Electromechanical system fault diagnosis system based on deep learning

Abstract

The invention relates to the technical field of deep learning algorithms, and provides an electromechanical system fault diagnosis system based on deep learning, and the system is characterized in that a data collection and preprocessing module collects the operation data of an electromechanical system in real time, and carries out the dynamic window length setting, cleaning, noise reduction and standardization processing on the operation data; the interpretable deep learning diagnosis module carries out feature screening and decoupling learning by means of causal gating and a double-branch network, extracts fault related features and generates a diagnosis result containing a causal path and an abnormal prompt, and the physical constraint fusion module obtains physical principle data of the electromechanical system and parameter data of the electromechanical system in normal work. Carrying out physical constraint on the interpretable deep learning diagnosis module through an electromechanical system physical principle; according to the method, causal feature screening, dynamic causal mask generation and path extraction, multi-physical field law modeling and physical constraint injection are fused, and time sequence instantaneous causal analysis is combined, so that causal dominant expression and physical logic consistency guarantee of fault diagnosis is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of deep learning algorithms, and more specifically, to an electromechanical system fault diagnosis system based on deep learning. Background Art

[0002] The electromechanical system fault diagnosis technology has gone through the development stages from rule-based, signal processing to shallow machine learning. Early methods relied on expert experience to build rule bases, such as expert systems, or extracted fault features through signal processing techniques, such as Fourier transform and wavelet analysis.

[0003] With the increase in data volume, shallow machine learning methods such as support vector machines and random forests have been introduced, but still require manual feature design and have limited processing capabilities for complex non-linear problems.

[0004] Deep learning automatically extracts data features through multi-layer neural networks, solving the limitations of traditional methods relying on manual feature engineering. For example, convolutional neural networks are good at processing images and time series data and are often used in motor vibration signal analysis; recurrent neural networks capture the dynamic features of time series and are applied to power system dynamic fault diagnosis; autoencoder networks extract latent features through unsupervised learning and are used for noise reduction and feature extraction of complex equipment.

[0005] However, the existing technologies still have the following deficiencies: 1. Black box characteristics. Existing deep learning models are insufficient when explaining the causes of faults. For example, for "air conditioner shutdown", whether it is a sensor fault or a compressor damage, the existing deep learning models cannot make a clear judgment, making it difficult for maintenance personnel of China Railway to make quick decisions and still relying on manual re-inspection, resulting in delays in emergency repair time and affecting the operation safety of high-speed railway station buildings.

[0006] 2. Most existing deep learning models in the prior art rely on data-driven and lack the integration of physical principles of electromechanical systems, such as mechanical dynamics and thermodynamics laws, resulting in diagnostic results that may violate physical common sense. For example, the model may determine that "the sudden drop in bearing temperature is a fault", but in actual working conditions, such a phenomenon is physically impossible without the intervention of a cooling system. Summary of the Invention

[0007] To solve the above technical problems, the present invention provides an electromechanical system fault diagnosis system based on deep learning, which realizes the causal explicit expression of fault diagnosis and ensures the physical logic consistency through causal feature screening and decoupling, dynamic causal mask generation and path extraction, integrating multi-physical field law modeling and residual constraint injection, combined with time-series instantaneous causal analysis and feature purification, so as to solve the problems in the existing technology.

[0008] The technical solution of the present invention is as follows: A fault diagnosis system for electromechanical systems based on deep learning, comprising: A data acquisition and preprocessing module, which is used to collect historical operation data and real-time operation data of the electromechanical system, and perform dynamic window length setting, cleaning, noise reduction and standardization processing on them; An interpretable deep learning diagnosis module, which is data-connected to the data acquisition and preprocessing module, performs feature screening and decoupling learning by means of causal gating and a dual-branch network, extracts fault-related features, and finally generates a diagnosis result including a causal path and an anomaly prompt, outputting a diagnosis basis while realizing fault diagnosis; A physical constraint fusion module, which is data-connected to the interpretable deep learning diagnosis module, obtains the physical principle data of the electromechanical system and the parameter data under normal operation of the electromechanical system, and physically constrains the interpretable deep learning diagnosis module through the physical principles of the electromechanical system to ensure that the diagnosis result conforms to physical laws; A decision-making interaction module, which is data-connected to the interpretable deep learning diagnosis module, is used to receive the result and explanation information output by the interpretable deep learning diagnosis module, present them to the operation and maintenance personnel, provide an interaction interface, support manual re-inspection and feedback, and optimize the diagnosis system.

[0009] Preferably, the data acquisition and preprocessing module includes a data acquisition unit and a data preprocessing unit; The data acquisition unit includes vibration sensors, temperature sensors, current sensors, voltage sensors and humidity sensors, which are used to collect real-time operation data of the electromechanical system. At the same time, the data acquisition unit obtains historical electromechanical system operation data and fault data through the historical operation log of the electromechanical system; The data preprocessing unit uses the sliding window technology to divide the time series data, sets the dynamic window length according to different device characteristics, uses the adaptive filtering algorithm to eliminate the noise interference in the data, and combines the normalization method to unify the data scale to ensure the data quality and consistency.

[0010] Preferably, the interpretable deep learning diagnosis module includes a feature calculation unit and an interpretable diagnosis unit; The feature calculation unit generates a sparse causal graph by fusing temporal and instantaneous causal analysis, combines causal gating and a dual-branch network to achieve feature screening and decoupling learning, and completes the effective extraction of fault-related features; The interpretable diagnosis unit outputs a fault probability distribution, generates a contribution heat map based on the fused features and dynamic causal masks, and identifies unknown anomalies through clustering analysis, and finally generates a diagnosis report including a causal path and an anomaly prompt.

[0011] Preferably, the feature calculation unit includes an offline calculation subunit and an online calculation subunit; The offline calculation subunit receives the input multi-sensor time-series data matrix and the fault label y, where represents the set of real numbers, T×D represents the dimension of the matrix, T is the total length of the time series, and D is the sensor dimension; Adopting a time-domain causal discovery algorithm, fusing time-series causality and instantaneous causality analysis, for each sensor data , using a 1D causal convolution kernel sliding window, calculate its causal strength score with the fault label marked based on binary faults: ; where represents the causal strength score of the sensor data with respect to the fault label . The smaller the score, the stronger the causality. is the time-series causality term, is the instantaneous causality term, represents the balance factor, used to control the weight ratio of the time-series causality term and the instantaneous causality term, through cross-validation; The time-series causality term is specifically calculated as: ; where the smaller the score, the stronger the causality. l represents the length of the causal convolution kernel, that is, the size of the time window. represents the weight parameter of the learnable convolution kernel k, used to weight the historical signal , represents the fault label at time step , 0 or 1, 0 indicates no fault, 1 indicates a fault, i represents the i-th sensor, and j represents the j-th fault; The instantaneous causality term represents the mutual information between the sensor data and the fault label , used to quantify the instantaneous statistical dependence relationship between the two, independent of the time order. The calculation formula is: ; where x represents the discretized value of a certain sensor signal, and y represents a certain value state of the fault label, that is, 0 or 1; represents the sensor data being in x and the fault being in y joint probability, represents the marginal probability of the sensor data being in x, Indicates a fault label In the state Marginal probability; After calculating the score Then, through significance testing, i.e., permutation testing, false associations are filtered to generate a sparse causal graph G, where the nodes are sensor i and fault j, and the edges are significant causal links ; Then the causal graph G is converted into a binary mask matrix , where C is the number of fault categories, Indicates that sensor i has a causal contribution to fault j; then at each specified training cycle, the causal graph is recalculated and M is updated to adapt to the change in data distribution; The online calculation subunit receives the input, sensor data and the mask matrix M, and performs channel-level screening on the original features through the causal gating layer: ; Among them, Represents the Hadamard product, forcing the weights of irrelevant channels to zero. At the same time, only the sensor signals causally related to the current fault category are retained for subsequent calculations; Then, decoupled feature learning is performed using a dual-branch network structure: Causal branch, processing , and extracting temporal causal features H through causal convolution; Non-causal branch, processing data , and extracting global features through ordinary CNN ; Output the fused feature .

[0012] Preferably, the interpretable diagnosis unit includes an inference diagnosis subunit and a report generation subunit; The inference diagnosis subunit outputs a fault probability distribution based on the fused feature F and the dynamic causal mask M ; ; Among them, Is the weight matrix, with the dimension of the number of fault categories multiplied by the dimension of the fused feature F, b represents the bias term, and each category corresponds to a scalar bias, which is trained based on the historical operation data and fault data of the electromechanical system; the significance traceability of the gradient of the causal branch generates a fault contribution heat map A, that is, the fault To the causal feature The sum of the time dimension gradients; ; Then, for the non-causal branch features Perform K-means clustering. If the current sample deviates from the normal cluster center by more than the threshold, mark this sample with a non-causal anomaly flag and trigger an unknown anomaly warning; The report generation subunit is based on the failure probability , the contribution degree heat map A, and the non-causal anomaly flag. Combine with the dynamic causal graph G to extract the top-K causal paths of the fault category j. If a non-causal anomaly warning is triggered, append the prompt: "An unverified feature pattern is detected. It is recommended to manually check the device anomaly"; Preferably, the physical constraint fusion module includes a physical modeling unit, a physical optimization unit, and a physical verification unit; The physical modeling unit symbolizes physical laws by integrating the multi-physical field principles of the electromechanical system, generates a differentiable computational graph, uses PINN to handle complex physical relationships, and realizes cross-field coupling through tensor operations. Finally, it outputs the DPCU Library that supports gradient backpropagation, laying a foundation for subsequent calculations; The physical optimization unit injects dynamic physical constraints in real time during model training, performs physical state estimation based on the original sensor data and the DPCU Library, calculates the absolute residual between the actual and theoretical physical quantities, and dynamically adjusts the constraint weights by combining a sliding window and KL divergence to ensure that the model training conforms to physical laws; The physical verification unit verifies the physical feasibility of the prediction results based on the fault types and absolute residuals predicted by the model, and judges the validity of the results by retrieving the violated physical rule set; for infeasible results, screen high-probability candidate faults to correct the prediction. If it cannot be corrected, trigger an alarm and trace the suspicious sensor data to provide a basis for system maintenance; Preferably, the physical modeling unit inputs the multi-physical field principles of the electromechanical system, transforms physical laws into symbolic expressions, and uses the SymPy integration of the automatic differentiation tool PyTorch to generate a differentiable computational graph; For complex physical relationships that are difficult to explicitly express, use the physics-informed neural network PINN for surrogate modeling to ensure differentiability, construct an inter-field coupling relationship graph, define the interaction rules of physical quantities, and realize cross-field calculations through tensor operations. Finally, output the differentiable physical calculation unit library DPCU Library and support gradient backpropagation; Preferably, the physical optimization unit injects dynamic physical constraints during the online training of the feature calculation unit; First, input the original sensor data X and the DPCU Library for physical state estimation, and use the DPCU to calculate the derivative physical quantities from X , and at the same time, calculate the theoretical physical quantities based on the physical model , where z represents the number of types of physical quantities; Physical residual generation: Calculate the absolute residual between the actual and the theoretical: ; And statistically analyze the residual distribution through a sliding window , adaptively relax the constraints, and design dynamic constraint weights: ; Among them, KL represents the KL divergence, which is used to measure the difference degree between two distributions, represents the scaling factor, which controls the difference sensitivity, is the Sigmoid function, is the residual distribution under normal working conditions. The more the residual distribution deviates from the normal, the larger it is, and the stricter the physical constraints are; The actual application process is as follows: Sliding window statistics. Every specified period of time, calculate the distribution of the residuals of each physical quantity within the current window , KL divergence calculation. Calculate the KL divergence between and the pre-stored , and convert the KL divergence into a weight through the Sigmoid function . The greater the difference, the closer it is to 1, and the stronger the physical constraints are; Preferably, the physical verification unit is based on the absolute residual R when the model predicts the fault types y and t; Physical feasibility verification. According to the fault type y, retrieve the set of physical rules that it must violate ; If there is that is not satisfied, it is determined that the prediction result y is physically infeasible; Multi-hypothesis correction. Start the physically guided candidate set search: Screen the top three fault probabilities from the fault probability distribution and select the fault types that meet ; If no candidate is satisfied, trigger an unknown fault alarm, prompt manual intervention, physical parameter tracing. For physically infeasible predictions, trace back the sensor data that causes contradictions in reverse, and generate a list of suspicious sensors for maintenance reference; Preferably, the decision-making interaction module displays the contribution heat map, causal path map and original signal comparison through a Web interface, as well as the additional prompts generated in the report generation subunit.

[0013] Compared with the prior art, the present invention has the following beneficial effects: 1. Through the decoupling of feature screening between the causal gating and the dual-branch network in the interpretable deep learning diagnosis module, the generation of a dynamic causal mask to produce a heat map of fault contributions, and the extraction of causal paths, the present invention realizes the explicit presentation of the causal relationship between sensor signals and fault types during the fault diagnosis process, achieves a clear distinction between sensor faults and specific failure modes in a fault scenario, enables maintenance personnel to quickly locate the root cause of the fault based on the causal path and the contribution degree heat map, avoids delays in manual re-inspection, and improves the emergency repair efficiency and operation safety of the electromechanical system in the high-speed railway station building.

[0014] 2. Through the physical modeling unit in the physical constraint fusion module that transforms multi-physical field laws into differentiable computational graphs, the physical optimization unit that dynamically injects physical residual constraints, and the physical verification unit that verifies the prediction results based on physical rules, the present invention realizes the deep coupling of physical laws such as mechanical dynamics and thermodynamics with the deep learning model, achieves the automatic identification and correction of diagnostic results that violate physical common sense, ensures the physical feasibility of the diagnostic results in line with the actual working conditions, and avoids misoperations in maintenance caused by model misjudgment.

[0015] 3. Through the generation of a sparse causal graph by temporal and instantaneous causal analysis in the feature calculation unit, the filtering of non-causal associated features by the causal gating layer, and the decoupling of temporal causal features and global features by the dual-branch network, the present invention realizes the directional screening and efficient extraction of fault-related features in multi-sensor data such as vibration and temperature, achieves the feature purification effect of removing noise interference and irrelevant signals, improves the model's ability to capture early fault features, and significantly enhances the accuracy and robustness of the electromechanical system fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is a schematic diagram of the working process connection between the units of the present invention; Figure 2 is a schematic diagram of the calculation process of the interpretable deep learning diagnosis module of the present invention; Figure 3 is a schematic diagram of the calculation process of the physical constraint fusion module of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0017] The following further describes in detail the embodiments of the present invention in conjunction with 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.

[0018] The present invention provides an electromechanical system fault diagnosis system based on deep learning, including: A data acquisition and preprocessing module, which is used to collect the historical operation data and real-time operation data of the electromechanical system, and perform dynamic window length setting, cleaning, noise reduction, and standardization processing on them; An interpretable deep learning diagnosis module, which is connected to the data acquisition and preprocessing module, performs feature screening and decoupling learning with the help of causal gating and a dual-branch network, extracts fault-related features, and finally generates a diagnosis result containing a causal path and anomaly prompts, thereby realizing fault diagnosis and outputting a diagnosis basis at the same time; A physical constraint fusion module, which is connected to the data of the interpretable deep learning diagnosis module, obtains the physical principle data of the electromechanical system and the parameter data of the electromechanical system under normal working conditions, and physically constrains the interpretable deep learning diagnosis module through the physical principles of the electromechanical system to ensure that the diagnosis results comply with the laws of physics; A decision interaction module is connected to the data of the explainable deep learning diagnosis module, and is used to receive the results and explanation information output by the explainable deep learning diagnosis module and present them to the operation and maintenance personnel, provide an interactive interface, support manual re-inspection and feedback, and optimize the diagnosis system. Embodiment 1:

[0019] like Figures 1-3 As shown, in this embodiment, in a hub-level high-speed railway station, various key electromechanical systems such as ventilation, refrigeration, and power transmission are in a high-load operation state for a long time. During the peak passenger flow period during holidays, when multiple refrigeration units are linked to adjust the indoor temperature and humidity, the system monitoring interface suddenly shows that the equipment operation parameters in a certain area have abnormal fluctuations. At the same time, the equipment operation noise and vibration monitoring module synchronously captures non-periodic abnormal noises. The operation and maintenance personnel initially judged it as an equipment failure, but traditional monitoring methods cannot quickly distinguish whether it is a real failure caused by sensor signal interference, control logic abnormality, or mechanical component loss. Therefore, the present invention is introduced to avoid large-scale equipment shutdowns that affect passenger travel.

[0020] The data acquisition unit includes a vibration sensor, a temperature sensor, a current sensor, a voltage sensor and a humidity sensor, which are used to collect real-time operation data of the electromechanical system. At the same time, the data acquisition unit obtains historical electromechanical system operation data and fault data through the electromechanical system historical operation log; The data preprocessing unit uses sliding window technology to divide time series data, sets dynamic window length according to different equipment characteristics, uses adaptive filtering algorithm to eliminate noise interference in the data, and combines normalization method to unify data scale to ensure data quality and consistency.

[0021] The feature calculation unit generates a sparse causal graph by integrating time series and instantaneous causal analysis, and combines causal gating with a dual-branch network to achieve feature screening and decoupled learning, thus effectively extracting fault-related features. The feature calculation unit includes an offline calculation subunit and an online calculation subunit; The offline computing subunit receives the input multi-sensor time series data matrix and a fault label y, where, represents the set of real numbers, T×D represents the dimension of the matrix, T is the total length of the time series, and D is the sensor dimension; Adopt a time-domain causal discovery algorithm, fuse time-series causality and instantaneous causality analysis, and for each sensor data , use a 1D causal convolution kernel sliding window to calculate its causal strength score with the fault label marked based on binary faults : ; Among them, represents the sensor data for the fault label causal strength score, the smaller the score, the stronger the causality, is the time-series causal term, is the instantaneous causal term, represents the balance factor, which is used to control the weight ratio of the time-series causal term and the instantaneous causal term, and is verified by cross-validation; The time-series causal term The specific calculation is: ; Among them, the smaller the score, the stronger the causality, l represents the length of the causal convolution kernel, that is, the size of the time window, represents the weight parameter of the learnable convolution kernel k, which is used to weight the historical signal , represents the fault label at time step , 0 or 1, 0 means no fault, 1 means fault, i represents the i-th sensor, and j represents the j-th fault; The instantaneous causal term represents the sensor data and the fault label mutual information, which is used to quantify the instantaneous statistical dependence relationship between the two, and does not depend on the time order. The calculation formula is: ; Among them, x represents the discretized value of a certain sensor signal, and y represents a certain value state of the fault label, that is, 0 or 1; represents the sensor data is in x, and the fault is in the joint probability of y, represents the marginal probability that the sensor data is in x, represents the fault label is in the state marginal probability; After calculating the scores then, through significance testing, i.e., permutation testing, false associations are filtered to generate a sparse causal graph G, where the nodes are sensor i and fault j, and the edges are significant causal links ; Then the causal graph G is converted into a binary mask matrix , where C is the number of fault categories indicating that sensor i has a causal contribution to fault j; then at each specified training cycle, the causal graph is recalculated and M is updated to adapt to the change in data distribution; The online calculation subunit receives the input, sensor data and the mask matrix M, and through the causal gating layer, channel-level screening of the original features is performed: ; where represents the Hadamard product, forcing the weights of irrelevant channels to zero, and at the same time, only the sensor signals causally related to the current fault category are retained for subsequent calculations; Then decoupled feature learning is performed, using a dual-branch network structure: The causal branch processes and extracts the temporal causal features H through causal convolution; The non-causal branch processes the data and extracts the global features through ordinary CNN ; Output the fused features .

[0022] By generating a sparse causal graph through temporal and instantaneous causal analysis in the feature calculation unit, filtering non-causal associated features by the causal gating layer, and decoupling temporal causal features and global features by the dual-branch network, directional screening and efficient extraction of fault-related features in multi-sensor data such as vibration and temperature are achieved, achieving the feature purification effect of removing noise interference and irrelevant signals, improving the model's ability to capture early fault features, and significantly enhancing the accuracy and robustness of the fault diagnosis of the electromechanical system.

[0023] The interpretable diagnosis unit outputs the fault probability distribution based on the fused features and the dynamic causal mask, generates the contribution heat map, and identifies unknown anomalies through cluster analysis, and finally generates a diagnosis report including the causal path and anomaly prompt. The interpretable diagnosis unit includes an inference diagnosis subunit and a report generation subunit. The inference diagnosis subunit outputs the fault probability distribution based on the fused feature F and the dynamic causal mask M ; ; where is the weight matrix, with dimensions equal to the number of fault categories multiplied by the dimension of the fused feature F. b represents the bias term, and each category corresponds to a scalar bias, which is obtained through training based on the historical operation data and fault data of the electromechanical system; the gradient of the causal branch is traced for significance to generate the fault contribution heat map A, that is, the fault For the causal feature the sum of the time dimension gradients; ; Then, perform K-means clustering on the non-causal branch features If the current sample deviates from the normal cluster center by more than the threshold, mark this sample with a non-causal anomaly flag and trigger an unknown anomaly warning; The report generation subunit, based on the fault probability , the contribution heat map A, and the non-causal anomaly flag, combined with the dynamic causal graph G, extracts the top-K causal paths of the fault category j. If a non-causal anomaly warning is triggered, append the prompt: "An unvalidated feature pattern is detected. It is recommended to manually verify the equipment anomaly"; Through the feature screening and decoupling of the causal gating and double-branch network in the interpretable deep learning diagnosis module, the generation of the fault contribution heat map by the dynamic causal mask, and the extraction of the causal path, the explicit presentation of the causal relationship between the sensor signal and the fault type during the fault diagnosis process is realized, achieving a clear distinction between the sensor fault and the specific failure mode under the fault scenario. The maintenance personnel can quickly locate the root cause of the fault based on the causal path and the contribution heat map, avoiding delays in manual re-inspection and improving the emergency repair efficiency and operation safety of the electromechanical system in the high-speed railway station building. Example 2:

[0024] As Figures 1-3 shown, in this embodiment, in a certain period of time in the high-speed railway station building in Example 1, it was hit by continuous heavy rain for several days, and the outdoor rainwater backflowed into the equipment mezzanine, resulting in a sudden increase in the humidity of some electromechanical equipment compartments. The maintenance personnel found that the bearing temperature sensors of multiple pump units continuously reported anomalies, while the vibration sensor data did not change significantly, and the equipment operation power curve also remained stable. At this time, the model's preliminary diagnosis prompted "potential mechanical fault risk", but on-site maintenance experience shows that: under the condition that the active cooling system is not enabled and the equipment load does not change suddenly, a significant drop in the bearing temperature within a short time violates the basic laws of thermodynamics, and there may be signal distortion due to sensor moisture or conflicts in physical model predictions. It is necessary to verify whether the diagnostic results conform to the equipment dynamics principle to avoid the risks of over-maintenance or missed inspection caused by misjudgment.

[0025] The physical modeling unit symbolizes physical laws by integrating the multi - physical field principles of the electromechanical system, generates a differentiable computational graph, uses PINN to handle complex physical relationships, and realizes cross - field coupling through tensor operations. Finally, it outputs the DPCU Library that supports gradient backpropagation, laying a foundation for subsequent calculations; The physical modeling unit inputs the multi - physical field principles of the electromechanical system, such as the law of conservation of energy, and transforms these physical laws into symbolic expressions. Using the SymPy integration of the automatic differentiation tool PyTorch, it generates a differentiable computational graph; For complex physical relationships that are difficult to explicitly express, a physics - informed neural network PINN is used for surrogate modeling to ensure differentiability, construct a cross - field coupling relationship graph, define the interaction rules of physical quantities, and realize cross - field calculations through tensor operations. Finally, it outputs the differentiable physical calculation unit library DPCU Library and supports gradient backpropagation; The physical optimization unit injects dynamic physical constraints during the online training of the feature calculation unit; First, it inputs the original sensor data X and the DPCU Library for physical state estimation, and uses the DPCU to calculate the derived physical quantities from X , and at the same time, calculates the theoretical physical quantities based on the physical model , where z represents the number of types of physical quantities; Physical residual generation: Calculate the absolute residual between the actual and the theoretical: ; And statistically analyze the residual distribution through a sliding window , adaptively relax the constraints, and design the dynamic constraint weights: ; Among them, KL represents the KL divergence, which is used to measure the difference between two distributions, represents the scaling factor, which controls the difference sensitivity, is the Sigmoid function, is the residual distribution under normal conditions. The more the residual distribution deviates from the normal, the larger it is, and the stricter the physical constraints are; The actual application process is as follows: Sliding window statistics: Every specified time window, calculate the distribution of the residuals of each physical quantity within the current window , KL divergence calculation: Calculate the KL divergence between and the pre - stored , and convert the KL divergence into a weight through the Sigmoid function . The greater the difference, the closer it is to 1, and the stronger the physical constraints are; The physical verification unit is based on the absolute residual R when the model predicts the fault type y and t; Physical feasibility verification, according to the fault type y, retrieve the set of physical rules that it must violate ; If there is not satisfied, it is determined that the predicted result y is physically infeasible; Multi-hypothesis correction, start the physically guided candidate search: screen the top three highest fault probabilities from the fault probability distribution to screen the fault types that meet ; If no candidate is satisfied, trigger an unknown fault alarm, prompt manual intervention, physical parameter traceability, for physically infeasible predictions, trace back the sensor data that causes contradictions in reverse, and generate a list of suspicious sensors for maintenance reference; By converting the multi-physical field laws into a differentiable computational graph through the physical modeling unit in the physical constraint fusion module, dynamically injecting physical residual constraints through the physical optimization unit, and verifying the prediction results based on physical rules through the physical verification unit, the deep coupling of physical laws such as mechanical dynamics and thermodynamics with the deep learning model is realized, achieving the automatic identification and correction of diagnostic results that violate physical common sense, ensuring that the diagnostic results meet the physical feasibility of the actual working conditions, and avoiding maintenance misoperations caused by model misjudgment.

[0026] 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 mechatronic system fault diagnosis system based on deep learning, characterized in that: include: A data acquisition and preprocessing module, which is used to collect historical and real-time operation data of the electromechanical system, and set dynamic window length, cleaning, noise reduction and standardization processing; An interpretable deep learning diagnosis module, which is connected to the data acquisition and preprocessing module, performs feature screening and decoupling learning with the help of causal gating and a dual-branch network, extracts fault-related features, and finally generates a diagnosis result containing a causal path and anomaly prompts, thereby realizing fault diagnosis and outputting a diagnosis basis at the same time; A physical constraint fusion module, which is connected to the data of the interpretable deep learning diagnosis module, obtains the physical principle data of the electromechanical system and the parameter data of the electromechanical system under normal working conditions, and physically constrains the interpretable deep learning diagnosis module through the physical principles of the electromechanical system to ensure that the diagnosis results comply with the laws of physics; A decision interaction module is connected to the data of the explainable deep learning diagnosis module, and is used to receive the results and explanation information output by the explainable deep learning diagnosis module and present them to the operation and maintenance personnel, provide an interactive interface, support manual re-inspection and feedback, and optimize the diagnosis system.

2. A deep learning-based electromechanical system fault diagnosis system as claimed in claim 1, characterized in that: The data acquisition and preprocessing module includes a data acquisition unit and a data preprocessing unit; The data acquisition unit includes a vibration sensor, a temperature sensor, a current sensor, a voltage sensor and a humidity sensor, which are used to collect real-time operation data of the electromechanical system. At the same time, the data acquisition unit obtains historical electromechanical system operation data and fault data through the electromechanical system historical operation log; The data preprocessing unit uses sliding window technology to divide time series data, sets dynamic window length according to different equipment characteristics, uses adaptive filtering algorithm to eliminate noise interference in the data, and combines normalization method to unify data scale to ensure data quality and consistency.

3. The electromechanical system fault diagnosis system based on deep learning as claimed in claim 1, characterized in that: The explainable deep learning diagnosis module includes a feature calculation unit and an explainable diagnosis unit; The feature calculation unit generates a sparse causal graph by fusing time series and instantaneous causal analysis, combines causal gating with a dual-branch network to achieve feature screening and decoupling learning, and completes the effective extraction of fault-related features; The explainable diagnosis unit outputs the fault probability distribution based on the fusion features and dynamic causal masks, generates a contribution heat map, identifies unknown anomalies through cluster analysis, and finally generates a diagnosis report containing causal paths and anomaly prompts.

4. A deep learning-based electromechanical system fault diagnosis system as claimed in claim 3, characterized in that: The feature calculation unit includes an offline calculation subunit and an online calculation subunit; The offline calculation subunit receives the input multi-sensor time series data matrix and fault label y, where represents a set of real numbers, T×D represents the dimension of the matrix, T is the total length of the time series, and D is the sensor dimension; The time domain causal discovery algorithm is used to integrate the temporal causal and instantaneous causal analysis to analyze each sensor data. , using a 1D causal convolution kernel sliding window, calculate the fault label based on binary fault labeling The causal strength score : ； in, Representing sensor data Fault label The smaller the score, the stronger the causality. is the temporal causal term, is the instantaneous causal term, represents the balance factor, which is used to control the weight ratio of the temporal causal term and the instantaneous causal term, and is verified by cross selection; Temporal Causal Term The specific calculation is: ； Among them, the smaller the score, the stronger the causality, l represents the length of the causal convolution kernel, that is, the size of the time window, Represents the weight parameter of the convolution kernel k that can be learned, which is used to weight the historical signal , Indicates that at time step The fault label is 0 or 1, 0 means no fault, 1 means fault, i means the i-th sensor, and j means the j-th fault; Instantaneous causal term Representing sensor data With fault label The mutual information of is used to quantify the instantaneous statistical dependence between the two, which is independent of the time order. The calculation formula is: ； Among them, x represents the discretized value of a sensor signal, and y represents a certain value state of the fault label, that is, 0 or 1; Representing sensor data At x, and fault The joint probability of being in y, Representing sensor data The marginal probability of being at x, Indicates fault label In state The marginal probability of Calculate the score Then, the false associations are filtered out through the significance test, that is, the permutation test, to generate a sparse causal graph G, where the nodes are sensor i and fault j, and the edges are significant causal links. ; Then convert the causal graph G into a binary mask matrix , where C is the number of fault categories, Indicates that sensor i has a causal contribution to fault j; then, at every specified training cycle, the causal graph is recalculated and M is updated to adapt to changes in data distribution; The online computing subunit receives input, sensor data And the mask matrix M, through the causal gating layer, performs channel-level filtering on the original features: ; in, Represents the Hadamard product, forcing the weights of irrelevant channels to zero. At the same time, only the sensor signals causally related to the current fault category are retained to participate in subsequent calculations; Next, we conduct decoupled feature learning using a dual-branch network structure: Cause and effect branching, processing , extract the temporal causal feature H through causal convolution; The non-causal branch processes data X and extracts global features through ordinary CNN ; Output fusion features .

5. A deep learning-based electromechanical system fault diagnosis system as claimed in claim 4, characterized in that: The explainable diagnosis unit includes a reasoning diagnosis subunit and a report generation subunit; The reasoning diagnosis subunit outputs the fault probability distribution based on the fusion feature F and the dynamic causal mask M ; ; in, is a weight matrix, whose dimension is the number of fault categories multiplied by the dimension of the fusion feature F, b represents the bias term, each category corresponds to a scalar bias, which is obtained by training based on the operation data and fault data of the historical electromechanical system; the gradient of the causal branch is traced significantly to generate the fault contribution heat map A, that is, the fault Causal Features The sum of the time dimension gradients; ； Then, for the non-causal branch features Perform K-means clustering. If the current sample deviates from the normal cluster center by more than a threshold, a non-causal anomaly flag is marked on the sample and an unknown anomaly warning is triggered. The report generation subunit is based on the failure probability , the contribution heat map A and the non-causal anomaly signs are combined with the dynamic causal graph G to extract the Top-K causal paths of fault category j. If a non-causal anomaly warning is triggered, an additional prompt is added: "Unverified feature patterns are detected, and manual verification of equipment anomalies is recommended." 6. A deep learning-based electromechanical system fault diagnosis system as claimed in claim 5, characterized in that: The physical constraint fusion module includes a physical modeling unit, a physical optimization unit and a physical verification unit; The physical modeling unit symbolizes physical laws by integrating the multi-physics field principles of electromechanical systems, generates differentiable computational graphs, uses PINN to process complex physical relationships, and implements cross-field coupling with tensor operations. It finally outputs a DPCU Library that supports gradient backpropagation, laying the foundation for subsequent calculations. The physical optimization unit injects dynamic physical constraints in real time during model training, estimates the physical state based on the original sensor data and DPCULibrary, calculates the absolute residuals of actual and theoretical physical quantities, and dynamically adjusts the constraint weights in combination with the sliding window and KL divergence to ensure that the model training complies with physical laws; The physical verification unit verifies the physical feasibility of the prediction results based on the fault type and absolute residual predicted by the model, and determines the validity of the results by retrieving the violated physical rule set; for infeasible results, it screens high-probability candidate fault correction predictions, and if the correction is not possible, it triggers an alarm and traces back suspicious sensor data to provide a basis for system maintenance.

7. A deep learning-based electromechanical system fault diagnosis system as claimed in claim 6, characterized in that: The physical modeling unit inputs the multi-physics principle of the electromechanical system and converts the physical laws into symbolic expressions, and uses the SymPy integration of the automatic differentiation tool PyTorch to generate a differentiable computational graph; For complex physical relationships that are difficult to express explicitly, the physical information neural network PINN is used for proxy modeling to ensure differentiability, build a field coupling relationship diagram, define the interaction rules of physical quantities, realize cross-field calculations through tensor operations, and finally output the differentiable physical computing unit library DPCU Library, which supports gradient backpropagation.

8. A deep learning-based electromechanical system fault diagnosis system as claimed in claim 7, characterized in that: The physical optimization unit performs dynamic physical constraint injection during online training of the feature calculation unit; First, input the raw sensor data X and DPCU Library to estimate the physical state, and use DPCU to calculate the derivative physical quantity from X At the same time, theoretical physical quantities are calculated based on physical models , where z represents the number of types of physical quantities; Physical residual generation: Calculate the absolute difference between actual and theoretical residuals: ； And the residual distribution is counted through the sliding window , adaptive constraint relaxation, design dynamic constraint weights: ； Among them, KL represents KL divergence, which is used to measure the difference between two distributions. represents the scaling factor, controlling the difference sensitivity, is the Sigmoid function, is the residual distribution under normal working conditions. The more the residual distribution deviates from the normal condition, The larger it is, the tighter the physical constraints are; The actual application process is as follows: Sliding window statistics, calculate the distribution of the residuals of each physical quantity in the current window every specified time window , KL divergence calculation, With pre-stored Calculate the KL divergence and convert the KL divergence into weights through the Sigmoid function , the greater the difference, The closer it is to 1, the stronger the physical constraint.

9. A deep learning-based electromechanical system fault diagnosis system as claimed in claim 8, characterized in that: The physical verification unit predicts the absolute residual R of the fault types y and t based on the model; Physical feasibility check: according to the fault type y, retrieve the set of physical rules that must be violated ; If exists If it is not satisfied, the prediction result y is judged to be physically infeasible; Multi-hypothesis correction, enabling physics-guided candidate set search: from failure probability distribution Medium screening failure probability top three high screening compliance Type of fault; If no candidate meets the requirement, an unknown fault alarm is triggered, prompting manual intervention, physical parameter traceability, reverse tracing of contradictory sensor data for physically unfeasible predictions, and generation of a list of suspicious sensors for maintenance reference.

10. The electromechanical system fault diagnosis system based on deep learning as claimed in claim 5, characterized in that: The decision interaction module displays the contribution heat map, causal path map and original signal comparison, as well as the additional prompts generated in the report generation subunit by providing a web interface.

Images