An AI system for driving motor failure diagnosis and residual life grading early warning

Abstract

The present application relates to the technical field of intelligent motor operation and maintenance diagnosis, and particularly relates to a driving motor failure diagnosis and residual life grading early warning AI system, comprising: a multi-source data acquisition module, a feature fusion extraction module, a lightweight mixed AI judgment module, a grading early warning management and control module, and a vehicle-cloud collaborative iteration module; the multi-source data acquisition module is used for high-frequency synchronous acquisition of electrical parameters, vibration signals and rotating speed signals of the driving motor, and outputs multi-source original signals to the feature fusion extraction module. The present application performs high-frequency synchronous acquisition on multiple signals such as current, voltage, vibration and rotating speed, combines adaptive decomposition and feature mining algorithms, effectively strips aging-related feature components from background noise, so that the hidden degradation states such as winding micro-short circuit, early pitting of bearings and slight eccentricity of rotors can be identified in time, and a complete and reliable data basis is provided for subsequent diagnosis and early warning.

Description

Technical Field

[0001] This invention relates to the field of intelligent motor operation and maintenance diagnosis technology, specifically to an AI system for drive motor failure diagnosis and remaining life classification early warning. Background Technology

[0002] Drive motors are the core power components of new energy and industrial transmission equipment. Their operating status is directly related to equipment safety and maintenance costs. Motors that work under complex conditions for a long time are prone to hidden aging problems, which can easily lead to power decline or even sudden power interruption. Existing monitoring methods are mostly based on single data acquisition methods, which have limited dimensions and weak synchronization, making it difficult to capture early subtle aging characteristics and accurately assess the health status of motors.

[0003] Currently, the industry mostly uses a single AI model for motor fault diagnosis, which is only effective for identifying obvious faults. It is prone to misjudgment and omission of early hidden faults. Traditional high-precision models have complex structures and high computing power requirements, making them difficult to adapt to real-time deployment on vehicle terminals. At the same time, existing technologies can only qualitatively judge faults and cannot quantify failure probability and remaining lifespan, making it difficult to support refined health management.

[0004] Currently, motor maintenance is mostly a passive mode of manual inspection and post-event repair, lacking a hierarchical early warning and proactive power control mechanism. It is difficult to avoid the risk of power interruption in advance, and maintenance work relies heavily on manual experience. There is a lack of intelligent and standardized maintenance solutions, resulting in low overall maintenance efficiency and accuracy. Furthermore, traditional models are fixed and cannot be iterated, making them unable to adapt to the ever-changing actual operating conditions of motors.

[0005] Existing motor diagnostic and maintenance technologies suffer from problems such as insufficient monitoring dimensions, poor accuracy in identifying hidden faults and lifespan prediction, difficulty in vehicle-mounted deployment, insufficient intelligent early warning and maintenance, and inability to iteratively update models. These technologies make it difficult to meet the integrated needs of early fault prediction, accurate lifespan assessment, and hierarchical management and maintenance, and cannot effectively avoid the risk of power interruption. Their adaptability and maintenance effectiveness are poor, which restricts the development of refined and safe operation and maintenance throughout the entire life cycle of motors. Summary of the Invention

[0006] The purpose of this invention is to provide an AI system for drive motor failure diagnosis and remaining life grading early warning in order to solve the above-mentioned problems, thereby solving the problems mentioned in the background art.

[0007] To address the aforementioned issues, this invention provides a technical solution: an AI system for drive motor failure diagnosis and remaining life grading early warning, comprising: a multi-source data acquisition module, a feature fusion and extraction module, a lightweight hybrid AI judgment module, a grading early warning and control module, and a vehicle-cloud collaborative iteration module;

[0008] The multi-source data acquisition module is used to perform high-frequency synchronous acquisition of the electrical parameters, vibration signals and speed signals of the drive motor, and output the multi-source raw signals to the feature fusion and extraction module.

[0009] The feature fusion and extraction module is used to preprocess the received multi-source raw signals and extract multi-dimensional aging features, construct an aging feature library, and output multi-dimensional aging features to the lightweight hybrid AI judgment module.

[0010] The lightweight hybrid AI judgment module incorporates a pruned and optimized deep learning and machine learning fusion model and degradation trajectory prediction algorithm, used to output fault type and real-time failure probability based on the multi-dimensional aging characteristics. The remaining useful life (RUL) is also recorded and output to the graded early warning and control module.

[0011] The hierarchical early warning and control module is used to... RUL classifies risk levels and triggers matching control strategies, while generating a maintenance list;

[0012] The vehicle-cloud collaborative iteration module is used to transmit vehicle-mounted data back to the cloud and drive model iteration updates, send the updated model parameters to the vehicle-mounted terminal, and receive maintenance feedback results to transmit back to the cloud to form a closed loop.

[0013] Preferably, in the multi-source data acquisition module, the electrical parameters include at least three-phase current signals. , , and bus voltage signal The vibration signal includes at least the radial acceleration signal. With axial acceleration signal The rotational speed signal includes the real-time angular velocity signal. The sampling clocks of each channel are synchronized.

[0014] Preferably, the preprocessing of the feature fusion extraction module includes adaptive variational mode decomposition, which decomposes various signals from the multi-source original signal. Decomposed into K eigenmode components For k=1, 2, ..., K, its decomposition model is:

[0015] ;

[0016] ;

[0017] in, The center frequency of the Kth component is used. After decomposition, effective components are selected based on the permutation entropy threshold. The signal is reconstructed and the sideband energy ratio and harmonic distortion rate are extracted as latent aging characteristics.

[0018] Preferably, in the lightweight hybrid AI judgment module, the deep learning part uses a one-dimensional convolutional neural network with channel pruning to process temporal aging features, and the machine learning part uses an extreme gradient boosting tree to process statistical aging features.

[0019] The two are fused through adaptive weighting, with the fusion weight λ dynamically determined to minimize the validation set loss function. The final fault identification output is:

[0020] ;

[0021] Where 0 < λ < 1, This is the fault category probability vector output by a one-dimensional convolutional neural network. This is used to generate the fault category probability vector output by the extreme gradient boosting tree.

[0022] Preferably, the lightweight hybrid AI judgment module incorporates a dual-exponential degradation trajectory prediction algorithm, and the real-time failure probability... The comprehensive degradation index based on the output of the hybrid model The result obtained by mapping using the Sigmoid function is:

[0023] ;

[0024] in, , These are the model fitting parameters; the remaining useful life (RUL) is based on... The extrapolation of the bi-exponential degenerate trajectory yields the trajectory equation as follows:

[0025] ;

[0026] in, For the current moment, the comprehensive degradation index, , , These are the fitting coefficients. To predict the duration, the RUL is to satisfy... The smallest positive number value, This is a preset failure threshold.

[0027] Preferably, the comprehensive degradation index The following is obtained by weighted fusion of the multidimensional aging features:

[0028] , , ;

[0029] in, The m-th normalized aging characteristic value includes at least the electrical loss characteristic characterized by the sideband energy ratio, the mechanical wear characteristic characterized by the harmonic distortion rate, and the operating state characteristic characterized by the speed fluctuation. Let M be the weight corresponding to each feature, and M be the total number of features.

[0030] Preferably, the hierarchical early warning and control module will Mapped to RUL as a level 3 risk level:

[0031] Level I correspondence <0.4 and RUL>100 hours;

[0032] Level II corresponds to 0.4≤ <0.7 or 20 hours < RUL ≤ 100 hours;

[0033] Level III correspondence ≥0.7 or RUL≤20 hours;

[0034] Different levels correspond to different audible and visual alarm frequencies and torque limiting ratios. ,in For the rated torque, k is taken as 1.0, 0.7, and 0.4 respectively.

[0035] Preferably, the hierarchical early warning and control module also has a built-in fault type-maintenance measure mapping table, which is used to automatically generate a maintenance list based on the fault type and aging degree. The maintenance list includes the fault location component, suggested maintenance measures, and maintenance time window. , ,in, To preset the safety factor, 0 < ≤0.3.

[0036] Preferably, in the vehicle-cloud collaborative iteration module, the vehicle-mounted terminal uploads the original data segments, model output results, and maintenance feedback results within a preset time period before and after the fault triggering time to the cloud. After each accumulation of valid samples, the cloud retrains the benchmark model and sends the updated lightweight model parameters to the vehicle-mounted terminal through federated distillation, thereby realizing online model iteration.

[0037] Preferably, the cloud performs cluster analysis based on the raw data fragments, model output results and maintenance feedback results uploaded by each vehicle terminal, updates the global feature filtering rules and fault judgment thresholds, and sends the updated feature extraction parameters and model weights to each vehicle terminal.

[0038] Each vehicle terminal modifies its local feature mining algorithm and hierarchical warning threshold based on the issued parameters, and then sends the subsequent maintenance execution results back to the cloud, realizing a closed-loop iteration of collaborative evolution between the vehicle terminal and the cloud.

[0039] The beneficial effects of this invention are reflected in:

[0040] 1. This invention performs high-frequency synchronous acquisition of multiple signals such as current, voltage, vibration, and rotational speed. Combined with adaptive decomposition and feature mining algorithms, it effectively removes aging-related feature components from background noise, enabling the timely identification of latent degradation states such as winding micro-short circuits, early bearing pitting, and slight rotor eccentricity, providing a complete and reliable data foundation for subsequent diagnosis and early warning.

[0041] 2. This invention adaptively weights and fuses two types of models, balancing diagnostic accuracy and computational efficiency, and adapting to low-computing-power environments in vehicles. Simultaneously, it fits the entire lifecycle degradation trajectory of the motor based on a comprehensive degradation index, quantifying the remaining stable operating time, thus solving the problem of traditional solutions being able to diagnose faults but having uncertain lifespans, achieving a progress from qualitative to quantitative analysis.

[0042] 3. This invention consists of a two-layer architecture that combines real-time on-board diagnostics with deep cloud-based analysis. The on-board unit is responsible for real-time data collection and immediate early warning to ensure timely response. The cloud aggregates multi-vehicle operating data and maintenance feedback, dynamically optimizes feature selection rules and diagnostic thresholds through cluster analysis, and distributes the optimized parameters to each vehicle. The system continuously evolves with operational accumulation, constantly improving the accuracy of early warnings and gradually reducing false alarms and missed alarms. For new vehicle models or new fault modes, the cloud can also use transfer learning to assist in diagnosis with knowledge of similar vehicle models, shortening the performance ramp-up cycle in new scenarios. Attached Figure Description

[0043] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0044] Figure 1 This is a schematic diagram of the overall system workflow of the present invention;

[0045] Figure 2 This is a schematic diagram of the closed-loop iterative process of vehicle-cloud collaboration in this invention;

[0046] Figure 3 This is a schematic diagram of the internal judgment and early warning classification process of the AI ​​model of the present invention. Detailed Implementation

[0047] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are merely illustrative of the technical solution of the present invention and are therefore intended to limit the scope of protection of the present invention.

[0048] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application should have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0049] Example 1

[0050] like Figures 1-3 As shown, a drive motor failure diagnosis and remaining life grading early warning AI system includes: a multi-source data acquisition module, a feature fusion and extraction module, a lightweight hybrid AI judgment module, a grading early warning and control module, and a vehicle-cloud collaborative iteration module;

[0051] The multi-source data acquisition module is used to perform high-frequency synchronous acquisition of electrical parameters, vibration signals and speed signals of the drive motor, and output multi-source raw signals to the feature fusion and extraction module;

[0052] The feature fusion and extraction module is used to preprocess the received multi-source raw signals and extract multi-dimensional aging features, build an aging feature library, and output multi-dimensional aging features to the lightweight hybrid AI judgment module.

[0053] The lightweight hybrid AI judgment module incorporates a pruned and optimized deep learning and machine learning fusion model and degradation trajectory prediction algorithm, which is used to output fault type and real-time failure probability based on multi-dimensional aging characteristics. The remaining useful life (RUL) is also recorded and output to the hierarchical early warning and control module.

[0054] The tiered early warning and control module is used to... RUL classifies risk levels and triggers matching control strategies, while generating a maintenance list;

[0055] The vehicle-cloud collaborative iteration module is used to transmit vehicle-mounted data back to the cloud and drive model iteration updates. It sends the updated model parameters to the vehicle and receives maintenance feedback results back to the cloud to form a closed loop.

[0056] Furthermore, in the multi-source data acquisition module, the electrical parameters include at least three-phase current signals. , , and bus voltage signal The vibration signal includes at least the radial acceleration signal. With axial acceleration signal The rotational speed signal includes the real-time angular velocity signal. The sampling clocks of each channel are synchronized.

[0057] Furthermore, the preprocessing of the feature fusion extraction module includes adaptive variational mode decomposition, which converts various signals from the multi-source original signals. Decomposed into K eigenmode components For k=1, 2, ..., K, its decomposition model is:

[0058] ;

[0059] ;

[0060] in, The center frequency of the Kth component is used. After decomposition, effective components are selected based on the permutation entropy threshold. The signal is reconstructed and the sideband energy ratio and harmonic distortion rate are extracted as latent aging characteristics.

[0061] Furthermore, in the lightweight hybrid AI judgment module, the deep learning part uses a one-dimensional convolutional neural network with channel pruning to process temporal aging features, and the machine learning part uses an extreme gradient boosting tree to process statistical aging features.

[0062] The two are fused through adaptive weighting, with the fusion weight λ dynamically determined to minimize the validation set loss function. The final fault identification output is:

[0063] ;

[0064] Where λ is the fusion weight output by the deep learning model, 0 < λ < 1. This is the fault category probability vector output by a one-dimensional convolutional neural network. This is used to generate the fault category probability vector output by the extreme gradient boosting tree.

[0065] Furthermore, the lightweight hybrid AI judgment module incorporates a dual-exponential degradation trajectory prediction algorithm to predict failure probabilities in real time. Comprehensive degradation index based on hybrid model output The result obtained by mapping using the Sigmoid function is:

[0066] ;

[0067] in, , For model fitting parameters; Remaining useful life (RUL) is based on The extrapolation of the bi-exponential degenerate trajectory yields the trajectory equation as follows:

[0068] ;

[0069] in, For the current moment, the comprehensive degradation index, , , These are the fitting coefficients. To predict duration, RUL needs to meet... The smallest positive number value, This is a preset failure threshold.

[0070] Preferred comprehensive degradation index It is obtained by weighted fusion of multidimensional aging characteristics:

[0071] , , ;

[0072] in, Let m be the m-th normalized aging characteristic value. The normalized aging characteristic value includes at least the electrical loss characteristic characterized by the sideband energy ratio, the mechanical wear characteristic characterized by the harmonic distortion rate, and the operating state characteristic characterized by the speed fluctuation. Let M be the weight corresponding to each feature, and M be the total number of features.

[0073] Furthermore, the tiered early warning and control module will Mapped to RUL as a level 3 risk level:

[0074] Level I correspondence <0.4 and RUL>100 hours;

[0075] Level II corresponds to 0.4≤ <0.7 or 20 hours < RUL ≤ 100 hours;

[0076] Level III correspondence ≥0.7 or RUL≤20 hours;

[0077] Different levels correspond to different audible and visual alarm frequencies and torque limiting ratios. ,in For the rated torque, k is taken as 1.0, 0.7, and 0.4 respectively.

[0078] Furthermore, the tiered early warning and control module also includes a built-in fault type-maintenance measure mapping table, which is used to automatically generate a maintenance list based on the fault type and aging degree. The maintenance list includes the faulty component, recommended repair measures, and maintenance time window. , ,in, To preset the safety factor, 0 < ≤0.3.

[0079] Furthermore, in the vehicle-cloud collaborative iteration module, the vehicle-mounted terminal uploads the original data fragments, model output results, and maintenance feedback results within a preset time period before and after the fault triggering time to the cloud. After each valid sample is accumulated in the cloud, the baseline model is retrained, and the updated lightweight model parameters are sent to the vehicle-mounted terminal through federated distillation to realize online model iteration.

[0080] Furthermore, the cloud performs cluster analysis based on the raw data fragments, model output results, and maintenance feedback results uploaded by each vehicle terminal, updates the global feature selection rules and fault judgment thresholds, and sends the updated feature extraction parameters and model weights to each vehicle terminal.

[0081] Each vehicle terminal modifies its local feature mining algorithm and hierarchical warning threshold based on the issued parameters, and then sends the subsequent maintenance execution results back to the cloud, realizing a closed-loop iteration of collaborative evolution between the vehicle terminal and the cloud.

[0082] Example 2

[0083] This embodiment demonstrates real-time diagnostics and immediate hierarchical early warning for vehicle-mounted edge computing;

[0084] Specifically, the system is implemented as follows: A new energy commercial vehicle is equipped with this system and operates continuously under high-speed conditions. The system uses a multi-source data acquisition module to... The three-phase current of the drive motor is synchronously acquired at a sampling frequency of 10kHz. , , Bus voltage Radial vibration acceleration and angular velocity ;

[0085] The feature fusion extraction module extracts the original signal. Perform adaptive variational mode decomposition according to constraints. It is decomposed into K intrinsic mode components. After filtering the effective components based on the permutation entropy threshold, the sideband energy ratio is extracted as the electrical loss feature and the harmonic distortion rate is extracted as the mechanical wear feature.

[0086] In the lightweight hybrid AI judgment module, 1D-CNN and XGBoost are fused by weight λ, which is dynamically determined to be λ=0.62 by minimizing the cross-entropy loss on the validation set, and outputs a fault category probability vector. It can identify early micropitting failures in bearings;

[0087] Comprehensive degradation index The failure probability is obtained through Sigmoid mapping:

[0088] ,

[0089] in, =2.3、 =0.5, current =0.52, belonging to Level II risk, with a double exponential degradation trajectory. = + Extrapolation yields RUL = 68 hours;

[0090] The graded early warning and control module triggers a Level II audible and visual warning, with an alarm frequency of 1Hz and a torque limiting ratio. Simultaneously, a maintenance list is automatically generated: the fault is identified as bearing wear, the recommended repair measure is to inspect and replace the bearing, and the maintenance time window is specified. =RUL*0.2=13.6 hours. The vehicle-cloud collaborative iteration module uploads the original data fragments and diagnostic results within 30 seconds before and after the fault triggering time to the cloud.

[0091] Example 3

[0092] This embodiment is a cloud-based collaborative optimization model driven by cross-team data iteration and closed-loop evolution;

[0093] A logistics company owns 50 electric heavy trucks of the same model, all of which are deployed with this system. After the cloud server receives N=1500 valid samples from each vehicle terminal, it triggers global model retraining. The samples include fault trigger data fragments, intermediate model output results, and maintenance feedback results.

[0094] Cloud-based cluster analysis of multi-vehicle data revealed a systematic bias in the system's threshold for detecting winding aging faults under mountainous conditions, with a 12% higher false negative rate compared to plains conditions. Based on this, the cloud-based system updated its global feature filtering rules, adding the current harmonic component to the comprehensive degradation index under mountainous conditions. Weighting coefficients in , The characteristic weight for electrical loss is adjusted from 0.35 to 0.48, and corresponding adjustments are made accordingly. , To meet ;

[0095] Subsequently, the cloud distributed the updated lightweight model parameters and feature extraction rules to all 50 vehicle terminals via differential privacy federated distillation. After receiving the data, the vehicle terminals corrected the optimization strategy of the adaptive variational mode decomposition layer K in the local feature mining algorithm, changing it from a fixed K=5 to an adaptive selection of K=4-7 in mountainous conditions, and updated the graded warning threshold. In subsequent operation, a vehicle successfully warned of winding insulation aging fault 120 hours in advance in a long downhill mountainous condition, which is about 40% more advance warning than before the model update.

[0096] Example 4

[0097] This embodiment demonstrates how data augmentation and transfer learning can improve diagnostic accuracy for few-sample faults.

[0098] In the early stages of a new type of drive motor's market launch, the number of fault samples was scarce, with only 17 sets of rotor eccentricity fault data collected, which was insufficient to support high-performance model training. To address this, our system introduces a time-sliding window-based data augmentation strategy in the vehicle-cloud collaborative iteration module: Assuming the fault data sequence length is N=500 sampling points, and setting the window size W=3 and step size S=2, we extract... Each sample retains its temporal context, expanding the effective sample to more than 200 groups.

[0099] Meanwhile, the cloud uses a large amount of operating data collected under normal working conditions of the same vehicle model to pre-train the LSTM network model, obtain the health status representation knowledge of the motor, and then transfers the weight parameters of the encoder in the pre-trained model as the initial value to the fault prediction model through transfer learning, and uses the enhanced fault samples for fine-tuning training.

[0100] After the above processing, the fault diagnosis model of the new motor improved the identification accuracy of rotor eccentricity fault from the initial 67.3% to 94.1%, and the remaining life prediction error was reduced from ±18% to ±7.2%, achieving a diagnostic accuracy comparable to that of mature models.

[0101] Comparative Example 1

[0102] Based on traditional motor fault diagnosis systems, a single signal threshold alarm scheme is used.

[0103] In existing technologies, some motor fault diagnosis systems rely solely on a single current signal or vibration signal for fault determination. Their typical workflow involves analyzing the three-phase current signals of the motor... , , Extract the effective value of current. ,when Exceeding the preset fixed threshold Trigger an overcurrent alarm; or only monitor the total energy of vibration acceleration. ,when Exceeding the preset threshold The vibration exceeding the standard alarm is triggered at any time;

[0104] Problems with the above solution:

[0105] (1) The single signal dimension is limited and cannot distinguish different fault types such as winding aging, bearing wear, and rotor eccentricity. It is easy to misjudge non-faulty operating condition fluctuations as faults, and the false alarm rate is usually between 15% and 25%.

[0106] (2) Fixed threshold lacks adaptive capability. The operating parameters of the motor vary significantly under different working conditions such as start-up, stop, load change, and high temperature. Fixed threshold cannot adapt to changes in working conditions, resulting in a false negative rate of more than 30% under complex working conditions.

[0107] (3) It can only identify obvious faults that have exceeded the threshold, and has no ability to perceive early latent degradation features such as slow changes in current harmonic components and gradual rise of vibration spectrum sideband energy.

[0108] The comparison results with Example 2 are shown in the table below:

[0109] Comparison indicators Single signal threshold scheme This system signal source Single current or single vibration Synchronization of multiple sources including current, voltage, vibration, and rotation speed Feature extraction Manually set fixed threshold Adaptive variational mode decomposition + permutation entropy filtering Fault type identification Not supported support Early degradation identification Not available have Accuracy of bearing micro-pitting diagnosis Approximately 62% Approximately 94% Missed detection rate under complex working conditions More than 30% Less than 8%

[0110] Conclusion: Comparative Example 1, due to its single signal source and reliance on a fixed threshold, cannot achieve early degradation identification and fault type differentiation. Example 2, through multi-source synchronous acquisition, adaptive variational mode decomposition, and multi-dimensional aging feature fusion, achieves accurate capture of latent degradation features and high-precision identification of fault types, with a diagnostic accuracy rate improved by approximately 32% compared to Comparative Example 1.

[0111] Comparative Example 2

[0112] A CNN / RNN heavy-duty model solution for motor fault diagnosis systems based on traditional single AI models;

[0113] In existing technologies, some solutions employ a single deep learning model for motor fault diagnosis. A typical approach involves converting the collected current and vibration signals into a two-dimensional time-frequency spectrum using a short-time Fourier transform, then inputting this spectrum into a heavy-duty convolutional neural network such as ResNet-50 for fault classification. Alternatively, a three-layer stacked LSTM network is used, with the input time-series sampling sequence. Through hidden layer states The fault category is output after recursive calculation;

[0114] Problems with the above solution:

[0115] (1) Heavy models have a large number of parameters. ResNet-50 has about 25M parameters and LSTM three-layer stack has about 15M parameters. They cannot be deployed on vehicle-mounted low-computing embedded platforms. The computing power is usually <5TOPS. They can only run in the cloud or industrial control computer and cannot meet the needs of vehicle-mounted real-time diagnosis.

[0116] (2) The single model is acceptable for identifying obvious faults, such as short circuits and open circuits, but for early and weak faults such as early micro-pitting of bearings and slight short circuits between winding turns, the misjudgment and omission rate is high, usually between 20% and 35%, because the features are not obvious and the model lacks a multi-modal feature hierarchical analysis mechanism.

[0117] (3) A single model cannot simultaneously handle both classification and regression tasks. That is, it is difficult to quantify the failure probability and remaining service life while outputting the failure type. An additional independent regression model needs to be trained, which increases the system complexity.

[0118] The comparison results with Example 3 are shown in the table below:

[0119] Comparison indicators Single heavy AI model This system Model type Single model Hybrid Model Model parameter count 15M-25M Approximately 2.1 meters after pruning. Feasibility of vehicle-mounted deployment Not feasible feasible Early weak fault missed rate 20%-35% Less than 10% Should the fault type + RUL be output simultaneously? no yes Single-frame inference latency Approximately 180ms Approximately 35ms

[0120] Conclusion: Comparative Example 2 cannot be deployed in vehicles due to its excessively large model parameters, and its single model architecture lacks multimodal hierarchical parsing capabilities, resulting in insufficient early and subtle fault identification capabilities. Example 3 corresponds to the lightweight hybrid AI model of this system. Through feature-level fusion of 1D-CNN and XGBoost and model pruning, the number of parameters is compressed by about 85%-90% without compromising diagnostic accuracy, enabling real-time inference at the vehicle edge and reducing the early fault false negative rate by about 15-25%.

[0121] Comparative Example 3

[0122] The traditional local fixed-model motor fault diagnosis system lacks a collaborative iteration scheme.

[0123] In the existing technology, some motor fault diagnosis systems adopt a local fixed model scheme, that is, the vehicle terminal has a pre-trained fixed diagnostic model built in, and the model parameters are not updated after leaving the factory. The typical practice is to collect motor fault data under laboratory or standard working conditions before the vehicle leaves the factory, train the diagnostic model and fix it in the vehicle ECU. During the vehicle operation, fault identification and alarm are performed only by relying on the local fixed model.

[0124] Problems with the above solution:

[0125] (1) The model training data comes from laboratory standard working conditions or limited road conditions, which cannot cover the complex working conditions in actual vehicle operation, such as long downhill in mountainous areas, high temperature and high humidity, and frequent start-stop in cities, resulting in insufficient generalization ability of the model in actual scenarios.

[0126] (2) The model cannot use newly generated fault data and maintenance feedback results during vehicle operation to evolve itself. The diagnostic accuracy gradually decreases with the increase of mileage. Typically, the diagnostic accuracy drops by 15%-20% after 50,000 kilometers of operation.

[0127] (3) The fault data and maintenance experience of each vehicle are independent of each other, making it impossible to achieve cross-team knowledge sharing and collaborative evolution, resulting in the recurrence of the same fault on different vehicles without early warning.

[0128] (4) When a new type of motor or a new model is put into the market, a large amount of fault data needs to be collected again and the model needs to be trained from scratch. The development and verification cycle is usually 6-12 months, which affects the rapid iteration and deployment of products.

[0129] The comparison results with Example 3 are shown in the table below:

[0130] Comparison indicators Local fixed model This system Model update mechanism Fixed at the factory, will remain unchanged for life. Continuous cloud-based iteration and online updates Cross-team knowledge sharing Not available have Operating condition adaptability Difference powerful Diagnostic accuracy decreases after 50,000 kilometers Decrease of 15%-20% The decline shall not exceed 3% New model adaptation period 6-12 months 1-2 months Fault warning lead time Approximately 25 hours on average Approximately 68 hours on average

[0131] Conclusion: Comparative Example 3, due to its use of a local fixed model scheme, cannot achieve cross-vehicle knowledge sharing and continuous model evolution, and the diagnostic accuracy decreases significantly with running time. Example 3, through a vehicle-cloud collaborative architecture and federated distillation mechanism, achieves continuous iteration of the global model driven by multi-vehicle data, reducing the rate of decline in diagnostic accuracy by about 85%, shortening the adaptation cycle for new vehicle models by about 80%, and increasing the fault warning advance by about 2.7 times.

[0132] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. An AI system for drive motor failure diagnosis and remaining life grading early warning, characterized in that, include: Multi-source data acquisition module, feature fusion and extraction module, lightweight hybrid AI analysis module, hierarchical early warning and control module, and vehicle-cloud collaborative iteration module; The multi-source data acquisition module is used to perform high-frequency synchronous acquisition of the electrical parameters, vibration signals and speed signals of the drive motor, and output the multi-source raw signals to the feature fusion and extraction module. The feature fusion and extraction module is used to preprocess the received multi-source raw signals and extract multi-dimensional aging features, construct an aging feature library, and output multi-dimensional aging features to the lightweight hybrid AI judgment module. The lightweight hybrid AI judgment module incorporates a pruned and optimized deep learning and machine learning fusion model and degradation trajectory prediction algorithm, used to output fault type and real-time failure probability based on the multi-dimensional aging characteristics. The remaining useful life (RUL) is also recorded and output to the graded early warning and control module. The hierarchical early warning and control module is used to... RUL classifies risk levels and triggers matching control strategies, while generating a maintenance list; The vehicle-cloud collaborative iteration module is used to transmit vehicle-mounted data back to the cloud and drive model iteration updates, send the updated model parameters to the vehicle-mounted terminal, and receive maintenance feedback results to transmit back to the cloud to form a closed loop.

2. The AI ​​system for drive motor failure diagnosis and remaining life grading early warning according to claim 1, characterized in that: In the multi-source data acquisition module, the electrical parameters include at least three-phase current signals. , , and bus voltage signal The vibration signal includes at least the radial acceleration signal. With axial acceleration signal The rotational speed signal includes the real-time angular velocity signal. The sampling clocks of each channel are synchronized.

3. The AI ​​system for drive motor failure diagnosis and remaining life grading early warning according to claim 1, characterized in that: The preprocessing of the feature fusion extraction module includes adaptive variational mode decomposition, which decomposes various types of signals from the multi-source original signals. Decomposed into K eigenmode components For k=1, 2, ..., K, its decomposition model is: ； ； in, The center frequency of the Kth component is used. After decomposition, effective components are selected based on the permutation entropy threshold. The signal is reconstructed and the sideband energy ratio and harmonic distortion rate are extracted as latent aging characteristics.

4. The AI ​​system for drive motor failure diagnosis and remaining life grading early warning according to claim 1, characterized in that: In the lightweight hybrid AI judgment module, the deep learning part uses a one-dimensional convolutional neural network with channel pruning to process temporal aging features, and the machine learning part uses an extreme gradient boosting tree to process statistical aging features. The two are fused through adaptive weighting, with the fusion weight λ dynamically determined to minimize the validation set loss function. The final fault identification output is: ； Where 0 < λ < 1, This is the fault category probability vector output by a one-dimensional convolutional neural network. This is used to boost the fault category probability vector of the extreme gradient boosting tree output.

5. The AI ​​system for drive motor failure diagnosis and remaining life grading early warning according to claim 1, characterized in that: The lightweight hybrid AI judgment module incorporates a dual-exponential degradation trajectory prediction algorithm, and the real-time failure probability... The comprehensive degradation index based on the output of the hybrid model The result obtained by mapping using the Sigmoid function is: ； in, , These are the model fitting parameters; the remaining useful life (RUL) is based on... The extrapolation of the bi-exponential degenerate trajectory yields the trajectory equation as follows: ； in, For the current moment, the comprehensive degradation index, , , These are the fitting coefficients. To predict the duration, the RUL is to satisfy... The smallest positive number value, This is a preset failure threshold.

6. The AI ​​system for drive motor failure diagnosis and remaining life grading early warning according to claim 5, characterized in that: The comprehensive degradation index The following is obtained by weighted fusion of the multidimensional aging features: ， ， ； in, The m-th normalized aging characteristic value includes at least the electrical loss characteristic characterized by the sideband energy ratio, the mechanical wear characteristic characterized by the harmonic distortion rate, and the operating state characteristic characterized by the speed fluctuation. Let M be the weight corresponding to each feature, and M be the total number of features.

7. The AI ​​system for drive motor failure diagnosis and remaining life grading early warning according to claim 1, characterized in that: The hierarchical early warning and control module will Mapped to RUL as a level 3 risk level: Level I correspondence <0.4 and RUL>100 hours; Level II corresponds to 0.4≤ <0.7 or 20 hours < RUL ≤ 100 hours; Level III correspondence ≥0.7 or RUL≤20 hours; Different levels correspond to different audible and visual alarm frequencies and torque limiting ratios. ,in For the rated torque, k is taken as 1.0, 0.7, and 0.4 respectively.

8. The AI ​​system for drive motor failure diagnosis and remaining life grading early warning according to claim 7, characterized in that: The hierarchical early warning and control module also has a built-in fault type-maintenance measure mapping table, which is used to automatically generate a maintenance list based on the fault type and aging degree. The maintenance list includes the fault location component, suggested maintenance measures, and maintenance time window. , ,in, To preset the safety factor, 0 < ≤0.

3.

9. The AI ​​system for drive motor failure diagnosis and remaining life grading early warning according to claim 1, characterized in that: In the vehicle-cloud collaborative iteration module, the vehicle-mounted terminal uploads the original data segments, model output results, and maintenance feedback results within a preset time period before and after the fault triggering time to the cloud. After each valid sample is accumulated, the cloud retrains the benchmark model and sends the updated lightweight model parameters to the vehicle-mounted terminal through federated distillation, thereby realizing online model iteration.

10. The AI ​​system for drive motor failure diagnosis and remaining life grading early warning according to claim 9, characterized in that: The cloud performs cluster analysis based on the raw data fragments, model output results and maintenance feedback results uploaded by each vehicle terminal, updates the global feature screening rules and fault judgment thresholds, and sends the updated feature extraction parameters and model weights to each vehicle terminal. Each vehicle terminal modifies its local feature mining algorithm and hierarchical warning threshold based on the issued parameters, and then sends the subsequent maintenance execution results back to the cloud, realizing a closed-loop iteration of collaborative evolution between the vehicle terminal and the cloud.

Images