Electromechanical system fault diagnosis system based on vibration law analysis

Abstract

The invention relates to the technical field of electromechanical fault diagnosis, and provides an electromechanical system fault diagnosis system based on vibration law analysis, a multi-source data fusion module is used for acquiring multi-dimensional data in an electromechanical system, fusing the multi-source data by taking the time-frequency characteristic of a vibration signal as a core, generating a weighted fusion tensor, and analyzing the weighted fusion tensor; the diagnosis model module extracts working condition invariant features based on the weighted fusion tensor generated by the multi-source data fusion module, solves the problem of model generalization and outputs fused feature vectors, and the diagnosis module performs fault diagnosis of the electromechanical system in combination with the vibration signal physical law and the feature vectors output by the diagnosis model module. And carrying out visual feedback. According to the invention, the multi-source data fusion module fuses the multi-source data for weighted tensor modeling, dual-channel feature learning of the diagnosis model module and adversarial domain adaptive gradient inversion training, efficient integration and feature extraction of the multi-source data are realized, and the problems of low information utilization rate and insufficient model generalization ability are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of electromechanical fault diagnosis, and specifically to an electromechanical system fault diagnosis system based on vibration law analysis. Background Art

[0002] The analysis of the vibration law of electromechanical equipment is one of the core means of fault diagnosis. Traditional methods in the past relied on manual feature extraction and engineering experience, such as wavelet packet transform, fast Fourier transform, etc., but were vulnerable to noise interference and inefficient under complex working conditions; for example, the process of manual feature extraction was vulnerable to background noise interference, resulting in the loss of key fault information; second, the algorithm had poor adaptability to non-stationary signals and was difficult to capture transient impact features; third, the diagnostic process had hysteresis and could not meet the real-time requirements.

[0003] In recent years, with the progress of intelligent sensing technology, fault monitoring methods based on vibration signal analysis have made breakthroughs. New systems achieve global perception of equipment status by constructing distributed sensor networks, and directly mine fault characterization information from the original vibration signals in combination with deep learning methods. Among them, the intelligent monitoring system based on high-frequency vibration sensors and big data analysis realizes the online evaluation of equipment status, and the end-to-end diagnosis method based on deep neural networks significantly improves the diagnostic accuracy through automatic feature extraction, showing important application value in industrial scenarios.

[0004] However, the prior art still has the following deficiencies: 1. There are defects in multi-source data fusion. Different types of data are difficult to effectively integrate, resulting in low information utilization rate and unable to comprehensively reflect the operating state of the electromechanical system.

[0005] 2. The model learning ability and generalization ability of feature learning are poor. The constructed models often can only be applicable to specific scenarios or data sets. When facing new working conditions, different equipment parameters or complex and changeable operating environments, the diagnostic accuracy of the models drops significantly, making it difficult to meet the diverse fault diagnosis needs in actual engineering. Summary of the Invention

[0006] To solve the above technical problems, the present invention provides an electromechanical system fault diagnosis system based on vibration law analysis, which realizes the efficient integration and feature extraction of multi-source data through the improved Hilbert-Huang transform and envelope spectrum entropy weighted tensor modeling of the multi-source data fusion module, and the adversarial domain adaptation gradient reversal training of the diagnostic model module, so as to solve the problems in the prior art.

[0007] An electromechanical system fault diagnosis system based on vibration law analysis includes: Multi-source data fusion module, which is used to obtain multi-dimensional data in the electromechanical system, take the time-frequency characteristics of vibration signals as the core, fuse multi-source data, and generate a weighted fusion tensor; Diagnostic model module, which is data-connected to the multi-source data fusion module, extracts condition-invariant features from the weighted fusion tensor generated by the multi-source data fusion module, solves the problem of model generalization, and outputs the fused feature vector; Diagnostic module, which is data-connected to the diagnostic model module and the multi-source data fusion module, combines the physical laws of vibration signals with the feature vector output by the diagnostic model module to perform fault diagnosis on the electromechanical system and provide visual feedback.

[0008] Preferably, the multi-source data fusion module includes a vibration signal processing unit and a weight calculation unit; The signal processing unit collects multi-source data during historical or laboratory faults and real-time multi-source data of the electromechanical system, uses the improved Hilbert-Huang transform on the vibration signal data to generate a time-frequency spectrum matrix as the fusion reference feature, and constructs a multi-dimensional tensor in combination with multi-source auxiliary data; The weight calculation unit, based on the constructed multi-dimensional tensor, uses the envelope spectrum entropy as a weight index to calculate the importance weight of the frequency bands of the vibration time-frequency spectrum, and performs multi-source data tensor modeling based on the frequency band importance weight, and constructs the time-frequency spectrum matrix of the vibration signal and multi-source auxiliary parameters such as temperature and current into a weighted fusion tensor.

[0009] Preferably, the signal processing unit receives the original vibration signal to generate a vibration time-frequency spectrum, where is the total number of t, represents a matrix of dimension , and the specific calculation process is as follows; First, perform CEEMDAN decomposition: decompose into n IMF components and a residual: ; where t represents the sampling time point for obtaining samples, represents the i-th IMF component, is the residual component; Hilbert transform: calculate the analytic signal for each IMF component: ; where is the Hilbert transform operator, and j represents the imaginary unit, which is used to place the signal after the Hilbert transform in the imaginary part of the complex plane; After the Hilbert transform, based on the analytic signal the instantaneous amplitude and the instantaneous phase are calculated to obtain the time-frequency spectrum matrix : ; wherein, represents the frequency variable, that is, the set of continuous frequency variables or all discrete frequency components in the spectrum analysis. When f has a subscript, that is represents the specific frequency value of the i-th discretization, which is determined by the Fourier transform resolution; Finally, multi-source data alignment is performed. Linear interpolation is performed on the low-frequency signals of temperature and current and aligned to the time stamps of the vibration signals to construct a multi-dimensional tensor : where D represents the dimension containing the frequency, amplitude, temperature, and current of the vibration time-frequency spectrum.

[0010] Preferably, the weight calculation unit uses the envelope spectrum entropy as a single weight index to simultaneously characterize the impact characteristics and frequency domain complexity of the vibration signal. The specific calculation process is as follows: Envelope signal extraction: Perform Hilbert envelope demodulation on the original vibration signal to obtain the envelope signal : ; Then, perform envelope spectrum entropy calculation, and calculate the Fourier spectrum of the envelope signal ; Define the probability of spectrum energy distribution : ; Calculate the spectrum entropy H: ; Calculate the band weight of : ; wherein, strengthen the high-energy frequency band, corresponding to the fault characteristic frequency, suppress broadband noise. A high entropy value indicates a dispersed spectrum, and a low entropy value indicates concentrated energy; Then scale the dimension of the vibration time-frequency spectrum: ; Output the weighted fusion tensor and the original time-frequency spectrum .

[0011] Preferably, the diagnostic model module includes a feature extraction unit and a generalization enhancement unit; The feature extraction unit is used to extract time-frequency features through a dual-channel 1DCNN based on the weighted data tensor generated by the multi-source data fusion module, and obtain fault-related features. Among them, the first channel extracts resonance band features through frequency-domain slices of the vibration time-frequency spectrum, and the second channel extracts impact waveform features through the time-domain vibration amplitude sequence; The generalization enhancement unit is used to process the feature data of the multi-source data based on the history of the electromechanical system or laboratory faults extracted by the feature extraction unit, establish and enhance the generalization ability of the features to adapt to different working conditions and data distributions.

[0012] Preferably, the feature extraction unit first inputs a historical dataset of multi-source data based on the history of the electromechanical system or laboratory faults, including vibration signals and fault labels under multiple working conditions; Perform dual-channel convolution training calculations. The frequency-domain channel learns resonance patterns, and the time-domain channel learns impact waveforms. The specific process of feature extraction through a dual-channel 1DCNN is as follows: Channel 1, frequency-domain features: Input , convolution kernel : ; Channel 2, time-domain features: Input the vibration amplitude sequence , convolution kernel : ; Among them, amp represents the amplitude, h represents the output feature, represents the resonance feature matching degree at the t-th at time t, and the output feature represents the impact feature matching degree at time t, represents the sliding time offset of the convolution kernel, represents the bias term, represents the bias term in the calculation based on frequency-domain features, That is, it represents the bias term in the calculation based on time-domain features, k represents the time width of the convolution kernel, is the non-linear activation function; Perform attention mechanism fusion to generate attention weights : ; Perform feature fusion: ; Among them, is the Sigmoid function, is the weight of the fully connected layer, Represents the bias term in the attention mechanism fusion calculation, and finally outputs the fused feature vector. .

[0013] Preferably, the generalization enhancement unit performs adversarial domain adaptation on the feature extraction unit to solve the generalization problem. Use the device data from the laboratory or known working conditions as the source domain data, and use the new device or variable working condition data from the field as the target domain data. Adversarial training: Train the feature extraction unit through gradient reversal to extract fault features independent of working conditions. In the real-time inference stage, directly input the weighted fusion tensor based on real-time multi-source data, use the trained feature extraction unit to extract features, output the weighted fusion tensor based on real-time multi-source data, and calculate the fused feature vector. .

[0014] Preferably, the diagnosis module generates a diagnosis result based on the feature vector and the vibration physical law. First, obtain the fault feature vector output by the diagnosis model module. And the original vibration time-frequency spectrum generated by the multi-source data fusion module. Fault classification and location: Construct a normal state hypersphere through support vector data description, calculate the distance from the feature vector to the hypersphere boundary to achieve anomaly detection, and then combine the vibration signal characteristic frequency library or the bearing fault characteristic frequency calculation formula to perform frequency band deconvolution on the abnormal feature vector to locate the fault component.

[0015] Preferably, the diagnosis module can also perform visual feedback, overlay the fault frequency band energy heat map on the original time-frequency spectrum, mark the position deviation between the characteristic frequency and the actual resonance peak, and output the fault type, position confidence, and time-frequency domain visualization map.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. In the present invention, the signal processing unit in the multi-source data fusion module uses the improved Hilbert-Huang transform for the vibration signal to generate the time-frequency spectrum matrix as the fusion reference feature, constructs a multi-dimensional tensor in combination with multi-source auxiliary data, and the weight calculation unit uses the envelope spectrum entropy as the weight index to calculate the frequency band importance weight of the vibration time-frequency spectrum and perform multi-source data tensor modeling, realizing the effective integration of multi-source data of different types, different formats, different sampling frequencies, and dimensions, and solving the problems of low information utilization rate and inability to comprehensively reflect the operating state of the electromechanical system caused by the difficulty of effectively integrating multi-source data.

[0017] 2. In the present invention, the generalization enhancement unit in the diagnostic model module performs adversarial domain adaptation on the feature extraction unit. The device data from the laboratory or known working conditions is used as the source domain data, and the new device or variable working condition data from the field is used as the target domain data. The feature extraction unit is trained by gradient reversal to extract fault features independent of the working conditions, thereby enhancing the generalization ability of the features. This solves the problems that the model learning ability and generalization ability of feature learning are poor, the diagnostic accuracy decreases when facing new working conditions, different device parameters, or complex and variable operating environments, and it is difficult to meet the diverse fault diagnosis requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a schematic diagram of the data connection process when the modules and units of the present invention are working; Figure 2 is a schematic diagram of the process of calculating the weighted fusion tensor by the multi-source data fusion module of the present invention; Figure 3 is a schematic diagram of the calculation process of the diagnostic model module of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0019] The following further describes in detail the embodiments of the present invention with reference to the drawings and embodiments. The following embodiments are used to illustrate the present invention, but cannot be used to limit the scope of the present invention.

[0020] The present invention provides an electromechanical system fault diagnosis system based on vibration law analysis, including: A multi-source data fusion module, which is used to obtain multi-dimensional data in the electromechanical system, take the time-frequency characteristics of the vibration signal as the core, fuse the multi-source data, and generate a weighted fusion tensor; A diagnostic model module, which is data-connected to the multi-source data fusion module, extracts working condition-invariant features from the weighted fusion tensor generated by the multi-source data fusion module, solves the model generalization problem, and outputs a fused feature vector; A diagnostic module, which is data-connected to the diagnostic model module and the multi-source data fusion module, combines the physical laws of the vibration signal with the feature vector output by the diagnostic model module to perform fault diagnosis of the electromechanical system and provide visual feedback. Embodiment 1:

[0021] As Figures 1-3 shown, in this embodiment, a large railway station building covers key electromechanical equipment such as passenger elevators, escalators, ventilation and air-conditioning units, distribution transformers, and automatic door systems, and faces challenges such as high-frequency start and stop of equipment, periodic fluctuations in passenger flow, and complex environmental noise on a daily basis.

[0022] Traditional monitoring systems are difficult to capture the characteristics of equipment fatigue accumulation due to the scattered deployment of sensors and isolated data. There was once a situation where the local area was closed due to the fracture of the drive chain, seriously affecting the passenger passage efficiency.

[0023] At the same time, this station building faces extreme passenger flow pressure during special periods such as the Spring Festival travel rush and Golden Week. The mechanical and electrical equipment needs to operate continuously at full load, and there are also temporary changes in the operation mode. Traditional models are prone to misjudgment when the working conditions change drastically and cannot adapt to the parameter differences of temporarily enabled equipment, resulting in misallocation of operation and maintenance resources. Therefore, a fault diagnosis system for mechanical and electrical systems based on vibration law analysis of the present invention is introduced for fault diagnosis.

[0024] First, the signal processing unit collects multi-source data and real-time multi-source data during historical or laboratory faults of the mechanical and electrical system of the railway station building, uses the improved Hilbert-Huang transform for the vibration signal data to generate a time-frequency spectrum matrix as the fusion reference feature, and constructs a multi-dimensional tensor in combination with multi-source auxiliary data; The signal processing unit receives the original vibration signal to generate a vibration time-frequency spectrum, where, is the total number of t, represents a matrix with dimension , and the specific calculation process is as follows; Perform CEEMDAN decomposition: Decompose into n IMF components and a residual: ; where, t represents the sampling time point for obtaining samples, represents the i-th IMF component, is the residual component; Hilbert transform: Calculate the analytic signal for each IMF component: ; where, is the Hilbert transform operator, j represents the imaginary unit, and is used to place the signal after Hilbert transform in the imaginary part of the complex plane; After the Hilbert transform, based on the analytic signal calculate the instantaneous amplitude and instantaneous phase to obtain the time-frequency spectrum matrix : ; where, represents the frequency variable, that is, the set of continuous frequency variables or all discrete frequency components in spectrum analysis. When f has a subscript, that is, Represents the specific frequency value of the $i$-th discretization, which is determined by the Fourier transform resolution; Finally, multi-source data alignment is performed. Linear interpolation is carried out on the low-frequency signals of temperature and current, and they are aligned to the time stamps of the vibration signals to construct a multi-dimensional tensor : Among them, $D$ represents the dimensions including the frequency, amplitude, temperature, and current of the vibration time-frequency spectrum.

[0025] The weight calculation unit uses the envelope spectrum entropy as a single weight index, which simultaneously characterizes the impact characteristics and frequency-domain complexity of the vibration signal. The specific calculation process is as follows: Envelope signal extraction is performed on the original vibration signal to obtain the envelope signal through Hilbert envelope demodulation : ; Then, envelope spectrum entropy calculation is carried out. The Fourier spectrum of the obtained envelope signal is calculated ; The probability distribution of the spectrum energy is defined : ; The spectrum entropy $H$ is calculated: ; Calculate the band weight of : ; Among them, strengthen the high-energy frequency bands, corresponding to the fault characteristic frequencies, suppress the broadband noise. A high entropy value indicates a dispersed spectrum, and a low entropy value indicates concentrated energy; Then, the vibration time-frequency spectrum dimension is scaled: ; Output the weighted fusion tensor and the original time-frequency spectrum .

[0026] By performing CEEMDAN decomposition on the vibration signals of equipment such as escalator drive chains and elevator traction machines, a time-frequency spectrum matrix is generated as the fusion reference feature to solve the problem of weak early fault signals in a strong noise environment.

[0027] Vibration, current, rotational speed, temperature, humidity and other multi-source data are synchronously collected. The low-frequency data is aligned to the time stamps of the vibration signals through linear interpolation, and a multi-dimensional tensor including frequency, amplitude and operating parameters is constructed to achieve the spatio-temporal unity of multi-dimensional data.

[0028] By calculating the importance weights of each frequency band of the vibration signal using envelope spectrum entropy, the fault feature frequency band is strengthened, such as suppressing the environmental noise frequency band, to achieve the effect of improving the signal-to-noise ratio.

[0029] The diagnostic model module includes a feature extraction unit and a generalization enhancement unit; The feature extraction unit is used to perform time-frequency feature extraction through a two-channel 1DCNN based on the weighted data tensor generated by the multi-source data fusion module, and obtain fault-related features. Among them, the first channel extracts resonance frequency band features through the frequency domain slices of the vibration time-frequency spectrum, and the second channel extracts impact waveform features through the time-domain vibration amplitude sequence; The feature extraction unit first inputs a multi-source data historical dataset based on the history of the electromechanical system or laboratory faults, including vibration signals and fault labels under multiple working conditions; Perform two-channel convolution training calculations. The frequency domain channel learns resonance patterns, and the time domain channel learns impact waveforms. The specific process of feature extraction through a two-channel 1DCNN is as follows: Channel 1, frequency domain features: Input , convolution kernel : ; Channel 2, time domain features: Input vibration amplitude sequence , convolution kernel : ; Among them, amp represents amplitude, h represents the output feature, represents the resonance feature matching degree at the t-th at time t, and the output feature represents the impact feature matching degree at time t, represents the sliding time offset of the convolution kernel, represents the bias term, represents the bias term in the calculation based on frequency domain features, That is, it represents the bias term in the calculation based on time domain features, k represents the time width of the convolution kernel, is a non-linear activation function; Perform attention mechanism fusion to generate attention weights : ; Perform feature fusion: ; Among them, is the Sigmoid function, is the weight of the fully connected layer, represents the bias term in the calculation of attention mechanism fusion, and finally outputs the fused feature vector 。

[0030] The generalization enhancement unit performs adversarial domain adaptation on the feature extraction unit to solve the generalization problem; Use the device data from the laboratory or known working conditions as the source domain data, and use the new device or variable working condition data from the field as the target domain data; Adversarial training: Train the feature extraction unit through gradient reversal to extract fault features independent of the working conditions; In the real-time inference stage, directly input the weighted fusion tensor based on real-time multi-source data, use the trained feature extraction unit to extract features, output the weighted fusion tensor based on real-time multi-source data, and calculate the fused feature vector 。

[0031] Feature extraction through dual-channel 1DCNN: Construct a frequency domain channel to extract resonance band features, such as the gearbox meshing frequency, and a time domain channel to extract impact waveform features, such as the instantaneous amplitude mutation of bearing spalling, and automatically learn the deep features of the device health state.

[0032] The diagnosis module generates a diagnosis result based on the feature vector and the vibration physical law; First, obtain the fault feature vector output by the diagnosis model module and the original vibration time-frequency spectrum generated by the multi-source data fusion module; Fault classification and location: Construct a normal state hypersphere through support vector data description, calculate the distance from the feature vector to the hypersphere boundary to achieve anomaly detection, and then combine the vibration signal characteristic frequency library or the bearing fault characteristic frequency calculation formula to perform frequency band deconvolution on the abnormal feature vector to locate the fault component.

[0033] Construct an SVDD hypersphere based on historical normal data, calculate the distance from the current feature vector to the hypersphere boundary in real time to achieve quantitative detection of device anomalies; combine the vibration characteristic frequency library to locate the fault component through frequency band deconvolution.

[0034] At the same time, the diagnosis module can also perform visual feedback, overlay the fault frequency band energy heat map on the original time-frequency spectrum, mark the position deviation between the characteristic frequency and the actual resonance peak, and output the fault type, position confidence, and time-frequency domain visualization map.

[0035] The embodiments of the present invention are given for the purpose of illustration and description. Although the embodiments of the present invention have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those of ordinary skill in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention.

Claims

1. A mechanical and electrical system fault diagnosis system based on vibration law analysis, characterized in that: include: A multi-source data fusion module, which is used to obtain multi-dimensional data in the electromechanical system, fuse the multi-source data with the time-frequency characteristics of the vibration signal as the core, and generate a weighted fusion tensor; A diagnosis model module, wherein the diagnosis model module is data-connected with the multi-source data fusion module, extracts working condition invariant features from the weighted fusion tensor generated by the multi-source data fusion module, solves the model generalization problem, and outputs a fused feature vector; The diagnosis module is connected with the diagnosis model module and the multi-source data fusion module, combines the physical law of the vibration signal with the characteristic vector output by the diagnosis model module, performs fault diagnosis of the electromechanical system, and performs visual feedback.

2. The electromechanical system fault diagnosis system based on vibration law analysis as claimed in claim 1, characterized in that: The multi-source data fusion module includes a vibration signal processing unit and a weight calculation unit; The signal processing unit collects multi-source data and real-time multi-source data of the electromechanical system in history or laboratory failure, uses the improved Hilbert-Huang transform to generate a time-frequency spectrum matrix for the vibration signal data as a fusion benchmark feature, and constructs a multi-dimensional tensor in combination with multi-source auxiliary data; The weight calculation unit calculates the frequency band importance weights of the vibration time-frequency spectrum based on the constructed multidimensional tensor and adopts the envelope spectrum entropy as the weight index, and performs multi-source data tensor modeling based on the frequency band importance weights, and constructs the time-frequency spectrum matrix of the vibration signal and the multi-source auxiliary parameters of temperature and current into a weighted fusion tensor.

3. A mechanical and electrical system fault diagnosis system based on vibration law analysis as claimed in claim 2, characterized in that: The signal processing unit receives the original vibration signal Generate a vibration time spectrum where is the total number of t, The dimension is The specific calculation process is as follows: First, perform CEEMDAN decomposition: Decomposed into n IMF components and residuals: ； Among them, t represents the sampling time point of obtaining the sample, represents the i-th IMF component, is the residual component; Hilbert transform: Calculate the analytical signal for each IMF component : ； in, is the Hilbert transform operator, j represents the imaginary unit, which is used to place the Hilbert transformed signal in the imaginary part of the complex plane; After the Hilbert transform is completed, based on the analytical signal The instantaneous amplitude and instantaneous phase Calculate the time-frequency spectrum matrix : ； in, It represents the frequency variable, that is, the continuous frequency variable or the set of all discrete frequency components in spectrum analysis. When f has a subscript, that is, Represents the specific frequency value of the i-th discretization, which is determined by the Fourier transform resolution; Finally, multi-source data alignment is performed, and the temperature and current low-frequency signals are linearly interpolated and aligned to the vibration signal timestamp to construct a multidimensional tensor. : Where D represents the dimension including the spectral frequency, amplitude, temperature and current during vibration.

4. The electromechanical system fault diagnosis system based on vibration law analysis as claimed in claim 3, characterized in that: The weight calculation unit uses envelope spectrum entropy as a single weight index to characterize the impact characteristics and frequency domain complexity of the vibration signal. The specific calculation process is as follows: Envelope signal extraction for original vibration signal Perform Hilbert envelope demodulation to obtain the envelope signal : ； Then, the envelope spectrum entropy is calculated to obtain the envelope signal The Fourier spectrum of ; Defining the probability of spectrum energy distribution : ； Calculate the spectral entropy H: ； calculate Band weight : ； in, Strengthen the high energy frequency band, corresponding to the fault characteristic frequency, Suppress broadband noise. A high entropy value indicates a dispersed spectrum, while a low entropy value indicates concentrated energy. Then scale the vibration spectrum dimension: ； Output weighted fusion tensor and the original time spectrum .

5. The electromechanical system fault diagnosis system based on vibration law analysis as claimed in claim 1, characterized in that: The diagnostic model module includes a feature extraction unit and a generalization enhancement unit; The feature extraction unit is used to extract time-frequency features based on the weighted data tensor generated by the multi-source data fusion module by establishing a dual-channel 1DCNN to obtain fault-related features, wherein the first channel extracts the resonance frequency band features through the frequency domain slices of the vibration time-frequency spectrum, and the second channel extracts the impact waveform features through the time domain vibration amplitude sequence; The generalization enhancement unit is used to process the feature data extracted by the feature extraction unit based on multi-source data of the electromechanical system history or laboratory failure, and to establish and enhance the generalization ability of the feature to adapt to different working conditions and data distribution.

6. A mechanical and electrical system fault diagnosis system based on vibration law analysis as claimed in claim 5, characterized in that: The feature extraction unit first inputs a historical data set of multi-source data based on the history of the electromechanical system or laboratory failure, including vibration signals and fault labels of multiple working conditions; Perform dual-channel convolution training calculations, learn the resonance mode in the frequency domain channel, and learn the impact waveform in the time domain channel. The specific process of feature extraction by establishing a dual-channel 1DCNN is as follows: Channel 1, frequency domain characteristics: enter , convolution kernel : ； Channel 2, time domain features: input vibration amplitude sequence , convolution kernel : ； Among them, amp represents the amplitude, h represents the output characteristics, Represents the matching degree of the tth resonance feature at time t, and the output feature represents the impact feature matching degree at time t, Represents the sliding time offset of the convolution kernel, represents the bias term, Represents the bias term in the frequency domain feature calculation, That is, it represents the bias term in the time domain feature calculation, k represents the time width of the convolution kernel, is a nonlinear activation function; Perform attention mechanism fusion to generate attention weights : ； Perform feature fusion: ； in, is the Sigmoid function, is the weight of the fully connected layer, Represents the bias term in the attention mechanism fusion calculation, and finally outputs the fused feature vector .

7. The electromechanical system fault diagnosis system based on vibration law analysis as claimed in claim 6, characterized in that: The generalization enhancement unit performs adversarial domain adaptation on the feature extraction unit to solve the generalization problem; The equipment data from the laboratory or known working conditions is used as the source domain data, and the new equipment or variable working condition data from the field is used as the target domain data; Adversarial training: Extract fault features that are unrelated to the working conditions by training the feature extraction unit through gradient reversal; In the real-time inference stage, the weighted fusion tensor based on real-time multi-source data is directly input, the features are extracted using the trained feature extraction unit, the weighted fusion tensor based on real-time multi-source data is output, and the fused feature vector is calculated. .

8. The electromechanical system fault diagnosis system based on vibration law analysis as claimed in claim 1, characterized in that: The diagnostic module generates a diagnostic result based on the characteristic vector and the physical law of vibration; First, obtain the fault feature vector output by the diagnosis model module and the original vibration time-frequency spectrum generated by the multi-source data fusion module; Fault classification and location: construct a normal state hypersphere through support vector data description, calculate the distance from the feature vector to the hypersphere boundary to achieve anomaly detection, and then combine the vibration signal characteristic frequency library or the bearing fault characteristic frequency calculation formula to perform frequency band deconvolution on the abnormal feature vector to locate the faulty component.

9. A mechanical and electrical system fault diagnosis system based on vibration law analysis as claimed in claim 8, characterized in that: The diagnostic module can also provide visual feedback, superimpose the energy heat map of the fault frequency band on the original time-frequency spectrum, mark the position deviation between the characteristic frequency and the actual resonance peak, and output the fault type, position confidence and time-frequency domain visualization map.

Images