to select effective characteristic pressure time series from the collected data, and the null values ​​in the actual operation data are filled by averaging the values ​​of the previous and next moments.

7. The solenoid valve fault diagnosis method based on domain adversarial transfer learning according to claim 1 is characterized in that: The feature generator and domain classifier adopt fully connected neural network, and the fault classifier adopts convolutional neural network.

8. A solenoid valve fault diagnosis system based on domain adversarial transfer learning, characterized in that: It includes data acquisition module, balanced air cylinder control system test bench module, feature extraction module, domain confrontation module, and fault diagnosis module; The data acquisition module is used to use solenoid valves with different fault types in the real train balancing module to collect the actual operating pressure time series. In the I / O data acquisition, the pressure sensor collects the pressure time series on the main air cylinder, the balancing air cylinder and the train pipe. The data of each channel is synchronized in time, with a sampling rate of up to 250kSa / s, a resolution of 12 bits, and an input and output buffer size of 16K. The pressure time series is transmitted to the industrial computer through the network interface; The balanced air cylinder control system test bench module is used to simulate different fault types of the solenoid valve by adjusting parameters to obtain a large number of pressure time series under different fault types of the solenoid valve; The feature extraction module is used to extract common features from the mixed real train equalization module pressure time series and the equalization air cylinder control test bench pressure time series; The domain adversarial module is used to identify which pressure time series set the input feature comes from; The fault diagnosis module is used to train a convolutional neural network model using a mixed real train balancing module pressure time series and a balanced air cylinder control test bench pressure time series as input and a real train balancing module fault type as output to obtain a solenoid valve fault diagnosis model.

Images