Electric drive system fault diagnosis based on reversible neural network assisted canonical correlation analysis
By using reversible neural networks to assist canonical correlation analysis, a residual generator was constructed to solve the nonlinear fault diagnosis problem of electric drive systems, achieving high-precision fault detection and location.
Patent Information
- Application Number
- CN202511256690.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-04
- Publication Date
- 2026-01-27
AI Technical Summary
Existing fault diagnosis methods for electric drive systems struggle to handle nonlinear characteristics, resulting in low detection accuracy and difficulty in fault location. Traditional methods cannot meet the needs of practical applications.
A reversible neural network is used to assist canonical correlation analysis. By constructing a residual generator, the reversible mapping of nonlinear variables is realized, and fault diagnosis is performed in combination with test statistics.
It improves the accuracy of fault detection and fault location capabilities in electric drive systems, enabling rapid and accurate fault diagnosis.
Smart Images
Figure CN121412480A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of fault diagnosis, and particularly discloses a method for power drive system fault diagnosis based on reversible neural network assisted canonical correlation analysis, which is suitable for the fault diagnosis task of the power drive system. BACKGROUND
[0002] The power drive system is the core equipment in the fields of industrial production and rail transportation, and its safety and reliability are crucial. As a key component of the power drive system state monitoring, the sensor may cause control imbalance or even system collapse once it fails. Therefore, an efficient fault diagnosis technology is a core requirement to ensure the safe operation of the power drive system.
[0003] The existing power drive system fault diagnosis methods are mainly divided into model-driven methods and data-driven methods. The model-driven fault diagnosis method relies on an accurate mathematical model. However, due to the strong nonlinearity and uncertainty of the power drive system, it is difficult to establish an accurate mathematical model thereof. This limits the application of the model-driven fault diagnosis method.
[0004] The data-driven fault diagnosis method uses sensor signals to analyze the power drive system, and can realize the fault diagnosis of the power drive system. In the data-driven fault diagnosis method, the fault diagnosis based on multivariate statistical analysis is concerned. Among them, the canonical correlation analysis method, which is the mainstream of multivariate statistical analysis, is widely used in the fault diagnosis task of the power drive system. However, the traditional fault diagnosis method based on canonical correlation analysis needs to assume that the data obeys the Gaussian distribution and the variables are linearly related, which cannot adapt to the power drive system with nonlinear characteristics, resulting in low fault diagnosis performance.
[0005] In order to solve the problem of fault diagnosis of the power drive system with nonlinear characteristics, in recent years, a method based on neural network assisted canonical correlation analysis has been proposed. However, this method has the defect of irreversible mapping, which causes the attenuation of fault information in the projection process and affects the accuracy of fault diagnosis. In addition, the existing method based on neural network assisted canonical correlation analysis lacks the direct positioning ability of fault location, which is difficult to meet the actual application requirements.
[0006] The present application aims to overcome the shortcomings of the existing power drive system fault diagnosis technology, and proposes a method for power drive system fault diagnosis based on reversible neural network assisted canonical correlation analysis. The reversible neural network is used to realize the reversible mapping of nonlinear variables, the residual signal is generated by combining the canonical correlation analysis, and the test statistic is used to realize the rapid diagnosis of the fault. SUMMARY
[0007] This invention provides a fault diagnosis method for electric drive systems based on reversible neural network-assisted canonical correlation analysis, which solves the problems of low detection accuracy and difficulty in fault location in nonlinear electric drive systems.
[0008] The core of this invention is to create a framework based on reversible neural network-assisted canonical correlation analysis, through which a residual generator is built to perform fault diagnosis tasks in electric drive systems. The method of this invention mainly includes two key stages: constructing a residual generator based on reversible neural network-assisted canonical correlation analysis and implementing fault diagnosis.
[0009] The specific process of the residual generator based on invertible neural network-assisted canonical correlation analysis is as follows: Electric drive system model based on signal form : , in, and They are The input and output of the electric drive system at all times This represents an offline dataset for electric drive systems. Let be the total number of samples in the dataset; if a fault occurs in the electric drive system, the data collected by the sensors will change, and the sensor data with fault information can be represented as: , Among them, Split into and ,Right now: , , , , These respectively indicate that there is a fault. In , These respectively indicate that there is a fault. In Under fault-free conditions, and It will follow the following distribution: , in, and They are and The mean of discrete data. and It is autocovariance. It is mutual covariance; Generalized objective function based on canonical correlation analysis Defined as: , in, This represents the absolute value operator. Let represent the trace operator; perform singular value decomposition on the above expression, i.e.: ,in, , , , Indicates the number of principal components. , The canonical correlation coefficient; A typical vector can be defined as , in, It is the number of non-zero singular values. and They are The former column sum The former Okay; therefore, the optimization problem is defined as: , in, Indicates to make When the maximum value is obtained and Values, Indicates to make When the maximum value is obtained and Values; A reversible neural network can be represented as: , in, , , , , , , , , and Represents a nonlinear mapping. Represent the mapping relationship of invertible neural networks; construct a loss function for canonical correlation analysis assisted by invertible neural networks. : ; Based on the above formula, the mapping relationship of the invertible neural network can be trained. Therefore, the residual signal based on canonical correlation analysis assisted by a reversible neural network is: .
[0010] The specific process of fault diagnosis is as follows: Assume the fault in the electric drive system occurs The residual signal obtained from canonical correlation analysis assisted by a reversible neural network is as follows: ; Using Taylor's formula Analysis reveals: , in, Indicates to Find the partial derivative. express The Hessian matrix; therefore, we can obtain: ; make, , Then, it becomes: ; Assume the fault in the electric drive system occurs The residual signal obtained from canonical correlation analysis assisted by a reversible neural network is as follows: , Then, the fault isolation task is carried out; Design Group 4 Test statistic: ; Design four groups under fault-free conditions. Threshold for the test statistic: , in, This indicates retrieving the maximum value from the set. If a fault occurs in the electric drive system If above, then the following conditions are met: ; If a fault occurs in the electric drive system If above, then the following conditions are met: ; If the electric drive system failure occurs simultaneously and If above, then the following conditions are met: .
[0011] In summary, compared with existing fault diagnosis technologies, the advantages of this invention are:
[0012] Firstly, this invention utilizes a reversible neural network to construct... and The invertible mapping between them.
[0013] Secondly, this invention utilizes reversible neural networks to assist canonical correlation analysis in extracting the correlation between nonlinear variables.
[0014] Thirdly, this invention constructs four sets of residual signals based on canonical correlation analysis assisted by reversible neural networks to achieve fault diagnosis and location of electric drive systems. Attached Figure Description
[0015] Figure 1-2 is a block diagram of the reversible neural network described in this invention;
[0016] Figure 3 is a schematic diagram of the residual generator based on reversible neural network-assisted canonical correlation analysis as described in this invention;
[0017] Figure 4 is a flowchart of the fault diagnosis process based on reversible neural network-assisted canonical correlation analysis as described in this invention.
[0018] Figure 5 is a waveform diagram of the sensor signal of the electric drive system described in this invention;
[0019] Figure 6-8 shows the sensor signal waveforms of the electric drive system after the injection of the three types of faults described in this invention;
[0020] Figure 9-11 shows the fault diagnosis results based on reversible neural network-assisted canonical correlation analysis as described in this invention. Detailed Implementation
[0021] The present invention will be further described below with reference to the embodiments and the accompanying drawings:
[0022] The experimental object of this invention is a permanent magnet synchronous motor drive system employing a vector control strategy. This system mainly includes: a control unit, an execution unit, a sensing unit, and a data acquisition unit. The core parameters of the electric drive system are shown in Table 1, including the motor's moment of inertia (…). ), electromagnetic torque ( ), motor coil resistance ( ), motor coil inductance ( ),magnetic flux( ), sampling time ( This ensures that the experimental conditions are consistent with the actual industrial scenario.
[0023] Table 1 Main parameters of the traction system
[0024] The sensing unit includes three phase current sensors ( , , Two stator voltage sensors , ), a speed sensor ( Under normal system operation, the fault-free data waveform is shown in Figure 5. The data is grouped as follows:
[0025] To verify the fault diagnosis capability of the method of the present invention, three typical sensor faults were injected into the electric drive system to simulate common sensor anomalies in real-world scenarios:
[0026] Bias fault A bias fault with an amplitude of 0.025 A is injected into the electric drive system at 50.4 s. In the middle; when the electric drive system has been running for 50.7 s, a bias fault with an amplitude of 0.05 A is injected. As shown in Figure 6, after the bias fault was injected, the signal amplitudes of the six sensor groups did not change significantly.
[0027] Synchronous bias fault When the electric drive system has been running for 50.4 seconds, a synchronization bias fault with an amplitude of 0.05 A is injected. , , In Figure 7, after a synchronous bias fault is injected, the current sensor ( , , ) and voltage sensor ( , The magnitude of the change was not significant.
[0028] Intermittent fault When the electric drive system had been running for 50.4 s, an intermittent fault with an amplitude of 20 km / h was injected into speed sensor S. The duration of the intermittent fault was 0.3 s. As shown in Figure 8, after the intermittent fault was injected, the speed sensor (S) showed a relatively obvious amplitude fluctuation.
[0029] Figure 9-11 shows the fault diagnosis results based on canonical correlation analysis assisted by a reversible neural network. The solid line represents the waveform of the test statistic, and the dashed line represents the threshold.
[0030] A bias fault occurs in the electric drive system. back, and Rapidly exceeded the threshold and As shown in Figure 9, the diagnostic results based on canonical correlation analysis assisted by reversible neural networks are consistent with... The diagnostic logic for the fault occurred. Additionally, after the magnitude of the bias fault increased to 0.05 A, the value of the test statistic also increased.
[0031] Synchronization bias fault occurs in the electric drive system Then, as shown in Figure 10, and Rapidly exceeded the threshold and . and Some exceeded the threshold and The main reason is a synchronous bias fault. By disrupting the closed-loop control logic of the electric drive system, it leads to the source of the fault. Indirectly related Abnormal fluctuations occur, causing the statistics of the inverse mapped residual signal to exceed a threshold. However, the method of this invention can still provide accurate diagnostic results.
[0032] Intermittent faults occur in the electric drive system Then, as shown in Figure 11, and Rapidly exceeded the threshold and The diagnostic results based on canonical correlation analysis assisted by reversible neural networks are consistent with... Diagnostic logic for fault occurrence.
[0033] This invention is compared with existing methods (Local Linear Generalized Autoencoder (LLGAE), Canonical Correlation Analysis (CCA), and Single-side Neural Network-aided Canonical Correlation Analysis (SsCCA)). False Alarm Rates (FARs) and Missed Detection Rates (MDRs) are used as evaluation metrics, and the results are shown in Table 2. Where "—" indicates an empty string; Table 2 Performance Comparison of Four Methods
Claims
1. A method for fault diagnosis of an electric drive system using reversible neural network-assisted canonical correlation analysis, characterized in that, The method includes: A residual generator based on invertible neural network-assisted canonical correlation analysis; Fault diagnosis.
2. The method for fault diagnosis of electric drive systems using reversible neural network-assisted canonical correlation analysis according to claim 1, characterized in that, The residual generator based on reversible neural network-assisted canonical correlation analysis includes: Electric drive system model based on signal form : ， in, and They are The input and output of the electric drive system at all times This represents an offline dataset for electric drive systems. Let be the total number of samples in the dataset; if a fault occurs in the electric drive system, the data collected by the sensors will change, and the sensor data with fault information can be represented as: ， Among them, Split into and ,Right now: , , , , These respectively indicate that there is a fault. In , These respectively indicate that there is a fault. In Under fault-free conditions, and It will follow the following distribution: ， in, and They are and The mean of discrete data. and It is autocovariance. It is mutual covariance; Generalized objective function based on canonical correlation analysis Defined as: ， in, This represents the absolute value operator. Let represent the trace operator; perform singular value decomposition on the above expression, i.e.: ,in, , , , Indicates the number of principal components. , The canonical correlation coefficient; A typical vector can be defined as ， in, It is the number of non-zero singular values. and They are The former column sum The former Okay; therefore, the optimization problem is defined as: ， in, Indicates to make When the maximum value is obtained and Values, Indicates to make When the maximum value is obtained and Values; A reversible neural network can be represented as: ， in, , , , , , , , , and Represents a nonlinear mapping. Represent the mapping relationship of invertible neural networks; construct a loss function for canonical correlation analysis assisted by invertible neural networks. : ； Based on the above formula, the mapping relationship of the invertible neural network can be trained. Therefore, the residual signal based on canonical correlation analysis assisted by a reversible neural network is: 。 3. The method for fault diagnosis of electric drive systems using reversible neural network-assisted canonical correlation analysis according to claim 1, characterized in that, The fault diagnosis includes: Assume the fault in the electric drive system occurs The residual signal obtained from canonical correlation analysis assisted by a reversible neural network is as follows: ； Using Taylor's formula Analysis reveals: ， in, Indicates to Find the partial derivative. express The Hessian matrix; therefore, we can obtain: ； make, , Then, it becomes: ； Assume the fault in the electric drive system occurs The residual signal obtained from canonical correlation analysis assisted by a reversible neural network is as follows: ， Then, the fault isolation task is carried out; Design Group 4 Test statistic: ； Design four groups under fault-free conditions. Threshold for the test statistic: ， in, This indicates retrieving the maximum value from the set. If a fault occurs in the electric drive system If above, then the following conditions are met: ； If a fault occurs in the electric drive system If above, then the following conditions are met: ； If the electric drive system failure occurs simultaneously and If above, then the following conditions are met: 。