Hydraulic motor fault diagnosis method and system based on heterogeneous asynchronous data fusion

Abstract

The invention discloses a hydraulic motor fault diagnosis method and system based on heterogeneous asynchronous data fusion, and the method comprises the steps: collecting heterogeneous asynchronous data of a sensor network, carrying out the preprocessing of the data, inputting the data into a parallel dynamic pruning residual network, extracting features, and carrying out the preliminary fusion, thereby obtaining an initial feature plane; and inputting the initial feature plane into a multi-connection neural network, fusing data, extracting features, obtaining a final feature plane, using the fused feature data as a training set, and building a feature classification model to test and evaluate the performance of the feature classification model. According to the method, the limitation of a traditional fusion model in asynchronous processing of high-frequency vibration signals and low-frequency thermodynamic data of a hydraulic system is effectively overcome, the robustness of key fault features in a strong noise environment is remarkably improved, and rapid virtual-real mapping of bench test simulation data and online monitoring data is realized; typical faults such as plunger pair abrasion and valve plate cavitation of the hydraulic motor can be accurately supported, and safety guarantee is provided for equipment life prediction and safety control.

Description

Technical Field

[0001] The present invention relates to a technology in the field of industrial big data and information fusion, and specifically to a hydraulic motor fault diagnosis method and system based on heterogeneous asynchronous data fusion. Background Art

[0002] As a core component for power transmission in construction machinery, the operation state monitoring of a hydraulic motor relies on the data fusion of multi-source heterogeneous sensors such as vibration, oil pressure, and temperature. However, existing heterogeneous asynchronous data fusion solutions are mostly in the theoretical research stage and have poor effects under the complex working conditions of hydraulic systems. For example, there is a sampling rate difference of more than 4 orders of magnitude between the vibration signal (10 kHz) and the oil temperature data (0.1 Hz) of a hydraulic motor. Traditional time alignment methods generate a phase lag of more than 20% within a 10 ms real-time control window. The strong noise environment of the hydraulic system results in a low signal-to-noise ratio of conventional feature extraction networks. In recent years, research has attempted to optimize through the attention mechanism, but there is still a lack of targeted design for the multi-physical field coupling characteristics of hydraulic systems. Therefore, developing a special fusion framework for hydraulic motors with real-time performance, noise resistance, and cross-model adaptability to achieve high-accuracy hydraulic motor fault diagnosis has become an urgent need for the intelligent operation and maintenance of industrial equipment. Summary of the Invention

[0003] Aiming at the problems of time series asynchrony and cross-modal data heterogeneity in the multi-source heterogeneous data fusion of hydraulic motors, the present invention proposes a hydraulic motor fault diagnosis method and system based on heterogeneous asynchronous data fusion. Through the collaborative architecture of a dynamic pruning residual network and a spatio-temporal cross-attention mechanism, it realizes the efficient spatio-temporal alignment and feature enhancement of multi-physical field coupling data such as vibration, oil pressure, and temperature, effectively overcomes the limitations of traditional fusion models in the asynchronous processing of high-frequency vibration signals and low-frequency thermodynamic data in hydraulic systems, significantly improves the robustness of key fault features in a strong noise environment, and realizes the rapid virtual-real mapping of bench test simulation data and online monitoring data, which can accurately support typical hydraulic motor faults such as plunger pair wear and valve plate cavitation, and provide safety guarantees for equipment life prediction and safety control.

[0004] The present invention realizes the hydraulic motor fault diagnosis based on heterogeneous asynchronous data through the following technical solutions:

[0005] The present invention relates to a hydraulic motor fault diagnosis method based on heterogeneous asynchronous data fusion, including:

[0006] Step 1: Collect heterogeneous asynchronous data of the sensor network and preprocess the data, specifically including:

[0007] 1.1: Collect vibration, temperature, and pressure data through sensors;

[0008] 1.2. Remove abnormal data values and fill in the missing data points using the time series interpolation method;

[0009] 1.3. Use the short-time Fourier transform to convert the time series data of the vibration sensor into a time-frequency diagram.

[0010] Step 2. Input the preprocessed data into the parallel dynamic pruning residual network to preliminarily fuse the data and extract features, obtaining an initial feature plane, specifically including:

[0011] 2.1. The parallel dynamic pruning residual network inputs different sensor data in different channels to preliminarily extract features;

[0012] 2.2. Calculate the vibration-image feature correlation matrix through the cross-channel attention mechanism to generate an initial fusion feature plane.

[0013] The described parallel dynamic pruning residual network adopts a multi-sensor signal channel heterogeneous architecture. Cross-modal interaction is carried out between the signal channels of each sensor through bidirectional feature distillation. The vibration signal channel transmits the frequency-domain attention mask to the temperature signal channel, and the temperature signal channel feeds back the temperature-frequency correlation matrix and dynamically adjusts the calculation resource allocation, including: a spectrum-sensitive gating unit deployed in the vibration signal channel and dynamically closing redundant convolution channels based on the main frequency energy.

[0014] The temperature signal channel adopts a temperature gradient-driven pruning strategy to dynamically adjust the calculation density through the gradient retention rate.

[0015] The threshold of the described closing redundant convolution channels where: N is the total number of frequency bands divided according to the vibration characteristics of the hydraulic system, f k is the center frequency of the k-th frequency band, corresponding to the gear meshing frequency of the hydraulic motor and its harmonics, E(f k ) is the energy density of the k-th frequency band calculated from the vibration signal STFT, σ(f k ) is the hydraulic fault sensitivity weight obtained from the statistics of historical fault data.

[0016] The described temperature-frequency correlation matrix where: A v,t is the vibration-temperature correlation matrix reflecting the temporal correlation strength between each frequency band and the temperature signal, Q v is the extracted vibration signal spectrum feature as the vibration feature query vector, K t , V t is the temperature feature key-value pair, and d is the feature vector dimension automatically optimized by the multi-head attention mechanism.

[0017] Step 3. After inputting the initial feature plane into the multi-connected neural network, fuse the data and extract features to obtain the final feature plane;

[0018] The multi-connected neural network described above includes: a bidirectional spatio-temporal attention module, a hierarchical feature aggregation module, and a dynamic weight decision module, where: the bidirectional spatio-temporal attention module establishes the physical coupling relationship between the vibration spectrum and the temperature spectrum, and generates a fault thermal coupling matrix through improved cross-attention; the hierarchical feature aggregation module uses multi-scale pyramid pooling to fuse macroscopic assembly errors and microscopic wear characteristics; the dynamic weight decision module constructs a two-factor gating function based on the vibration sensor readings and the housing surface temperature rise rate, dynamically allocates the fusion weights of vibration, temperature, and pressure modes, and suppresses non-physical correlation features through the temperature-vibration correlation function.

[0019] The improved cross-attention mentioned above refers to: where: A v→t is the enhanced temperature feature, W v is the learnable weight in the vibration frequency domain, W t is the learnable weight in the temperature time domain, F v is the original vibration frequency domain signal, F t is the original temperature time domain signal.

[0020] The hierarchical feature aggregation module extracts the deformation gradient and curvature distribution in the pressure signal features through polar coordinate-Cartesian dual-path convolution in parallel.

[0021] The two-factor gating function G = Sigmoid(α·C contam +β·ΔT), where: G is the sensor sensitivity adjustment coefficient, used to control the sampling frequency and signal gain, G ∈ [0,1], α is the noise weight coefficient, C contam is the noise intensity, β is the temperature change rate weight coefficient reflecting the demand for equipment temperature drift compensation, β ∈ [-0.2(℃ / min) -1 ,0.5(℃ / min) -1 ), ΔT is the temperature change rate.

[0022] The temperature-vibration correlation function mentioned above where: T is the working temperature of the hydraulic oil, f is the vibration signal frequency, u(f) is the optimal temperature of the vibration in this frequency band for the reference temperature, that is, the healthy state, and σ is the temperature distribution bandwidth parameter.

[0023] Step 4: Use the fused feature data as the training set to build a feature classification model;

[0024] The feature classification model mentioned above uses a multi-task classification head, corresponding to typical faults of the hydraulic system such as plunger pair wear and valve plate cavitation respectively. At the same time, the fault-sensitive features are strengthened through the channel-space dual-attention mechanism, and the gradient penalty term and dynamic sample weight strategy are integrated into the loss function.

[0025] Step 5. Test and evaluate the performance of the feature classification model, specifically including:

[0026] 5.1. Construct three types of datasets including normal, boundary, and fault injection;

[0027] 5.2. Quantify the evaluation metrics: detection rate > 95%, false alarm rate < 1.5%, and inference latency < 50 ms;

[0028] 5.2. Input the test sample set into the feature classification model to obtain the actual output of the feature classification model, and compare it with the ideal output of the model to obtain the performance test and evaluation results of the feature classification model. Technical effects

[0029] Based on the dynamic pruning residual network architecture of the spectrum-sensitive gating unit, the present invention dynamically closes redundant convolution channels through the main frequency energy of the vibration signal, combines the temperature gradient-driven pruning strategy and curvature feature sparsification processing, realizes the adaptive allocation of computing resources for multi-modal data, and at the same time, through the improved cross-attention calculation formula in the bidirectional spatio-temporal attention module, establishes the physical coupling relationship between the vibration spectrum and the temperature spectrum through the hydraulic condition adaptive weight t; based on the polar coordinate-Cartesian dual-path convolution architecture, realizes the cross-dimensional fusion of the pressure spectrum feature and the vibration spectrum feature, and combines the non-physical correlation feature suppression mechanism of the temperature-vibration correlation function. Compared with the prior art, the present invention reduces the phase lag from 20% to 5.2% within a 10 ms real-time control window, reduces the computational amount by the dynamic pruning strategy while maintaining 98.6% of the effective information; improves the accuracy of plunger pair wear detection to 96.3% and reduces the false alarm rate to 1.2%. Brief description of the drawings

[0030] Figure 1 It is a flowchart of the present invention;

[0031] Figure 2 It is a flowchart of the embodiment;

[0032] Figure 3 It is a schematic diagram of the system of the present invention. Detailed implementation manners

[0033] As Figure 1 and Figure 2 shown, this embodiment relates to a hydraulic motor fault diagnosis method based on heterogeneous asynchronous data fusion, including:

[0034] Step 1. Collect heterogeneous asynchronous data of the sensor network and preprocess the data, specifically including:

[0035] 1.1. Collect asynchronous data of vibration sensors, temperature sensors, and pressure sensors;

[0036] Using a data acquisition master controller equipped with a 16-bit high-precision ADC, the clock synchronization of the sensor network is achieved through the IEEE 1588 protocol, and the vibration of the hydraulic motor housing, the surface temperature field, and the oil pressure data are synchronously acquired, and the analog signals are converted into digital signals;

[0037] 1.2. Remove abnormal data values and fill in the missing data using the time series interpolation method;

[0038] For asynchronous data, obtain the timestamps of the data acquisition of each sensor, standardize the timestamps and determine the global time reference, align the timestamps according to the global time reference, use the spline interpolation method, construct different polynomials in each time interval, calculate the data values using the polynomials, and insert the calculated data into the original sequence to fill the data gaps.

[0039] 1.3. Use the short-time Fourier transform to convert the sensor time series data into a time-frequency diagram, retain the data spatial distribution characteristics, and unify the data dimensions;

[0040] Step 2. Input the preprocessed data into a parallel dynamic pruning residual network to initially fuse the data and extract features, and obtain an initial feature plane, specifically including:

[0041] 2.1. Input the data of different sensors into the parallel dynamic pruning residual network through channels to initially extract features, including: vibration signal channel: deploy a spectrum-sensitive gating unit to dynamically prune redundant channels based on the main frequency energy of the vibration signal; the temperature signal channel adopts a temperature gradient-driven pruning strategy to automatically skip large kernel calculations; pressure signal channel: activate 30% lightweight convolutional kernels and sparsify non-critical regions.

[0042] Input multi-source heterogeneous data through channels, and each channel performs cross-modal interaction through bidirectional feature distillation. The vibration signal channel transmits a frequency domain attention mask to the temperature signal channel to suppress the temperature features in the frequency bands not related to faults. The temperature signal channel feeds back the temperature-frequency correlation matrix M t→v to strengthen the analysis of the vibration frequency bands corresponding to abnormal temperature rises;

[0043] 2.2. Achieve feature fusion through a cross-channel attention mechanism;

[0044] Calculate the correlation matrix A v between the vibration feature F v,t and the temperature feature F;

[0045] Adopt channel splicing and 1×1 convolution for dimensionality reduction, and fuse the pressure feature F p ;

[0046] Output the initial feature plane F init ;

[0047] Step 3: After inputting the initial feature plane into the multi-connected neural network, fuse the data and extract features to obtain the final feature plane, specifically as follows:

[0048] Establish the physical coupling A of the vibration spectrum F v and the temperature spectrum F t through the bidirectional spatio-temporal attention module; v→t ;

[0049] Through the hierarchical feature aggregation module, first perform 64×64 pyramid pooling on the vibration features to extract the overall deformation features of the housing, perform 8×8 pooling on the pressure features to capture local abnormal oil pressure of the plunger pair, and fuse the spatial features through polar coordinate-Cartesian dual-path convolution;

[0050] Based on the oil contamination degree C contam and the shell temperature rise rate ΔT / Δt, calculate the model weights, and filter the abnormal responses through the temperature-vibration coupling function ;

[0051] Step 4: Use the fused feature data as the training set to build a feature classification model;

[0052] Step 5: Test and evaluate the performance of the feature classification model, specifically including:

[0053] 5.1. Construct three types of data sets including normal, boundary, and fault injection, covering working conditions: normal working conditions (oil temperature 40±5°C, pressure 20 MPa), boundary working conditions (oil temperature 65°C, pressure 28 MPa)

[0054] 5.2. Quantify the evaluation indicators: detection rate > 95%, false alarm rate < 1.5%, inference delay < 50 ms;

[0055] 5.2. Input the test sample set into the feature classification model to obtain the actual output of the feature classification model, and compare it with the ideal output of the model to obtain the performance test and evaluation results of the feature classification model.

[0056] Such as Figure 3As shown in the figure, the system for implementing the above method in this embodiment includes: a data acquisition and data preprocessing unit, a spatio-temporal reference unification and feature extraction unit, a cross-modal feature alignment and fusion unit, a feature classification model construction unit, and a feature classification unit. Among them: The data acquisition and data preprocessing unit acquires sensor data and preprocesses the data; the spatio-temporal reference unification and feature extraction unit unifies the spatio-temporal dimensions of the data, and through a parallel dynamic pruning residual network, extracts features and preliminarily fuses the features to obtain an initial feature plane; the cross-modal feature alignment and fusion unit inputs the initial feature plane into a multi-connected neural network to weight and fuse multi-source data; the feature classification model construction unit uses the cross-modal fused feature data as a training set to construct a feature classification model; the feature classification model classifies the features of multi-source data.

[0057] The spatio-temporal reference unification and feature extraction unit includes: a dynamic time synchronization, a spatial coordinate transformation matrix, and a cross-dimensional feature alignment module. Among them, the dynamic time synchronization compensates for the non-uniform sampling time series difference by combining cubic spline interpolation; the spatial coordinate transformation matrix is based on the short-time Fourier transform to uniformly transform the data into two dimensions; the cross-dimensional feature alignment module maps multi-modal data such as vibration signals, temperature signals, and pressure signals to a unified feature space, uses a cross-channel attention mechanism to capture the correlation between modalities, and adaptively extracts time-varying features by combining dynamic convolution kernels.

[0058] The cross-modal feature alignment and fusion unit includes: a bidirectional spatio-temporal attention module, a hierarchical feature aggregation module, and a physical coupling filtering module. Among them, the bidirectional spatio-temporal attention module establishes cross-modal associations between the vibration spectrum and the temperature spectrum through a multi-head cross-attention mechanism; the hierarchical feature aggregation module adopts a polar coordinate-Cartesian dual-path convolution architecture to extract pressure features in the polar coordinate system and realizes the splicing of dual-domain feature tensors through a spatial transformation network; the physical coupling filtering module constructs a temperature-vibration coupling function based on the oil contamination degree and the shell body temperature rise rate, and fuses multi-source features through a gating mechanism.

[0059] After specific actual experiments, in the environment of a hydraulic axial piston pump test bench (impact load condition with a rated pressure of 28 MPa and a rotational speed of 1500 r / min), the above device is run synchronously at a sampling frequency of 10 kHz for the vibration sensor and a sampling frame rate of 1 Hz for the temperature sensor. The experimental data shows that the cross-modal feature fusion unit achieves an accuracy of 96.3% in the detection of plunger pair wear, and the false alarm rate drops to 1.2%; the comparative experiment shows that the bidirectional spatio-temporal attention module improves the detection sensitivity by 12 dB, and the hierarchical aggregation module retains 98.6% of the effective information while compressing 80% of the feature dimensions.

[0060] During specific experiments, 20% of the total data is taken as the test set data to test the performance of the model. By comparing the actual output and the ideal output of the feature classification model, the performance test and evaluation results of the feature classification model can be obtained. Through this example, it can be found that when the present invention performs data fusion, it can effectively extract heterogeneous asynchronous data features. The hierarchical feature fusion can retain 98.6% of the information, and the fault diagnosis result shows that the classification accuracy of the plunger pair wear reaches 96.3%. By integrating vibration data, temperature data, and pressure data, it can extract diverse data features and improve the accuracy of feature classification.

[0061] Compared with the prior art, through the spectrum-sensitive dynamic pruning strategy, while maintaining the integrity of the 10 kHz vibration signal, the present invention reduces the feature extraction delay to 18 ms, meeting the 20 ms real-time monitoring requirement of the hydraulic system; the bidirectional spatio-temporal attention module improves the vibration-temperature feature alignment accuracy to 98.4%; the overall technical framework improves the plunger pair wear detection accuracy to 96.3% and reduces the false alarm rate to 1.2%.

[0062] The above specific implementation can be locally adjusted by those skilled in the art in different ways without departing from the principles and purposes of the present invention. The protection scope of the present invention is subject to the claims and is not limited by the above specific implementation. All implementation solutions within its scope are subject to the constraints of the present invention.

Claims

1. A hydraulic motor fault diagnosis method based on heterogeneous asynchronous data fusion, characterized in that, Including: Step 1: Collect heterogeneous asynchronous data of the sensor network and preprocess the data. Step 2: Input the preprocessed data into a parallel dynamic pruning residual network to preliminarily fuse the data and extract features, obtaining an initial feature plane. Step 3: After inputting the initial feature plane into a multi-connected neural network, fuse the data and extract features to obtain the final feature plane. Step 4: Use the fused feature data as a training set to build a feature classification model. Step 5: Test and evaluate the performance of the feature classification model.

2. The hydraulic motor fault diagnosis method based on heterogeneous asynchronous data fusion according to claim 1, characterized in that, The specific content of Step 1 includes: 1.1 Collect vibration, temperature, and pressure data through sensors. 1.2 Remove abnormal data values and fill in missing data points using the time series interpolation method. 1.3 Use the short-time Fourier transform to convert the vibration sensor time series data into a time-frequency diagram.

3. The hydraulic motor fault diagnosis method based on heterogeneous asynchronous data fusion according to claim 1, characterized in that, The specific content of Step 2 includes: 2.1 The parallel dynamic pruning residual network inputs different sensor data through channels and preliminarily extracts features. 2.2 Calculate the vibration-image feature correlation matrix through a cross-channel attention mechanism to generate an initial fused feature plane.

4. The hydraulic motor fault diagnosis method based on heterogeneous asynchronous data fusion according to claim 3, characterized in that The parallel dynamic pruning residual network adopts a multi-sensor signal channel heterogeneous architecture. Cross-modal interaction is carried out between sensor signal channels through bidirectional feature distillation. The vibration signal channel transmits a frequency domain attention mask to the temperature signal channel, and the temperature signal channel feeds back the temperature-frequency correlation matrix and dynamically adjusts the calculation resource allocation, including: a spectrum-sensitive gating unit deployed in the vibration signal channel and dynamically closing redundant convolution channels based on the main frequency energy.

5. The hydraulic motor fault diagnosis method based on heterogeneous asynchronous data fusion according to claim 3, characterized in that, The temperature signal channel adopts a temperature gradient-driven pruning strategy to dynamically adjust the calculation density through the gradient retention rate. The threshold for closing redundant convolution channels , where: N is the total number of frequency bands divided according to the vibration characteristics of the hydraulic system, f k is the center frequency of the k-th frequency band, corresponding to the gear meshing frequency of the hydraulic motor and its harmonics, E(f k ) is the energy density of the k-th frequency band calculated from the vibration signal STFT, and σ(f k ) is the hydraulic fault sensitivity weight obtained from the statistics of historical fault data; The temperature-frequency correlation matrix , where: A v,t is the vibration-temperature correlation matrix reflecting the correlation strength between each frequency band and the temperature signal time series, Q v is the spectral feature of the vibration signal obtained as the vibration feature query vector, K t , V t is the temperature feature key-value pair, and d is the feature vector dimension automatically optimized by the multi-head attention mechanism.

6. The hydraulic motor fault diagnosis method based on heterogeneous asynchronous data fusion according to claim 1, characterized in that, The multi-connected neural network includes: a bidirectional spatio-temporal attention module, a hierarchical feature aggregation module, and a dynamic weight decision module. Among them: the bidirectional spatio-temporal attention module establishes the physical coupling relationship between the vibration spectrum and the temperature spectrum, and generates a fault thermal coupling matrix through an improved cross-attention; the hierarchical feature aggregation module uses multi-scale pyramid pooling to fuse the macroscopic assembly error and microscopic wear features; the dynamic weight decision module constructs a two-factor gating function based on the vibration sensor readings and the housing surface temperature rise rate, dynamically allocates the fusion weights of the vibration, temperature, and pressure modalities, and suppresses non-physical correlation features through the temperature-vibration correlation function.

7. The hydraulic motor fault diagnosis method based on heterogeneous asynchronous data fusion according to claim 6, characterized in that The improved cross-attention mentioned above refers to: , where: A v→t is the enhanced temperature feature, W v is the learnable weight in the vibration frequency domain, W t is the learnable weight in the temperature time domain, F v is the original vibration frequency domain signal, F t is the original temperature time domain signal; The hierarchical feature aggregation module extracts the deformation gradient and curvature distribution in the pressure signal features through polar coordinate-Cartesian dual-path convolution in parallel. The described two-factor gating function , where: G is the sensor sensitivity adjustment coefficient, used to control the sampling frequency and signal gain, G ∈ [0, 1], α is the noise weight coefficient, C contam is the noise intensity, β is the temperature change rate weight coefficient reflecting the demand for device temperature drift compensation, β ∈ [−0.2 (°C / min)⁻¹, 0.5 (°C / min)⁻¹]), ΔT is the temperature change rate; The temperature-vibration correlation function described above , where: T is the working temperature of the hydraulic oil, f is the frequency of the vibration signal, u(f) is the optimal temperature of the vibration in this frequency band for the reference temperature, that is, the healthy state, and σ is the temperature distribution bandwidth parameter; The feature classification model adopts a multi-task classification head, corresponding to typical faults of the hydraulic system such as plunger pair wear and valve plate cavitation respectively. At the same time, the fault-sensitive features are strengthened through the channel-space dual attention mechanism, and the gradient penalty term and dynamic sample weight strategy are integrated into the loss function.

8. The hydraulic motor fault diagnosis method based on heterogeneous asynchronous data fusion according to claim 1, characterized in that The specific content of Step 5 includes: 5.1 Construct a three-category dataset including normal, boundary, and fault injection. 5.2 Quantify the evaluation indicators: detection rate > 95%, false alarm rate < 1.5%, inference delay < 50ms. 5.

2. Input the test sample set into the feature classification model, obtain the actual output of the feature classification model, compare it with the ideal output of the model, and obtain the performance test and evaluation results of the feature classification model.

9. A hydraulic motor fault diagnosis system based on heterogeneous asynchronous data fusion for implementing the method according to any one of claims 1-8, characterized in that, Including: Data acquisition and data preprocessing unit, spatio-temporal reference unification and feature extraction unit, cross-modal feature alignment and fusion unit, feature classification model construction unit, and feature classification unit. Among them: The data acquisition and data preprocessing unit acquires sensor data and preprocesses the data; The spatio-temporal reference unification and feature extraction unit unifies the spatio-temporal dimensions of the data, extracts features and preliminarily fuses the features through a parallel dynamic pruning residual network to obtain an initial feature plane; The cross-modal feature alignment and fusion unit inputs the initial feature plane into a multi-connected neural network to weight and fuse multi-source data; The feature classification model construction unit uses the cross-modal fused feature data as a training set to construct a feature classification model; The feature classification model classifies the features of multi-source data.

10. The hydraulic motor fault diagnosis system based on heterogeneous asynchronous data fusion according to claim 9, characterized in that, The spatio-temporal reference unification and feature extraction unit includes: dynamic time synchronization, spatial coordinate transformation matrix, and cross-dimensional feature alignment module. Among them, dynamic time synchronization combines cubic spline interpolation to compensate for the non-uniform sampling time series difference; The spatial coordinate transformation matrix is based on the short-time Fourier transform to uniformly convert the data into two dimensions; The cross-dimensional feature alignment module maps multi-modal data such as vibration signals, temperature signals, and pressure signals to a unified feature space, uses a cross-channel attention mechanism to capture the correlation between modalities, and adaptively extracts time-varying features in combination with a dynamic convolution kernel.

11. The hydraulic motor fault diagnosis system based on heterogeneous asynchronous data fusion according to claim 9, characterized in that, The cross-modal feature alignment and fusion unit includes: a bidirectional spatio-temporal attention module, a hierarchical feature aggregation module, and a physical coupling filtering module. Among them, the bidirectional spatio-temporal attention module establishes cross-modal associations between vibration spectra and temperature spectra through a multi-head cross-attention mechanism; The hierarchical feature aggregation module adopts a polar coordinate-Cartesian dual-path convolution architecture to extract pressure features in the polar coordinate system and realizes the splicing of dual-domain feature tensors through a spatial transformation network; The physical coupling filtering module constructs a temperature-vibration coupling function based on the oil contamination degree and the shell body temperature rise rate, and fuses multi-source features through a gating mechanism.

Images