A fault diagnosis method and system based on electro-hydraulic linkage

Abstract

The application relates to the technical field of mechanical engineering and automatic control, and discloses a fault diagnosis method and system based on electro-hydraulic linkage, aiming to solve the problems of diagnosis model mismatch, early weak fault difficulty extraction and fault root location ambiguity caused by electrical and hydraulic data modal heterogeneity and deep fault feature coupling in the prior art. The method comprises the following steps: synchronously collecting electro-hydraulic multi-source signals; constructing an electro-hydraulic coupling dynamic state space model based on health data; generating a multi-dimensional residual sequence through a Kalman filter; performing multi-scale wavelet time-frequency decomposition on the residual and extracting features; fusing cross-domain features by using canonical correlation analysis; and finally realizing accurate fault identification through a support vector machine. The system comprises a synchronous acquisition module, a modeling module, a residual generation module, a feature fusion module and a fault identification module. The application significantly improves early fault sensitivity and coupling fault distinguishability through mechanism modeling and data driving fusion.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of mechanical engineering and automatic control technology, and particularly relates to a fault diagnosis method and system based on electro-hydraulic linkage. BACKGROUND

[0002] With the increasing demand for system reliability in fields such as high-end equipment, aerospace and heavy machinery, the fault diagnosis technology of electro-hydraulic linkage system has become a key link to ensure the safe operation of equipment. Electro-hydraulic linkage system drives hydraulic actuators through electrical control signals to realize high-precision and large-load power output, and has the advantages of fast response speed and large output torque under complex working conditions. However, its multi-physical domain coupling characteristics (electrical signals, hydraulic fluids, mechanical movements) make the internal state of the system highly nonlinear and interwoven, and single sensor or single domain monitoring means cannot fully reflect the real running state of the system, resulting in problems such as difficulty in extracting fault features, high false alarm rate and diagnosis lag.

[0003] Among them, the fault diagnosis method based on electro-hydraulic linkage focuses on the fusion of multi-source heterogeneous data such as current, voltage, control instruction signals on the electrical side and pressure, flow, displacement on the hydraulic side, and identifies potential abnormalities through cross-domain correlation analysis. The core of this technology is to establish a dynamic mapping relationship between electro-hydraulic responses, and on this basis to capture the coupling deviation caused by element wear, leakage, blockage or control mismatch. However, existing technologies generally use independent modeling or simple threshold comparison strategies, and fail to fully exploit the deep coupling characteristics of electro-hydraulic signals in the time domain, frequency domain and energy dimension.

[0004] In the prior art, some schemes rely on fixed rule library or offline training classification model for fault judgment, lacking self-adaptive ability to dynamic changes of working conditions; although some methods introduce data-driven models, they do not consider the non-stationary characteristics of electro-hydraulic systems under different loads, temperatures or running stages, resulting in poor model generalization ability. In addition, most diagnosis systems only focus on a single fault mode, and are difficult to meet the identification needs of complex faults or early weak anomalies. Therefore, there is an urgent need for a fault diagnosis method and system that can deeply integrate electro-hydraulic multi-source information, dynamically perceive the running state, and have strong robustness and self-adaptive ability. SUMMARY

[0005] The technical problem to be solved by the present application is to provide a fault diagnosis method and system based on electro-hydraulic linkage, aiming to overcome the technical bottlenecks of diagnosis model mismatch, difficulty in extracting early weak fault features and fuzzy fault root location caused by data modal heterogeneity of electrical and hydraulic domains and deep coupling of fault features in the prior art.

[0006] To address the aforementioned technical problems, this invention provides a fault diagnosis method and system based on electro-hydraulic linkage. This invention utilizes high-precision, highly synchronous data acquisition to construct a coupled state-space model that accurately characterizes the interaction between electrical and hydraulic dynamics, serving as a health benchmark. Real-time acquired electro-hydraulic data is input into a Kalman filter state observer based on this model, generating a multi-dimensional residual sequence reflecting the deviation between the actual system operating state and the health benchmark model prediction. This invention no longer analyzes single residual signals in isolation, but instead performs synchronous multi-scale time-frequency decomposition on the residual sequences of the electrical and hydraulic domains, extracting their deep-seated time-varying and frequency-varying features. The core lies in introducing a canonical correlation analysis method to maximize the correlation projection of the extracted cross-domain features, generating a fusion feature vector highly sensitive to electro-hydraulic coupling faults. Finally, a pre-trained support vector machine classifier is used to perform pattern recognition on this fusion feature vector, achieving accurate identification and localization of specific fault types.

[0007] According to one aspect of the present invention, a fault diagnosis method based on electro-hydraulic linkage is provided, which includes the following steps:

[0008] Simultaneously acquire multiple electrical domain sensor signals and multiple hydraulic domain sensor signals of the electro-hydraulic linkage system to form a raw dataset with a unified high-precision timestamp.

[0009] Based on historical data collected when the electro-hydraulic linkage system is operating in a healthy state, an electro-hydraulic coupling dynamic state space model is constructed using a subspace identification algorithm.

[0010] A Kalman filter state observer based on the electro-hydraulic coupling dynamic state-space model is constructed, and the real-time synchronously acquired sensor signals are used as input to estimate the system state vector online.

[0011] The system output estimated by the state observer is subtracted point by point from the actual measured value of the real-time sensor signal to generate a multidimensional residual sequence vector.

[0012] Multi-level discrete wavelet transform time-frequency decomposition is performed on each residual signal in the multi-dimensional residual sequence vector, wavelet coefficients at each decomposition level are extracted, and time-frequency feature set is calculated based on the wavelet coefficients;

[0013] The time-frequency characteristics of all electrical domain residual signals are integrated into an electrical domain feature matrix, and the time-frequency characteristics of all hydraulic domain residual signals are integrated into a hydraulic domain feature matrix. The normalized correlation analysis algorithm is used to calculate the normalized transformation basis that maximizes the correlation after projection of the two matrices, and the projection result is used as the electro-hydraulic coupling fault fusion feature vector.

[0014] The fusion feature vector of electro-hydraulic coupling faults is input into a pre-trained multi-class support vector machine model, and the fault category diagnosis result of the electro-hydraulic linkage system is output.

[0015] Furthermore, the simultaneous acquisition of multiple electrical domain sensor signals and multiple hydraulic domain sensor signals of the electro-hydraulic linkage system is specifically achieved through a data acquisition unit with a central synchronous clock source.

[0016] The central synchronization clock source provides a unified trigger acquisition signal and timestamp reference for all data acquisition channels, ensuring that the timestamp error of different physical quantity sensing signals is within the microsecond range. The acquisition frequency of the electrical domain sensing signal is set to 10 kHz, and the acquisition frequency of the hydraulic domain sensing signal is set to 1 kHz. The hydraulic domain data is upsampled to the same frequency as the electrical domain data through a linear interpolation algorithm to form strictly aligned time-series data pairs.

[0017] Furthermore, a dynamic state-space model of electro-hydraulic coupling is constructed using a subspace identification algorithm, specifically employing the N4SID algorithm based on random subspace identification. The long-term input-output data collected under healthy conditions is divided into past and future data blocks. By performing oblique projection and singular value decomposition on the Hankel matrix constructed from the data, the system state sequence is directly estimated, and the optimal estimates of the system state matrix, input matrix, output matrix, and direct transfer matrix are obtained. The order of the model is determined by analyzing the distribution of singular values ​​in the Hankel matrix.

[0018] Furthermore, time-frequency decomposition is performed using multi-level discrete wavelet transform, specifically employing the fourth-order Dobbets wavelet basis function. Five-level decomposition is performed on each residual signal to obtain an approximate coefficient sub-band and five detail coefficient sub-bands. The time-frequency feature set specifically includes: the energy value, energy entropy, kurtosis factor, and standard deviation of the wavelet coefficients at each decomposition level.

[0019] Furthermore, the time-frequency features extracted from all electrical domain residual signals are integrated into an electrical domain feature matrix, and the time-frequency features extracted from all hydraulic domain residual signals are integrated into a hydraulic domain feature matrix. A normalized correlation analysis algorithm is used to calculate a set of normalized transformation bases that maximize the correlation between the electrical domain feature matrix and the hydraulic domain feature matrix after projection, and the projection results of the two matrices on the transformation base are used as the final electro-hydraulic coupling fault fusion feature vector.

[0020] Furthermore, the training process of the pre-trained multi-class support vector machine model includes:

[0021] On the physical prototype or high-fidelity simulation model of the electro-hydraulic linkage system, a corresponding fault dataset is collected by actively injecting a single, deterministic fault mode. The fault modes cover typical electrical, hydraulic and electromechanical coupling faults such as short circuit between turns of motor winding, zero drift of pressure sensor, internal leakage of actuator, and stuck valve core of directional valve.

[0022] For each type of fault dataset collected, the steps of generating multidimensional residual sequence vectors, extracting time-frequency features, and performing normative correlation analysis are repeated to obtain electro-hydraulic coupling fault fusion feature vector samples that correspond one-to-one with each fault mode. Using the fusion feature vector samples of all fault modes and the fusion feature vector samples of the healthy state, a training set with category labels is constructed.

[0023] Radial basis functions are used as kernel functions. The optimal penalty coefficient and kernel function parameters are determined by grid search and cross-validation methods. The multi-class support vector machine is trained until the classification accuracy of the model on the validation set reaches a preset threshold.

[0024] Furthermore, a Kalman filter state observer is constructed, including:

[0025] The state equations of the electro-hydraulic coupling dynamic state-space model are used as the state equations of the Kalman filter, and the observation equations are defined by the output matrix.

[0026] Initialize the state vector as a zero vector and the covariance matrix as the identity matrix;

[0027] The observed noise covariance matrix is ​​determined based on the accuracy specifications in the sensor datasheet and the measured noise level, and the process noise covariance matrix is ​​set based on the theoretical model of the system's dynamic response.

[0028] Within each sampling period, two sub-steps, prediction and update, are performed to estimate the system state vector online at a frequency of 10 kHz.

[0029] According to another aspect of the present invention, a fault diagnosis system based on electro-hydraulic linkage is provided, comprising:

[0030] The multimodal data synchronous acquisition module is used to synchronously acquire multiple electrical domain sensing signals and multiple hydraulic domain sensing signals of the electro-hydraulic linkage system to form a raw dataset with a unified high-precision timestamp. The electrical domain sensing signals include the three-phase current, terminal voltage, rotor angular velocity and winding temperature of the drive motor, and the hydraulic domain sensing signals include the hydraulic pump outlet pressure, load actuator chamber pressure, main oil circuit flow rate, actuator displacement and hydraulic oil temperature.

[0031] The electro-hydraulic coupling dynamic state space modeling module contains an electro-hydraulic coupling dynamic state space model built based on the health status data of the electro-hydraulic linkage system. The model includes a system state matrix, input matrix, output matrix, and direct transfer matrix that describe the internal dynamics of the electrical subsystem and the hydraulic subsystem and their mutual coupling.

[0032] The real-time state observation and residual generation module is connected to the multimodal data synchronous acquisition module and the electro-hydraulic coupling dynamic state space modeling module. It is used to receive real-time sensing signals and health models, construct a Kalman filter state observer, estimate the system state online, and generate a multidimensional residual sequence vector. The residual sequence vector represents the deviation between the actual operating trajectory of the system and the trajectory predicted by the health model.

[0033] The cross-domain residual feature extraction and fusion module is connected to the real-time state observation and residual generation module. It is used to perform multi-level discrete wavelet transform on the input multi-dimensional residual sequence vector to extract time-frequency features, and to fuse the extracted cross-domain features using a normalized correlation analysis algorithm to generate an electro-hydraulic coupling fault fusion feature vector.

[0034] The fault mode identification and localization module is connected to the cross-domain residual feature extraction and fusion module. It has a pre-trained multi-class support vector machine model embedded in it, which is used to receive the electro-hydraulic coupling fault fusion feature vector and output the specific fault category diagnosis result.

[0035] Furthermore, the multimodal data synchronous acquisition module includes: multiple electrical domain sensors, specifically three Hall effect current sensors and two voltage sensors installed at the output of the motor driver, as well as a platinum resistance temperature sensor mounted on the surface of the motor winding and an optical encoder mounted on the motor shaft;

[0036] Multiple hydraulic field sensors, specifically a piezoresistive pressure sensor installed at the hydraulic pump outlet and the two chambers of the actuator, a turbine flow meter installed in the main oil pipeline, a magnetostrictive displacement sensor installed on the actuator piston rod, and a sheathed thermocouple temperature sensor immersed in the oil tank.

[0037] A data acquisition card has at least sixteen analog input channels and a built-in high-stability temperature-compensated crystal oscillator as a central synchronization clock source. Its analog-to-digital conversion resolution is 24 bits, which is used to convert the analog signals of all sensors into digital signals and attach a uniform timestamp.

[0038] Furthermore, the electro-hydraulic coupling dynamic state space modeling module, the real-time state observation and residual generation module, the cross-domain residual feature extraction and fusion module, and the fault mode identification and location module are all deployed in an industrial control computer.

[0039] The industrial control computer is equipped with a multi-core processor running a real-time operating system and has a large-capacity random access memory. It communicates with the multimodal data synchronous acquisition module at high speed via an industrial Ethernet bus. The processor is responsible for performing all computational tasks, including state-space model calculations, Kalman filter iterations, wavelet transforms, canonical correlation analysis, and support vector machine classification.

[0040] In summary, this application includes at least one of the following beneficial technical effects:

[0041] (1): By constructing an electro-hydraulic coupled dynamic state space model, this invention captures the bidirectional coupling relationship between the electrical domain and the hydraulic domain from the mechanism level. Compared with the diagnostic method that regards the two domains as independent subsystems, it greatly improves the accuracy and fidelity of the system benchmark model and provides a high signal-to-noise ratio basis for subsequent residual analysis.

[0042] (2): The present invention uses a state observer based on a Kalman filter to generate residuals, which can effectively filter out the interference of measurement noise, making the generated residual sequence more sensitive to the weak system dynamic changes caused by faults, thereby realizing the effective detection of early and latent faults.

[0043] (3): This invention creatively introduces normative correlation analysis to deeply fuse cross-domain features. It is no longer a simple feature splicing, but rather it digs out the most essential correlation structure between electrical domain features and hydraulic domain features. The resulting fused feature vector significantly enhances the distinguishability of different faults, solving the problem that traditional methods have difficulty distinguishing coupled faults with different root causes but similar appearances.

[0044] (4): By combining the time-frequency analysis capability of wavelet transform with the nonlinear classification capability of support vector machine, the present invention can process highly nonlinear and non-stationary fault signals. The diagnostic system has strong robustness and generalization ability and is highly adaptable to fault diagnosis under complex working conditions. Attached Figure Description

[0045] Fig. 1 This is a schematic diagram of the overall technical solution architecture of the present invention;

[0046] Fig. 2 This is a schematic diagram of the collaborative generation of multidimensional residual sequences according to the present invention;

[0047] Fig. 3 This is a schematic diagram of the logic flow for generating fusion features in this invention; Detailed Implementation

[0048] This invention provides a fault diagnosis method and system based on electro-hydraulic linkage, aiming to solve the technical bottlenecks in existing technologies, such as diagnostic model mismatch, difficulty in extracting early weak fault features, and ambiguity in fault root cause localization caused by the heterogeneity of data modes between the electrical and hydraulic domains and deep coupling of fault features. The specific implementation methods of the method and system will be described in detail below with reference to embodiments of this invention.

[0049] Reference Figs. 1 to 3 As shown, the fault diagnosis method based on electro-hydraulic linkage includes the following steps: S1: Synchronously collect multiple electrical domain sensing signals and multiple hydraulic domain sensing signals of the electro-hydraulic linkage system to form an original dataset with a unified high-precision timestamp.

[0050] S2: Based on historical data collected when the electro-hydraulic linkage system is operating in a confirmed healthy state, an electro-hydraulic coupling dynamic state space model is constructed using a subspace identification algorithm to characterize the dynamic characteristics of the system during normal operation.

[0051] S3: Construct a Kalman filter state observer based on an electro-hydraulic coupling dynamic state space model. Use the real-time synchronously acquired electrical domain sensing signals and hydraulic domain sensing signals as inputs to the state observer, and use the state observer to estimate the current state vector of the electro-hydraulic linkage system online.

[0052] S4: Perform point-by-point subtraction on the system output estimated by the state observer and the actual measured values ​​of the sensor signals acquired in real time to generate a multidimensional residual sequence vector corresponding to the dimension of the sensor signal.

[0053] S5: For each electrical domain residual signal and each hydraulic domain residual signal in the multidimensional residual sequence vector, multi-level discrete wavelet transform is used for time-frequency decomposition. Wavelet coefficients at each decomposition level are extracted, and time-frequency feature sets characterizing the signal energy distribution, statistical properties and non-stationarity are calculated based on the wavelet coefficients.

[0054] S6: Integrate the time-frequency features extracted from all electrical domain residual signals into an electrical domain feature matrix, and integrate the time-frequency features extracted from all hydraulic domain residual signals into a hydraulic domain feature matrix; use the normalized correlation analysis algorithm to calculate a set of normalized transformation bases that maximize the correlation between the electrical domain feature matrix and the hydraulic domain feature matrix after projection, and use the projection results of the two matrices on the transformation base as the final electro-hydraulic coupling fault fusion feature vector.

[0055] S7: Input the electro-hydraulic coupling fault fusion feature vector into a pre-trained multi-class support vector machine model, and output the current fault category diagnosis result of the electro-hydraulic linkage system from the multi-class support vector machine model.

[0056] Specifically, in step S1, multiple electrical domain sensing signals and multiple hydraulic domain sensing signals of the electro-hydraulic linkage system are collected simultaneously to form an original dataset with a unified high-precision timestamp.

[0057] The electrical domain sensing signals include the three-phase current, terminal voltage, rotor angular velocity, and winding temperature of the drive motor; the hydraulic domain sensing signals include the hydraulic pump outlet pressure, load actuator chamber pressure, main oil flow rate, actuator displacement, and hydraulic oil temperature. This synchronous acquisition process is achieved through a data acquisition unit with a central synchronous clock source.

[0058] A central synchronous clock source provides a unified trigger signal and timestamp reference for all data acquisition channels, ensuring that the timestamp error of different physical quantity sensing signals is within the microsecond range. The acquisition frequency of electrical domain sensing signals is set to 10000 Hz, and the acquisition frequency of hydraulic domain sensing signals is set to 1000 Hz. A linear interpolation algorithm is used to upsample the hydraulic domain data to the same frequency as the electrical domain data to form strictly aligned time-series data pairs.

[0059] Specifically, the data acquisition card has a built-in high-stability temperature-compensated crystal oscillator as a central synchronization clock source, with an analog-to-digital conversion resolution of 24 bits, used to convert the analog signals of all sensors into digital signals and attach a uniform timestamp.

[0060] The electrical domain sensors include three Hall effect current sensors and two voltage sensors mounted at the motor driver output, a platinum resistance temperature sensor mounted on the motor winding surface, and an optical encoder mounted on the motor shaft. The hydraulic domain sensors include piezoresistive pressure sensors mounted at the hydraulic pump outlet and in the actuator chambers, a turbine flow meter mounted in the main oil line, a magnetostrictive displacement sensor mounted on the actuator piston rod, and a sheathed thermocouple temperature sensor immersed in the oil tank. All sensor signals are transmitted to the data acquisition card via shielded cables to suppress the impact of electromagnetic interference on signal integrity.

[0061] In step S2, based on historical data collected when the electro-hydraulic linkage system is operating in a confirmed healthy state, an electro-hydraulic coupling dynamic state space model is constructed using a subspace identification algorithm to characterize the dynamic characteristics of the system's normal operation.

[0062] The model includes system state matrices, input matrices, output matrices, and direct transfer matrices describing the internal dynamics and coupling effects of the electrical and hydraulic subsystems. Specifically, the N4SID algorithm based on random subspace identification is employed. Long-term input-output data collected under healthy conditions is divided into past and future data blocks. By performing oblique projection and singular value decomposition on the Hankel matrix constructed from the data, the system state sequence is directly estimated, and then the optimal estimates of the system state matrix, input matrix, output matrix, and direct transfer matrix are obtained.

[0063] The model order is determined by analyzing the distribution of singular values ​​in the Hankel matrix. Health status data must cover the entire operating range of the system under typical conditions, including no-load, half-load, full-load, and variable-load conditions, to ensure the model has sufficient generalization ability. Data acquisition lasts for at least 30 minutes, with the sampling frequency consistent with step S1. During model building, the input signals are the control voltage command of the motor driver and the set flow command of the hydraulic system; the output signals are all sensor signals acquired in step S1. The model parameters are iteratively optimized until convergence by minimizing the mean square error between the predicted and actual outputs.

[0064] In step S3, a Kalman filter state observer based on an electro-hydraulic coupling dynamic state space model is constructed. The electrical domain sensing signals and hydraulic domain sensing signals acquired in real time are used as inputs to the state observer, and the current state vector of the electro-hydraulic linkage system is estimated online using the state observer.

[0065] The state equation of the Kalman filter is directly given by the state-space model constructed in step S2, while the observation equation is defined by the output matrix. During filter initialization, the state vector is set to zero, and the covariance matrix is ​​set to the identity matrix. The process noise covariance matrix and the observation noise covariance matrix are determined through offline calibration: the process noise covariance matrix is ​​set according to the theoretical model of the system's dynamic response, while the observation noise covariance matrix is ​​calculated based on the accuracy specifications in the sensor datasheet and the measured noise level.

[0066] Within each sampling period, the Kalman filter performs two sub-steps: prediction and update. The prediction step calculates the prior state estimate for the current time step based on the state estimate from the previous time step and the current input. The update step then uses the actual measurement value at the current time step to correct the prior estimate, obtaining the posterior state estimate. This state observer operates at a frequency of ten kHz, synchronized with the data acquisition frequency.

[0067] In step S4, the system output estimated by the state observer is subtracted point by point from the actual measured values ​​of the real-time synchronously acquired sensor signals to generate a multidimensional residual sequence vector corresponding to the dimension of the sensor signals. The residual sequence vector represents the deviation between the actual trajectory of the system and the trajectory predicted by the health model. The dimension of the residual vector is equal to the total number of sensor signals, i.e., nine dimensions.

[0068] Each residual signal is calculated independently, and the calculation formula is as follows: ,in, Let be the residual vector at time k. This is the actual measurement output vector at time k. Let be the output estimate vector of the Kalman filter at time k. This residual sequence vector should randomly fluctuate around zero under healthy system conditions, and its statistical characteristics are determined by observation noise. When the system fails, the residual sequence will show a systematic shift or a sudden increase in energy, reflecting the mismatch between the model and the actual system.

[0069] In step S5, for each electrical domain residual signal and each hydraulic domain residual signal in the multidimensional residual sequence vector, time-frequency decomposition is performed using multi-level discrete wavelet transform. Wavelet coefficients at each decomposition level are extracted, and time-frequency feature sets characterizing the signal energy distribution, statistical properties, and non-stationarity are calculated based on the wavelet coefficients.

[0070] Specifically, the fourth-order Dobbets wavelet basis function is used. A five-level decomposition is performed on each residual signal, yielding one approximate coefficient sub-band and five detail coefficient sub-bands. The time-frequency feature set specifically includes: the energy value, energy entropy, kurtosis factor, and standard deviation of the wavelet coefficients at each decomposition level.

[0071] The formula for calculating energy value is: ,in, The energy of the detail coefficients at the j-th level. For the i-th wavelet coefficient of the j-th layer, This represents the total number of coefficients in this layer. Energy entropy is defined as: ,in, The formula for calculating the kurtosis factor is: ,in and Let be the mean and standard deviation of the coefficients in the j-th layer, respectively. The standard deviation is... .

[0072] A total of 24-dimensional time-frequency features were extracted from each residual signal (four-dimensional features for each of the five detail coefficients, and approximation coefficients were not included in the feature calculation). A total of 96-dimensional features were generated from the four signals in the electrical domain, and a total of 120-dimensional features were generated from the five signals in the hydraulic domain.

[0073] In step S6, the time-frequency features extracted from all electrical domain residual signals are integrated into an electrical domain feature matrix, and the time-frequency features extracted from all hydraulic domain residual signals are integrated into a hydraulic domain feature matrix. A normalized correlation analysis algorithm is used to calculate a set of normalized transformation bases that maximize the correlation between the electrical domain feature matrix and the hydraulic domain feature matrix after projection, and the projection results of the two matrices on the transformation base are used as the final electro-hydraulic coupling fault fusion feature vector.

[0074] Let the characteristic matrix of the electrical domain be... The hydraulic domain feature matrix is Where n is the number of samples, p is 96, and q is 120. The goal of canonical correlation analysis is to find the projection vector. and , making the projection variable and Maximize the correlation coefficient. This problem is equivalent to solving the generalized eigenvalue problem:

[0075]

[0076] in, , , And Cyx are the covariance matrix and cross-covariance matrix of X and Y, respectively. Specifically, Let X be the covariance matrix of the electrical domain characteristic matrix. Let Y be the covariance matrix of the hydraulic domain characteristic matrix. Let X be the cross-covariance matrix of X and Y, and Cyx be the cross-covariance matrix of Y and X. To normalize the correlation coefficients, 'a' is the projection vector for the electrical domain feature matrix X, and 'b' is the projection vector for the hydraulic domain feature matrix Y. The first k pairs of normalized variables (k is determined by the cumulative contribution rate of the normalized correlation coefficients, typically above 90%) are selected to form a fused feature vector with a dimension of 2k. This fused feature vector retains the most relevant coupling information between the electrical and hydraulic domain features and exhibits high sensitivity to electro-hydraulic linkage faults.

[0077] In step S7, the electro-hydraulic coupling fault fusion feature vector is input into a pre-trained multi-class support vector machine model, and the multi-class support vector machine model outputs the current fault category diagnosis result of the electro-hydraulic linkage system.

[0078] The training process of the multi-class support vector machine model includes: collecting corresponding fault datasets by actively injecting a single, deterministic fault mode into the physical prototype or high-fidelity simulation model of the electro-hydraulic linkage system. The fault modes cover typical electrical, hydraulic and electromechanical coupling faults such as short circuit between turns of motor winding, zero drift of pressure sensor, internal leakage of actuator, and valve core jamming of directional valve. For each type of fault dataset collected, the residual generation, time-frequency feature extraction and standard correlation analysis steps are repeatedly executed to obtain electro-hydraulic coupling fault fusion feature vector samples corresponding to each fault mode.

[0079] A training set with category labels is constructed using fused feature vector samples from all fault modes and fused feature vector samples from the healthy state. A radial basis function is used as the kernel function, and the optimal penalty coefficient and kernel function parameters are determined through grid search and cross-validation. The multi-class support vector machine is trained until the model's classification accuracy on the validation set reaches a preset threshold. The support vector machine employs a one-to-many strategy to achieve multi-class classification, and the output is a predefined fault category label, such as "healthy," "motor inter-turn short circuit," and "actuator internal leakage."

[0080] The fault diagnosis system based on electro-hydraulic linkage includes a multimodal data synchronous acquisition module, an electro-hydraulic coupling dynamic state space modeling module, a real-time state observation and residual generation module, a cross-domain residual feature extraction and fusion module, and a fault mode identification and location module.

[0081] The multimodal data synchronous acquisition module is used to synchronously acquire multiple electrical domain sensor signals and multiple hydraulic domain sensor signals of the electro-hydraulic linkage system, forming a raw dataset with a unified high-precision timestamp. The electro-hydraulic coupling dynamic state space modeling module internally contains an electro-hydraulic coupling dynamic state space model constructed based on the health state data of the electro-hydraulic linkage system.

[0082] The real-time state observation and residual generation module is connected to the multimodal data synchronous acquisition module and the electro-hydraulic coupling dynamic state-space modeling module. It receives real-time sensor signals and health models, constructs a Kalman filter state observer, estimates the system state online, and generates a multidimensional residual sequence vector. The cross-domain residual feature extraction and fusion module is connected to the real-time state observation and residual generation module. It performs multi-level discrete wavelet transform on the input multidimensional residual sequence vector to extract time-frequency features, and uses a normalized correlation analysis algorithm to fuse the extracted cross-domain features, generating an electro-hydraulic coupling fault fusion feature vector.

[0083] The fault mode identification and localization module is connected to the cross-domain residual feature extraction and fusion module. It contains a pre-trained multi-class support vector machine model, which is used to receive the electro-hydraulic coupling fault fusion feature vector and output the specific fault category diagnosis result.

[0084] The multimodal data synchronous acquisition module includes multiple electrical domain sensors, multiple hydraulic domain sensors, and a data acquisition card. The electrical domain sensors specifically include a three-channel Hall effect current sensor and two-channel voltage sensors mounted on the motor driver output, a platinum resistance temperature sensor mounted on the motor winding surface, and an optical encoder mounted on the motor shaft. The hydraulic domain sensors specifically include a piezoresistive pressure sensor mounted at the hydraulic pump outlet and in the actuator chambers, a turbine flow meter mounted in the main oil pipeline, a magnetostrictive displacement sensor mounted on the actuator piston rod, and a sheathed thermocouple temperature sensor immersed in the oil tank.

[0085] The data acquisition card has at least sixteen analog input channels and a built-in high-stability temperature-compensated crystal oscillator as a central synchronization clock source. Its analog-to-digital conversion resolution is 24 bits, which is used to convert the analog signals of all sensors into digital signals and add a uniform timestamp.

[0086] The electro-hydraulic coupling dynamic state-space modeling module, the real-time state observation and residual generation module, the cross-domain residual feature extraction and fusion module, and the fault mode identification and location module are all deployed in an industrial control computer. The industrial control computer is equipped with a multi-core processor running a real-time operating system, has a large-capacity random access memory, and communicates with the multi-modal data synchronous acquisition module at high speed via an industrial Ethernet bus.

[0087] The processor is responsible for all computational tasks, including state-space model operations, Kalman filter iterations, wavelet transforms, canonical correlation analysis, and support vector machine classification. Upon system startup, the health state model and the trained support vector machine model are first loaded into memory. During operation, steps S3 to S7 are executed cyclically at a frequency of 10 kHz to achieve online fault diagnosis and location. Diagnostic results are displayed in real-time through a human-machine interface and can trigger alarm or shutdown protection logic.

[0088] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects.

[0089] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A fault diagnosis method based on electro-hydraulic linkage, characterized in that... This includes the following steps: Simultaneously acquire multiple electrical domain sensor signals and multiple hydraulic domain sensor signals from the electro-hydraulic linkage system to form a raw dataset with a unified high-precision timestamp; Based on historical data collected during the operation of the electro-hydraulic linkage system in a healthy state, an electro-hydraulic coupling dynamic state-space model is constructed using a subspace identification algorithm. This model includes a system state matrix, input matrix, output matrix, and direct transfer matrix describing the internal dynamics and mutual coupling of the electrical and hydraulic subsystems. Specifically, the N4SID algorithm based on random subspace identification is employed. The long-term sequence input and output data collected in the healthy state is divided into past and future data blocks. By performing oblique projection and singular value decomposition on the Hankel matrix constructed from the data, the system state sequence is directly estimated, and the optimal estimates of the system state matrix, input matrix, output matrix, and direct transfer matrix are obtained. The order of the electro-hydraulic coupling dynamic state-space model is determined by analyzing the distribution of singular values ​​in the Hankel matrix. A Kalman filter state observer based on the electro-hydraulic coupling dynamic state-space model is constructed, and the system state vector is estimated online by taking the real-time synchronously acquired sensor signals as input. The system output estimated by the state observer is subtracted point by point from the actual measured values ​​of the real-time sensor signals to generate a multidimensional residual sequence vector. For each residual signal in the multidimensional residual sequence vector, perform multi-level discrete wavelet transform time-frequency decomposition, extract wavelet coefficients at each decomposition level, and calculate the time-frequency feature set based on the wavelet coefficients; The time-frequency characteristics of all electrical domain residual signals are integrated into an electrical domain feature matrix, and the time-frequency characteristics of all hydraulic domain residual signals are integrated into a hydraulic domain feature matrix. The normalized correlation analysis algorithm is used to calculate the normalized transformation basis that maximizes the correlation after projection of the two matrices, and the projection result is used as the electro-hydraulic coupling fault fusion feature vector. The electro-hydraulic coupling fault fusion feature vector is input into a pre-trained multi-class support vector machine model, which outputs the fault category diagnosis result of the electro-hydraulic linkage system. The training process of the pre-trained multi-class support vector machine model includes: collecting corresponding fault datasets by actively injecting a single, deterministic fault mode on the physical prototype or high-fidelity simulation model of the electro-hydraulic linkage system. The fault modes include inter-turn short circuit of motor winding, zero-point drift of pressure sensor, internal leakage of actuator, and valve core jamming of directional valve. For each collected fault dataset, the steps of generating multi-dimensional residual sequence vectors, extracting time-frequency features, and performing normative correlation analysis are repeated to obtain electro-hydraulic coupling fault fusion feature vector samples corresponding to each fault mode. Using the fusion feature vector samples of all fault modes and the fusion feature vector samples in the healthy state, a training set with class labels is constructed. The radial basis function is used as the kernel function, and the optimal penalty coefficient and kernel function parameters are determined by grid search and cross-validation methods. The multi-class support vector machine is trained until the classification accuracy of the model on the validation set reaches a preset threshold.

2. The fault diagnosis method based on electro-hydraulic linkage according to claim 1, characterized in that... The system synchronously acquires multiple electrical domain sensor signals and multiple hydraulic domain sensor signals from the electro-hydraulic linkage system, specifically through a data acquisition unit with a central synchronous clock source. The central synchronization clock source provides a unified trigger acquisition signal and timestamp reference for all data acquisition channels, ensuring that the timestamp error of different physical quantity sensing signals is within the microsecond range. The acquisition frequency of the electrical domain sensing signal is set to 10 kHz, and the acquisition frequency of the hydraulic domain sensing signal is set to 1 kHz. The hydraulic domain data is upsampled to the same frequency as the electrical domain data through a linear interpolation algorithm to form strictly aligned time-series data pairs.

3. The fault diagnosis method based on electro-hydraulic linkage according to claim 1, characterized in that... The time-frequency decomposition is performed using multi-level discrete wavelet transform, specifically using the fourth-order Dobbets wavelet basis function. Five-level decomposition is performed on each residual signal to obtain an approximate coefficient sub-band and five detail coefficient sub-bands. The time-frequency feature set specifically includes: the energy value, energy entropy, kurtosis factor and standard deviation of the wavelet coefficients at each decomposition level.

4. The fault diagnosis method based on electro-hydraulic linkage according to claim 1, characterized in that... The time-frequency features extracted from all electrical domain residual signals are integrated into an electrical domain feature matrix, and the time-frequency features extracted from all hydraulic domain residual signals are integrated into a hydraulic domain feature matrix. A normalized correlation analysis algorithm is used to calculate a set of normalized transformation bases that maximize the correlation between the electrical domain feature matrix and the hydraulic domain feature matrix after projection. The projection results of the two matrices on the transformation base are used as the final electro-hydraulic coupling fault fusion feature vector.

5. The fault diagnosis method based on electro-hydraulic linkage according to claim 1, characterized in that... Construct a Kalman filter state observer, including: The state equations of the electro-hydraulic coupling dynamic state-space model are used as the state equations of the Kalman filter, and the observation equations are defined by the output matrix. Initialize the state vector as a zero vector and the covariance matrix as the identity matrix; The observed noise covariance matrix is ​​determined based on the accuracy specifications in the sensor datasheet and the measured noise level. The process noise covariance matrix is ​​then set based on the theoretical model of the system's dynamic response. Within each sampling period, two sub-steps, prediction and update, are performed to estimate the system state vector online at a frequency of 10 kHz.

6. A fault diagnosis system based on electro-hydraulic linkage, applied to the fault diagnosis method based on electro-hydraulic linkage as described in any one of claims 1-5, characterized in that... ,include: The multimodal data synchronous acquisition module is used to synchronously acquire multiple electrical domain sensing signals and multiple hydraulic domain sensing signals of the electro-hydraulic linkage system to form a raw dataset with a unified high-precision timestamp; An electro-hydraulic coupling dynamic state space modeling module, which internally contains an electro-hydraulic coupling dynamic state space model constructed based on the health status data of the electro-hydraulic linkage system; The real-time state observation and residual generation module is connected to the multimodal data synchronous acquisition module and the electro-hydraulic coupling dynamic state space modeling module. It is used to receive real-time sensing signals and health models, construct a Kalman filter state observer, estimate the system state online, and generate a multidimensional residual sequence vector. The cross-domain residual feature extraction and fusion module, connected to the real-time state observation and residual generation module, is used to perform multi-level discrete wavelet transform on the input multi-dimensional residual sequence vector to extract time-frequency features, and to fuse the extracted cross-domain features using a normalized correlation analysis algorithm to generate an electro-hydraulic coupling fault fusion feature vector; The fault mode identification and localization module is connected to the cross-domain residual feature extraction and fusion module. It has a pre-trained multi-class support vector machine model embedded in it, which is used to receive the electro-hydraulic coupling fault fusion feature vector and output the specific fault category diagnosis result.

Images