An AMT motor fault diagnosis method and system based on machine learning

Abstract

This invention relates to the interdisciplinary field of machine learning and vehicle equipment health management, and discloses a machine learning-based method and system for AMT motor fault diagnosis. The method acquires multimodal sensor signals and historical state features of the motor. The sensor signals are input into a contrastive learning network based on the motor's dynamic equations for feature decoupling, outputting a candidate feature set. Temperature parameters are explored using a reinforcement learning agent based on historical feature deviation mapping. Current operating parameters are input into the agent, and a routing mask vector is generated by combining the temperature parameters. After physical boundary correction, an activated feature set is obtained. This mask is used to activate the corresponding diagnostic sub-model, and the predicted probability distributions of each activated sub-model are fused to output the final diagnostic result. This invention achieves precise decoupling of multimodal features, adaptively adjusts decision routes in the event of hardware degradation, effectively reduces underlying computing power overhead, and significantly improves diagnostic accuracy under complex operating conditions.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of machine learning and vehicle equipment health management, specifically to a machine learning-based method and system for diagnosing AMT motor faults. Background Technology

[0002] As a core component of modern automotive transmission systems, the operating status of the internal drive motor of an automatic transmission (AMT) directly determines the smoothness of gear shifting and driving safety. AMT motors are typically exposed to harsh conditions such as frequent load shocks, severe wide-frequency mechanical vibrations, and complex electromagnetic interference, leading to frequent electromechanical coupling faults of various types, including stator winding insulation degradation, rotor eccentricity, and bearing wear. Therefore, developing accurate and robust motor fault diagnosis systems has always been a key research focus in vehicle engineering and control.

[0003] Currently, mainstream motor fault diagnosis solutions in the industry mostly rely on physical quantity sensing data combined with traditional spectrum analysis or basic deep learning networks for pattern recognition. However, in the complex service environment of actual vehicles, the fault mechanism of motors exhibits highly nonlinear and cross-modal characteristics. When processing multi-source heterogeneous sensing signals, existing methods often only perform simple dimensional concatenation in the data preprocessing stage, failing to achieve deep collaboration and alignment of cross-modal features in the hidden semantic space. At the same time, these purely data-driven models generally ignore the constraints of the inherent physical dynamics of motors, causing the features extracted by the network to be easily interfered with by high-frequency physical distortion noise, deviating from objective reality. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a machine learning-based AMT motor fault diagnosis method and system, which solves the technical problems of existing motor fault diagnosis models being prone to physical distortion due to multimodal feature decoupling under complex working conditions, difficulty in adapting to hardware time-varying degradation, and excessive consumption of on-board edge computing power due to global static inference.

[0005] To achieve the above objectives, the present invention provides a machine learning-based AMT motor fault diagnosis method, comprising the following steps: Acquire multimodal sensing signals and historical operating status characteristics of the AMT motor; The multimodal sensing signal input is combined with a contrastive learning network constrained by the motor dynamics equations to decouple features and output a candidate feature set; Based on the historical operating state characteristics, the characteristic distribution deviation of the motor is calculated, and the characteristic distribution deviation is mapped to the policy exploration temperature parameter of the reinforcement learning agent. The current operating parameters of the motor are input into the reinforcement learning agent, and the temperature parameters are explored in combination with the strategy to output a routing mask vector. The routing mask vector is used to filter the candidate feature set to obtain the activation feature set. Activate the corresponding diagnostic sub-model according to the routing mask vector, input the activation feature set into the activated diagnostic sub-model to obtain the predicted probability distribution of each sub-model, weight and fuse the predicted probability distribution to obtain the system probability distribution, and output the diagnostic result; Calculate the divergence degree between the predicted probability distributions of each activated diagnostic sub-model, calculate the closed-loop reward value based on the divergence degree, and update the reinforcement learning agent according to the closed-loop reward value.

[0006] The innovative principle of the contrastive learning network constrained by the motor dynamics equations in the method provided by this invention lies in mapping the multimodal sensing signals into hidden layer electrical features and hidden layer mechanical features, respectively, through an electrical signal encoder and a mechanical signal encoder containing feedforward neural network layers. During pre-training, the physical encoder incorporates motor dynamics equations representing the mapping relationship between the motor stator voltage, three-phase current, rotor speed, and electromagnetic torque. Sample operating parameters are input into the physical encoder, which outputs theoretical state features. By calculating the contrastive loss value, the hidden layer electrical features and hidden layer mechanical features are forced to not only align with each other in the semantic space but also converge simultaneously to the theoretical state features that conform to physical mechanisms. This mechanism effectively eliminates non-physical noise in the sensing signals and decouples the candidate feature set that conforms to physical intuition.

[0007] In the method provided by this invention, the construction principle of the reinforcement learning agent is as follows: the current operating parameters of the motor are used to construct a state space, and the generation of the initial routing vector is used to construct an action space. To avoid getting trapped in local optima, the signal of the motor under preset benchmark operating parameters is extracted to obtain a feature vector. The Mahalanobis distance between the feature vector and the preset factory benchmark feature distribution is calculated as the feature distribution deviation. A nonlinear hyperbolic tangent function is used to scale and map the deviation to the benchmark, and the specific mathematical expression is as follows: ; in The temperature parameter represents the policy exploration parameter, which serves as a control variable for adjusting the smoothness of the probability distribution in the action selection policy of the reinforcement learning agent. The deviation of the characteristic distribution is represented by the combined Mahalanobis distance of the covariance matrix of the characteristic vector of the current cycle and the preset factory benchmark characteristic distribution and the Euclidean distance. This represents the preset baseline constant, which is the reference exploration temperature when the motor is in a healthy, non-aging state. and This represents the preset scaling factor, configured during system initialization, used to control the mapping magnitude and gradient of the hyperbolic tangent function.

[0008] Through this mechanism, when the motor undergoes physical degradation over its service life... Increase, driving the strategy to explore temperature parameters This increases the scope of the reinforcement learning agent's exploration of different feature channel combinations in the action space during the complex degradation stage.

[0009] In the method provided by this invention, the innovative principle of the closed-loop update of the reinforcement learning agent lies in the following: After completing diagnostic inference, the relative entropy algorithm is used to measure the divergence value of the individual fault probability distribution output by each activated diagnostic sub-model with a multilayer perceptron structure relative to the system probability distribution. The mean of the divergence is taken as the discrete entropy, which is then used as the divergence degree. The system constructs a multi-dimensional closed-loop reward model and updates the network parameters based on a comprehensive feedback of accuracy, computational cost, and diagnostic divergence degree. The formula for calculating the closed-loop reward value is as follows: ; in This represents the closed-loop reward value, which serves as the objective function value for gradient backpropagation optimization performed by the reinforcement learning agent's policy network. This indicates a preset diagnostic accuracy bonus item, which is assigned a fixed positive scalar value by the system when the fault category corresponding to the maximum value of the system probability distribution is consistent with the actual state or the confidence level meets the standard. The number of feature channels is determined by counting the number of elements with a value of 1 in the routing mask vector in real time. This represents the preset cost coefficient, and... The product constitutes the penalty term for feature channel calculation, which is used to constrain the computational cost of the model; The discrete entropy is assigned by the degree of divergence obtained from real-time calculation. This represents a preset divergence penalty coefficient, when the discrete entropy When the value exceeds the preset divergence threshold, this penalty is triggered to increase the negative constraint.

[0010] This mechanism enables the diagnostic sub-models to exhibit high divergence when faced with complex and unknown variable operating conditions. During a sudden increase, the strong negative reward will disrupt the existing policy steady state, forcing the agent to change the activation combination of the routing mask vector under the same current operating parameters of the motor, thereby realizing the algorithm's adaptive diagnostic evolution of the time-varying characteristics of the motor.

[0011] Another aspect of the present invention provides an AMT motor fault diagnosis system based on machine learning, comprising: a processor, a memory, and a communication module; The memory stores a computer program, which, when executed by the processor, implements the steps of the machine learning-based AMT motor fault diagnosis method described in the first aspect of the present invention.

[0012] This invention provides a machine learning-based method and system for fault diagnosis of AMT motors. It has the following beneficial effects: 1. This invention breaks through the limitation that pure data-driven models are prone to fitting high-frequency physical distortion noise. It forces electrical signals and mechanical signals to cooperate with each other in the hidden semantic space and converge towards the anchor point that conforms to the objective operating law. This achieves precise decoupling of complex electromechanical coupling characteristics at the source and enhances the system's bottom-level anti-disturbance capability under strong electromagnetic interference and severe broadband vibration environment.

[0013] 2. This invention quantifies the characteristic distribution deviation of historical operating states and nonlinearly maps it to the policy exploration temperature parameter of the reinforcement learning agent. When the equipment gradually ages and causes the current operating state to deviate from the factory preset distribution, this mechanism can adaptively amplify the exploration temperature value, actively smooth out the existing probability distribution peaks and expand the agent's optimization field of vision, effectively avoiding the model from getting trapped in outdated local optima.

[0014] 3. This invention can perform precise channel-by-channel screening of candidate feature sets based on transient operating conditions, centrally schedule limited on-board underlying computing resources to high-confidence inference links, and completely intercept abnormal data streams that violate physical boundaries by using forced zeroing operations, thus safeguarding the physical bottom line of decision security while taking into account the economic efficiency of computing power.

[0015] 4. This invention relies on a concurrent multi-expert diagnostic sub-model array to perform independent forward mapping on the selected activation feature set, and then performs weighted aggregation of probability distributions based on the historical confidence weights of each model. This dynamic addressing and integrated architecture design effectively smooths out the variance fluctuations that are easily generated by single neural networks when dealing with nonlinear fault characteristics. Attached Figure Description

[0016] Figure 1 This is a flowchart of the present invention; Figure 2 This is a schematic diagram of the cross-modal contrastive learning network structure and feature decoupling of the present invention; Figure 3 This is a block diagram of the reinforcement learning dynamic route generation and physical boundary constraint mechanism of the present invention; Figure 4 This is a diagram of the multi-expert diagnostic sub-model inference and weighted fusion architecture of the present invention; Figure 5 This is a logic diagram of the reinforcement learning closed-loop reward calculation and update mechanism of the present invention; Figure 6This is a hardware structure block diagram of the machine learning-based AMT motor fault diagnosis system of the present invention. Detailed Implementation

[0017] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see the appendix Figure 1 -Appendix Figure 6 This invention provides a machine learning-based AMT motor fault diagnosis method and system, which aims to achieve high-precision fault identification of mechanical automatic transmission AMT motors under complex working conditions and time-varying degradation processes through cross-modal feature decoupling and reinforcement learning dynamic routing mechanism.

[0019] The machine learning-based AMT motor fault diagnosis method includes the following steps: Acquire multimodal sensing signals and historical operating status characteristics of the AMT motor; The multimodal sensing signal input is combined with a contrastive learning network constrained by the motor dynamics equations to decouple features and output a candidate feature set; Based on the historical operating state characteristics, the characteristic distribution deviation of the motor is calculated, and the characteristic distribution deviation is mapped to the policy exploration temperature parameter of the reinforcement learning agent. The current operating parameters of the motor are input into the reinforcement learning agent, and the temperature parameters are explored in combination with the strategy to output a routing mask vector. The routing mask vector is used to filter the candidate feature set to obtain the activation feature set. Activate the corresponding diagnostic sub-model according to the routing mask vector, input the activation feature set into the activated diagnostic sub-model to obtain the predicted probability distribution of each sub-model, weight and fuse the predicted probability distribution to obtain the system probability distribution, and output the diagnostic result; Calculate the divergence degree between the predicted probability distributions of each activated diagnostic sub-model, calculate the closed-loop reward value based on the divergence degree, and update the reinforcement learning agent according to the closed-loop reward value.

[0020] In the overall data flow logic of the system, the algorithm front end first performs noise reduction and feature mapping on the heterogeneous physical signals collected at the bottom layer. By introducing objective motor dynamics equation constraints, the neural network is forced to filter out redundant information that violates the physical mechanism, thereby decoupling a high-purity candidate feature set.

[0021] In the mid-range routing decision-making stage, the agent does not use static rule matching, but dynamically adjusts the exploration boundary of its action space based on the long-term historical service status of the motor. The algorithm transforms the macroscopic aging degradation into microscopic temperature control parameters, guiding the agent to generate corresponding routing mask vectors, and then accurately extracts the activation feature set most sensitive to the current operating conditions from the candidate feature set.

[0022] In the diagnostic execution and feedback phase, the system, based on the masked instructions generated previously, activates a specific combination of diagnostic sub-models from a pre-set model library. Each sub-model, after receiving the same activation feature set, performs independent inference calculations. The final system probability distribution is generated through a weighted fusion mechanism, and its calculation process is characterized by the following basic formula: ; in It represents the system probability distribution, and its form is a numerical vector containing the occurrence probability of each fault category in the target fault category space; The total number of activated diagnostic sub-models is determined by counting the number of elements with a value of 1 in the routing mask vector; Indicates the system is the first The computational weights assigned to each activated diagnostic sub-model are used to adjust the proportion of each sub-model in the global decision, and the sum of the weights of all activated sub-models is 1. This represents the activation feature set obtained after filtering by the routing mask vector. Indicates the first Each activated diagnostic sub-model receives an activation feature set. After input, the individual fault probability distribution is calculated and output by its own network.

[0023] After obtaining the system probability distribution, the algorithm extracts the fault category corresponding to the element with the largest value in the distribution as the diagnostic result output for the current period. Subsequently, the system quantifies the divergence differences between the outputs of each independent sub-model and transforms them into reward and punishment feedback signals, which are then fed back to the reinforcement learning agent.

[0024] The system first acquires electrical and mechanical signals in real time for the current cycle through a multi-channel sensor network deployed on the motor body and drive controller. The electrical signals include the three-phase current signal driving the stator and the DC bus voltage signal, while the mechanical signals include the vibration acceleration signal from the high-frequency excitation of the motor housing and the position angle signal reflecting the rotor's motion state. The system's underlying hardware performs timestamp alignment and resampling operations on the asynchronous data streams generated by the heterogeneous sensors based on a high-precision clock, eliminating phase shifts caused by inherent differences in the sensor's sampling rate. The synchronized electrical and mechanical signals are then multi-dimensionally spliced ​​and encapsulated to ultimately form the multimodal sensing signal.

[0025] In this invention, to extract historical operating state characteristics, the system continuously monitors the interactive commands on the control bus. When it is determined that the actual control commands and feedback state of the motor remain stably within a preset benchmark operating parameter domain, a long-cycle state measurement is triggered. The preset benchmark operating parameter domain corresponds to the constant speed and constant output torque range of the motor at a specified gear. After entering this steady-state range, the system extracts raw sensor data within a fixed time window, calculates its multi-order time-domain statistics and frequency-domain band energy distribution, and vectorizes these multi-dimensional indicators to obtain the feature vector of the current cycle.

[0026] After obtaining the feature vector for the current period, the system needs to quantify the current physical health baseline of the motor. The system retrieves the factory-referenced feature distribution parameters stored in non-volatile memory. Considering the strong nonlinear coupling and covariance effects between different physical dimensions such as motor vibration and current, conventional Euclidean distance cannot accurately reflect the true deviation in the state space. Therefore, the system uses the Mahalanobis distance algorithm to calculate the deviation of the current feature vector from the factory-referenced distribution, thereby eliminating the interference of correlations between multiple variables and achieving an objective assessment of the motor's long-term performance drift.

[0027] The specific calculation process of the Mahalanobis distance is represented by the following mathematical formula: ; in The Mahalanobis distance, representing the feature vector of the current period relative to the factory baseline feature distribution, is a non-negative scalar value. This indicates that the system extracts and quantizes high-dimensional feature vectors in real time under the preset benchmark operating parameters. The mean vector representing the distribution of factory baseline characteristics is obtained by statistically averaging the data collected multiple times during the factory calibration test of the target motor under the same preset baseline operating parameters. This matrix represents the inverse of the covariance matrix of the factory baseline characteristic distribution, used to characterize the inherent statistical correlation structure among the various dimensions of health characteristics during the calibration testing phase. (Superscript) This represents the matrix transpose operation.

[0028] After obtaining the Mahalanobis distance through the matrix operations described above, the system combines the scalar value with its time series rate of change. To further mitigate single-time deviation jumps caused by external ambient temperature or occasional measurement noise, the system employs a sliding time window mean filtering algorithm to smooth the generated sequence results, ultimately constructing the processed results as the historical operating state features.

[0029] During the online inference phase of the system, multimodal sensing signals are simultaneously fed into an electrical signal encoder and a mechanical signal encoder with a multi-layer feedforward neural network structure. These two parallel network branches extract deep data representations of the signals in the spatiotemporal domain, mapping them to hidden electrical and mechanical features in a unified dimension. Subsequently, the system concatenates and fuses the hidden electrical and mechanical features along the channel dimension to directly construct the high-dimensional candidate feature set.

[0030] In this invention, to ensure that the candidate feature set output online is not contaminated by sudden interference in complex industrial environments, the network structure parameters of both the electrical signal encoder and the mechanical signal encoder are obtained through an offline pre-training process that includes physical mechanism constraints. The system constructs an independent physical encoder within the training framework, which internally embeds the motor dynamics equations characterizing the relationship between the electromagnetic energy conversion and mechanical power transmission of the target motor.

[0031] During the offline pre-training phase, the system inputs sample operating parameters extracted from the historical operating database into the physical encoder. These sample operating parameters include stator voltage command, current speed, and external load torque. The physical encoder performs rigorous mechanism calculations based on preset dynamic differential equations, and its output reflects the theoretical state characteristics of the motor under fault-free ideal conditions.

[0032] To achieve deep decoupling between data features and physical properties, the system applies polygonal geometric alignment constraints to the three extracted features within the hidden layer semantic space. Specifically, the system constructs and calculates a contrast loss value with the goal of minimizing a first difference between the hidden layer electrical features and the hidden layer mechanical features, and a second difference between each of them and the theoretical state features. The system then continuously updates the parameters of the electrical signal encoder and the mechanical signal encoder using a backpropagation algorithm.

[0033] The calculation logic for the contrastive loss value is represented by the following mathematical formula: ; in This represents the calculated contrastive loss value, which serves as the scalar value of the cost function for optimizing the neural network weights during the pre-training phase. and These represent the hidden layer electrical feature and hidden layer mechanical feature vectors generated by forward calculation of input sample signals within the same time window through an electrical signal encoder and a mechanical signal encoder, respectively. This represents the theoretical state feature vector generated by directly mapping the synchronous sample running parameters through the physical encoder. The L2 norm operator is used to measure the squared Euclidean distance between two high-dimensional vectors. This represents the first difference weighting coefficient, which the system uses to control the penalty for mutual alignment between electrical and mechanical modal features. This represents the second difference weighting coefficient, which the system uses to control the penalty for the convergence of hidden features generated by pure data-driven approaches to features based on objective physical theories.

[0034] Through the aforementioned calculations and gradient iterative optimization, the cross-modal contrastive learning network forces the hidden features extracted based on pure data-driven methods to converge towards anchor points that conform to objective physical laws within the multidimensional manifold space. This design, which implicitly embeds the first-principles equations of the motor into the feature extraction space of the neural network, eliminates non-realistic physical noise introduced by factors such as vibration transmission path distortion or electromagnetic environment coupling. This fundamentally ensures that the pre-trained network can decouple and output a candidate feature set with high purity and physical interpretability during actual online diagnosis.

[0035] In this invention, when the AMT motor is in a healthy state at the beginning of its life cycle, the characteristic distribution of its multimodal signals is highly concentrated. The system needs the network to quickly converge to a fixed feature channel selection strategy to reduce inference latency and computational overhead. However, when the motor undergoes irreversible degradation due to long-term high-load operation, such as a decrease in winding insulation or an increase in mechanical clearance, the boundary of the original feature space will drift nonlinearly, and the original fixed routing strategy will no longer be applicable. To address this, the system introduces a dynamic temperature regulation mechanism based on cross-domain mapping of hardware physical biases into the reinforcement learning architecture.

[0036] The system directly extracts the Mahalanobis distance, which characterizes the long-term physical performance drift, obtained from the aforementioned calculations, and uses it as the feature distribution deviation. Subsequently, the system inputs the feature distribution deviation into a preset nonlinear hyperbolic tangent function for dimensionless scaling and mapping truncation, and adds the scaled calculation result to a preset fundamental constant to generate a policy exploration temperature parameter that is directly used to intervene in the agent's decision distribution.

[0037] The specific mathematical logic of this mapping mechanism is represented by the following formula: ; in The output policy exploration temperature parameter is used as a hyperparameter input to regulate the smoothness of the probability distribution of the action output layer of the reinforcement learning agent. The deviation of the feature distribution is represented by the Mahalanobis distance value relative to the reference state, which is passed in real time by the historical operating state feature construction module. This represents a preset basic constant, which characterizes the baseline temperature threshold required for the system to maintain the minimum optimization activity of the algorithm when the motor is under a hardware aging baseline condition. This represents the preset mapping amplitude control factor, which is configured during system initialization. It is used to limit the absolute upper limit of the exploration temperature rise caused by the increase of physical deviation, so as to prevent the model from falling into a completely random chaotic exploration state. This represents the preset gradient adjustment factor, used to control the curve response sensitivity of the hyperbolic tangent function to small degradation features in the early stage.

[0038] Through the aforementioned nonlinear function constraints, the strategy explores that the temperature parameter value strictly increases nonlinearly and monotonically with the increase of the feature distribution deviation. When a long-period deviation in the underlying hardware characteristics of the motor is detected, the increased temperature parameter will actively smooth out the peak value of the agent's output action distribution, increasing the probability of non-dominant actions being selected.

[0039] In this invention, the system constructs the current operating parameters of the motor, including bus torque commands, real-time stator three-phase current RMS values, and current rotor speed, as the observation state input for the agent. After receiving the state input, the agent's policy network calculates the action preference values ​​of the feature channel combinations of each dimension through a fully connected layer, and inputs the action preference values ​​along with the policy exploration temperature parameters generated in the previous step into the action selection policy function.

[0040] The action selection strategy function is an exponential probability distribution algorithm with a temperature scaling factor. Under the control of this function, the strategy network samples according to the probability distribution affected by the temperature parameter and outputs an initial routing vector consisting of 0s and 1s. The length of this vector is consistent with the number of candidate feature sets output by the aforementioned cross-modal contrastive learning network in the channel dimension, and each element in the vector independently determines whether the corresponding channel feature data is enabled or disabled.

[0041] After generating the initial routing vector, the system calls the previously constructed motor dynamics equation in parallel, inputs the current control parameters into the equation, and calculates the theoretical numerical range that each dimension of characteristic variables should follow under the healthy state.

[0042] Subsequently, the system performs a channel-by-channel boundary condition matching operation. The judgment logic is as follows: the system compares the feature values ​​of each channel in the candidate feature set with their corresponding theoretical value ranges in real time; if it detects that the data of a certain channel feature significantly exceeds the theoretical upper or lower limit determined by the motor dynamics equation, and the deviation does not conform to any known typical fault drift law, then the system determines that the channel feature has been severely contaminated by sporadic sensing noise or has physical distortion.

[0043] After obtaining the routing mask vector that has passed physical boundary constraint verification, the system performs the final feature activation extraction operation. Through tensor inner product operations, the routing mask vector and the candidate feature set are multiplied element-wise along the channel dimension. This operation ensures that the channel features corresponding to positions with a value of 1 in the routing mask vector are preserved without loss, while the channel features corresponding to positions with a value of 0 are completely cleared.

[0044] In this invention, the system pre-configures an expert diagnostic array composed of multiple independent feedforward neural networks at the inference and decision-making backend. Each diagnostic sub-model is configured with a uniform-dimensional input tensor interface and a standardized classification output layer. Upon receiving the mask instruction from the pre-reinforcement learning optimization stage, the system performs a strict hard routing addressing and matching operation. Specifically, the system iterates through and analyzes the routing mask vector, using the channel index of the element with a value of 1 as the physical trigger condition, activating the diagnostic sub-model corresponding to that index, while keeping other model networks with values ​​of 0 in the mask in a dormant state. This blocks invalid data flow and reduces the computational power consumption of the underlying hardware.

[0045] After the model is activated, each activated diagnostic sub-model reads the shared activation feature set in parallel. Each sub-model performs forward mapping and nonlinear classification calculations on the feature set within its independently defined weight space, thereby generating individual fault probability distributions covering the target fault category at its respective output layer. The individual outputs of all activated models together constitute the system's predicted probability distribution set.

[0046] Faced with the concurrent local inferences output by multiple diagnostic sub-models, the system employs a weighted summation algorithm to unify decisions and eliminate variance fluctuations in individual inferences. The system assigns basic computational weights to each activated diagnostic sub-model based on a preset storage configuration. These weights are derived from the normalized basic accuracy metrics obtained during the initial offline calibration phase based on validation set testing, representing the historical confidence of each expert network. Subsequently, the system performs a weighted summation of the individual failure probability distributions of all activated models, ultimately aggregating them to obtain the system probability distribution representing the globally optimal decision.

[0047] The specific mathematical logic of this weighted fusion operation is represented by the following formula: ; in The system probability distribution generated by the aggregation calculation is a numerical vector containing all predefined fault categories of the system and their corresponding probabilities of occurrence. This indicates the total number of diagnostic sub-models activated in the current period, which the system obtains by counting the number of elements with a value of 1 in the routing mask vector in real time. Indicates the system is the first The scalar computation weights assigned to each activated diagnostic sub-model are such that the sum of these weights for all activated models is strictly equal to 1. This represents the set of activation features that serves as a common input source for multiple models; Indicates the first Each activated diagnostic sub-model receives an activation feature set. After input, the individual fault probability distribution vector is calculated and output through its own network layer.

[0048] After completing the weighted fusion operation described above, the system performs extreme value analysis on the system probability distribution vector. The system locates and extracts the element with the largest value in the probability distribution vector. This largest value represents the highest confidence level of the current diagnostic architecture after integrating local and global opinions. The system directly extracts the category label mapped to this largest value, uses the corresponding fault category as the final diagnostic result of the current work cycle, and converts it into a standard format message for output to the vehicle's upper-level communication bus.

[0049] In this invention, the system retrieves the individual fault probability distributions independently output by each activated diagnostic sub-model, as well as the final fused system probability distribution. The system employs a relative entropy algorithm to calculate the divergence value of each individual fault probability distribution relative to the system probability distribution. After completing the traversal calculation, the system calculates the discrete entropy by arithmetically averaging all the acquired divergence values. The system directly uses this discrete entropy as the divergence degree to objectively reflect the level of decision conflict generated by the multi-expert model array when dealing with specific input features within the current execution cycle.

[0050] Based on the quantification of disagreement levels, the system constructs a multi-dimensional closed-loop reward mechanism encompassing decision accuracy, computational cost, and internal stability. First, the system extracts diagnostic results and compares them with the verification labels of the current period, obtaining a preset diagnostic accuracy reward based on the correctness of the comparison result. Simultaneously, to prevent the agent from indiscriminately waking up all models during optimization, the system continuously counts the number of elements with a value of 1 in the routing mask vector, using this count as the number of feature channels. The system multiplies the number of feature channels by a preset cost coefficient to calculate a feature channel computation penalty, used to strictly constrain the agent's strategy from unnecessary consumption of underlying hardware computing power.

[0051] Regarding the divergence index, the system incorporates a threshold-based dynamic penalty triggering mechanism in the closed-loop feedback loop. The system continuously checks whether the acquired divergence degree exceeds a preset divergence threshold. When the divergence degree exceeds the threshold, it indicates a serious inference contradiction between sub-models under the current feature combination. The system then proactively increases the value of the discrete entropy penalty term in the closed-loop reward value based on the current discrete entropy value. Finally, the system performs a weighted sum of the aforementioned diagnostic accuracy reward term, feature channel calculation penalty term, and discrete entropy penalty term to calculate the final closed-loop reward value.

[0052] The calculation logic for this comprehensive closed-loop reward value is represented by the following formula: ; in This represents the closed-loop reward value obtained after weighted summation, which serves as the scalar of the objective function for optimizing the policy network execution of the reinforcement learning agent. This represents the positive incentive value assigned by the system to the diagnostic accuracy reward item based on the correctness of the diagnosis; This indicates the number of feature channels obtained by the system in real time. This represents a preset cost coefficient, used to convert the number of channels into a feature channel calculation penalty term in the dimension of computational overhead. This represents the discrete entropy value assigned by the degree of divergence; This represents a preset divergence penalty coefficient. When the divergence degree is greater than a preset divergence threshold, the system actively increases the value of this coefficient, thereby significantly increasing the negative proportion of the discrete entropy penalty term in the global reward system.

[0053] After obtaining the comprehensive closed-loop reward value, the system uses it as a feedback signal to send back to the algorithm's middleware. Based on this increased closed-loop reward value, the system performs gradient backpropagation calculations and updates the weight parameters of the policy network in the reinforcement learning agent according to the policy gradient theorem.

[0054] In this invention, the machine learning-based AMT motor fault diagnosis system includes a processor, a memory, and a communication module. The hardware components are interconnected via a high-speed data bus within the system, enabling bidirectional data communication.

[0055] The communication module includes a hardware interface for connecting the onboard sensor network and the vehicle's underlying electronic control unit, specifically encompassing a multi-channel controller area network transceiver and a high-frequency analog-to-digital converter. This module is responsible for real-time access to the voltage signal from the AMT motor stator, the three-phase current signal, and the vibration acceleration signal from the motor housing. It performs necessary electrical isolation, low-pass filtering, and timestamp synchronization operations on these multimodal sensor signals, and then pushes the pre-processed data stream to the system's buffer storage area.

[0056] The memory includes volatile random access memory and non-volatile flash memory. The non-volatile memory stores computer program code for performing motor fault diagnosis, as well as basic configuration data such as factory baseline characteristic distribution parameters, motor dynamic equation constants, preset fundamental constants, and divergence thresholds. Simultaneously, the memory is responsible for continuously storing and dynamically updating the network weight parameters of the reinforcement learning agent and the multi-expert diagnostic sub-model during system operation.

[0057] The processor is the core computing unit of the system. It establishes deep interaction with the memory and communication module, and by calling and executing the computer program code stored in the memory, it sequentially completes logical operations such as reading multimodal sensor signals and mapping historical features, decoupling features from the cross-modal contrastive learning network, nonlinear calculation and mask generation for exploring temperature parameters using reinforcement learning strategies, and dynamic activation and probability fusion of the multi-expert diagnostic sub-model. After completing the above inferences, the processor transmits the fused system probability distribution and the message frame containing the final diagnostic results back to the vehicle's main control unit in real time via the communication module. Simultaneously, the processor independently calculates the closed-loop reward value based on the divergence quantification result, performs gradient backpropagation calculations, and updates the corresponding algorithm network parameters in the memory.

Claims

1. A machine learning-based fault diagnosis method for AMT motors, characterized in that, include: Acquire multimodal sensing signals and historical operating status characteristics of the AMT motor; The multimodal sensing signal input is combined with a contrastive learning network constrained by the motor dynamics equations to decouple features and output a candidate feature set; Based on the historical operating state characteristics, the characteristic distribution deviation of the motor is calculated, and the characteristic distribution deviation is mapped to the policy exploration temperature parameter of the reinforcement learning agent. The current operating parameters of the motor are input into the reinforcement learning agent, and the temperature parameters are explored in combination with the strategy to output a routing mask vector. The routing mask vector is used to filter the candidate feature set to obtain the activation feature set. Activate the corresponding diagnostic sub-model according to the routing mask vector, input the activation feature set into the activated diagnostic sub-model to obtain the predicted probability distribution of each sub-model, weight and fuse the predicted probability distribution to obtain the system probability distribution, and output the diagnostic result; Calculate the divergence degree between the predicted probability distributions of each activated diagnostic sub-model, calculate the closed-loop reward value based on the divergence degree, and update the reinforcement learning agent according to the closed-loop reward value.

2. The machine learning-based fault diagnosis method for AMT motors according to claim 1, characterized in that, The acquisition of multimodal sensing signals and historical operating status characteristics of the AMT motor includes: The multimodal sensing signal is composed of electrical and mechanical signals collected during the current period. The signal of the motor under preset benchmark operating parameters is extracted to obtain a feature vector. The Mahalanobis distance between the feature vector and the preset factory benchmark feature distribution is calculated. The historical operating state features are constructed based on the Mahalanobis distance.

3. The machine learning-based fault diagnosis method for AMT motors according to claim 1, characterized in that, The step involves decoupling the multimodal sensing signal input into a contrastive learning network constrained by the motor dynamics equations to output a candidate feature set, including: The multimodal sensing signals are mapped into hidden layer electrical features and hidden layer mechanical features respectively by an electrical signal encoder and a mechanical signal encoder. The candidate feature set is obtained by splicing and fusing the hidden layer electrical features and the hidden layer mechanical features; The electrical signal encoder and the mechanical signal encoder are obtained based on pre-training. The pre-training process includes: inputting the sample running parameters into the physical encoder, outputting theoretical state features based on the motor dynamics equation, calculating the contrast loss value with the goal of minimizing the first difference between the hidden layer electrical features and the hidden layer mechanical features, and the second difference between the two and the theoretical state features respectively, and updating the encoder parameters.

4. The machine learning-based fault diagnosis method for AMT motors according to claim 1, characterized in that, The process of mapping the feature distribution bias to a policy exploration temperature parameter for the reinforcement learning agent includes: The characteristic distribution deviation is input into a nonlinear hyperbolic tangent function for scaling. The scaled result is added to a preset base constant to generate the strategy exploration temperature parameter, and the value of the strategy exploration temperature parameter increases as the feature distribution deviation increases.

5. The machine learning-based fault diagnosis method for AMT motors according to claim 3, characterized in that, The step of filtering the candidate feature set using the routing mask vector to obtain the active feature set includes: The reinforcement learning agent inputs the current operating parameters of the motor and the policy exploration temperature parameters into the action selection policy function, and outputs an initial routing vector. If the channel features in the candidate feature set exceed the theoretical numerical range determined by the motor dynamics equation, then the value of the corresponding channel in the initial routing vector is forcibly set to zero to obtain the routing mask vector. The activation feature set is extracted by multiplying the routing mask vector with the corresponding elements of the candidate feature set.

6. The machine learning-based fault diagnosis method for AMT motors according to claim 1, characterized in that, The output diagnostic results include: The diagnostic sub-model corresponding to the element with a value of 1 in the routing mask vector is taken as the activated diagnostic sub-model; Each activated diagnostic sub-model independently performs classification calculations on the activated feature set and outputs the corresponding individual fault probability distribution as the predicted probability distribution; Assign computational weights to each activated diagnostic sub-model, and sum the weighted probabilities of all individual faults to obtain the system probability distribution; The fault category corresponding to the maximum value in the system probability distribution is taken as the diagnostic result.

7. The machine learning-based fault diagnosis method for AMT motors according to claim 6, characterized in that, The calculation of the closed-loop reward value based on the divergence degree includes: The divergence value of the probability distribution of each individual failure relative to the probability distribution of the system is calculated using the relative entropy algorithm; The mean of all divergence values ​​is calculated to obtain the discrete entropy, and the discrete entropy is used as the degree of divergence. The closed-loop reward value is obtained by weighting and summing the preset diagnostic accuracy reward term, feature channel calculation penalty term, and discrete entropy penalty term.

8. The machine learning-based AMT motor fault diagnosis method according to claim 7, characterized in that, The step of updating the reinforcement learning agent based on the closed-loop reward value includes: Determine whether the degree of divergence is greater than a preset divergence threshold; When the degree of divergence is greater than the divergence threshold, the value of the discrete entropy penalty term in the closed-loop reward value is increased based on the discrete entropy; Gradient backpropagation calculation is performed based on the increased closed-loop reward value to update the network parameters of the reinforcement learning agent.

9. The machine learning-based fault diagnosis method for AMT motors according to claim 7, characterized in that, The calculation method for the feature channel calculation penalty term includes: The number of elements with a value of 1 in the routing mask vector is counted and used as the number of feature channels; The feature channel calculation penalty term is calculated by multiplying the number of feature channels by a preset cost coefficient.

10. A machine learning-based AMT motor fault diagnosis system, characterized in that, include: The system includes a processor, a memory, and a communication module. The memory stores a computer program, which, when executed by the processor, implements a machine learning-based AMT motor fault diagnosis method as described in any one of claims 1-9.

Images