Fault diagnosis method of electric drive system and related product

Abstract

The invention discloses a fault diagnosis method of an electric driving system and a related product. The method comprises the following steps: acquiring multi-source data of the electric driving system; preprocessing the multi-source data to obtain the processed multi-source data; inputting the multi-source data into a diagnosis model to obtain a fault diagnosis result of the electric drive system output by the diagnosis model; the method comprises the following steps: constructing a diagnosis model comprising an image branch, a sequence branch, a structured branch and a fusion layer, wherein the image branch can extract image features of processed multi-source data; the sequence branch can extract sequence features of the processed multi-source data; the structured branches can extract the structural features of the processed multi-source data. And the fusion layer further performs adaptive integration on the three types of features, so that a fault diagnosis result is accurately output, automatic fault diagnosis of the electric drive system is realized, and the accuracy of fault diagnosis is improved.

Description

Technical Field

[0001] This application relates to the field of automotive technology, and in particular to a fault diagnosis method and related products for an electric drive system. Background Technology

[0002] As a core component of electric vehicles, the NVH (noise, vibration, and harshness) performance of the electric drive system directly impacts the overall vehicle quality. However, NVH issues in electric drive systems are highly complex, stemming from the strong coupling effects of multiple physical fields, including electromagnetic, mechanical, control, and thermal fields. Traditional diagnostic methods heavily rely on the experience of NVH experts, typically involving manual analysis of spectrum diagrams, order diagrams, and time-frequency color graphs, combined with theoretical models and empirical rules for fault location. These methods have significant limitations: first, they are highly subjective, with different experts potentially reaching different conclusions from the same data; second, they are inefficient, struggling to process massive amounts of test data quickly; third, knowledge is difficult to quantify, standardize, and transfer; and fourth, they are prone to missed or incorrect diagnoses when dealing with multiple coupled faults or novel fault modes. Summary of the Invention

[0003] To address the aforementioned issues, this application provides a fault diagnosis method and related products for electric drive systems, aiming to achieve automated fault diagnosis of electric drive systems and improve the accuracy of fault diagnosis.

[0004] The embodiments of this application disclose the following technical solutions:

[0005] The first aspect of this application provides a fault diagnosis method for an electric drive system, the method comprising:

[0006] Acquire multi-source data from the electric drive system;

[0007] The multi-source data is preprocessed to obtain processed multi-source data;

[0008] The multi-source data is input into a diagnostic model to obtain the fault diagnosis result of the electric drive system output by the diagnostic model. The diagnostic model includes an image branch, a sequence branch, a structured branch, and a fusion layer. The image branch is used to extract image features from the processed multi-source data. The sequence branch is used to extract sequence features from the processed multi-source data. The structured branch is used to extract structural features from the processed multi-source data. The fusion layer is used to obtain the fault diagnosis result of the electric drive system based on image features, sequence features, and structural parameters.

[0009] Optionally, acquiring multi-source data from the electric drive system specifically includes:

[0010] Collect basic data of the motor, reducer, and housing in the electric drive system; the basic data includes vibration acceleration data and voltage signals.

[0011] Collect data on motor speed, output torque, DC bus voltage, three-phase current, temperature, and operating mode flags in the electric drive system;

[0012] The system acquires speed phase data from a speed sensor or motor resolver in an electric drive system; the speed phase data includes instantaneous speed pulses and phase signals.

[0013] The basic data of the motor, the basic data of the reducer, the basic data of the housing, the motor speed, the output torque, the DC bus voltage, the three-phase current, the temperature data, the operating mode flag, and the speed phase data are used as the multi-source data.

[0014] Optionally, the preprocessing of the multi-source data to obtain processed multi-source data specifically includes:

[0015] The image modal data in the multi-source data is standardized to obtain standardized multi-source data;

[0016] The sequence modal data in the multi-source data are denoised and standardized to obtain denoised multi-source data.

[0017] Structured modal data is obtained by calculating the structural parameters in the multi-source data.

[0018] The structure parameters in the multi-source data are replaced with the structured modal data to obtain the processed multi-source data.

[0019] Optionally, the step of inputting the multi-source data into the diagnostic model to obtain the fault diagnosis result of the electric drive system output by the diagnostic model specifically includes:

[0020] The image modal data from the multi-source data is input into the image branch of the diagnostic model to obtain image features; the image branch includes a pre-trained 2D-CNN network.

[0021] The sequence modality data from the multi-source data is input into the sequence branch of the diagnostic model to obtain sequence features; the sequence branch includes a 1D-CNN network or a combined network; the combined network includes a 1D-CNN network and an LSTM network.

[0022] The structured modal data from the multi-source data are input into the structured branch of the diagnostic model to obtain structural features; the structured branch includes a fully connected network.

[0023] The image features, sequence features, and structural features are input into the fusion layer of the diagnostic model to obtain the fault diagnosis results of the electric drive system output by the diagnostic model.

[0024] Optionally, the training process of the diagnostic model specifically includes:

[0025] Obtain a training dataset; the training dataset includes historical multi-source data of the electric drive system and the corresponding fault diagnosis results of the historical multi-source data;

[0026] Constructing a diagnostic model to be trained based on deep learning;

[0027] The historical multi-source data in the training dataset is input into the diagnostic model to obtain the output of the diagnostic model to be trained;

[0028] The loss value is calculated based on the output of the diagnostic model to be trained and the fault diagnosis results corresponding to the historical multi-source data.

[0029] With the goal of minimizing the loss value, the parameters of the diagnostic model to be trained are adjusted using the backpropagation algorithm. If the training termination condition is met, the training ends, and the diagnostic model is obtained.

[0030] If the training termination condition is not met, the process returns to the step of inputting historical multi-source data from the training dataset into the diagnostic model to obtain the output of the diagnostic model to be trained.

[0031] Optionally, the method further includes:

[0032] Based on the fault diagnosis results, a diagnostic report is obtained from the diagnostic report database to repair the electric drive system.

[0033] A second aspect of this application provides a fault diagnosis device for an electric drive system, the fault diagnosis device for the electric drive system comprising:

[0034] The acquisition module is used to acquire multi-source data from the electric drive system;

[0035] The processing module is used to preprocess the multi-source data to obtain processed multi-source data;

[0036] A diagnostic module is used to input the multi-source data into a diagnostic model to obtain the fault diagnosis result of the electric drive system output by the diagnostic model. The diagnostic model includes an image branch, a sequence branch, a structured branch, and a fusion layer. The image branch is used to extract image features from the processed multi-source data. The sequence branch is used to extract sequence features from the processed multi-source data. The structured branch is used to extract structural features from the processed multi-source data. The fusion layer is used to obtain the fault diagnosis result of the electric drive system based on image features, sequence features, and structural parameters.

[0037] A third aspect of this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the fault diagnosis method for the electric drive system provided in the first aspect.

[0038] The fourth aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the fault diagnosis method for the electric drive system provided in the first aspect.

[0039] The fifth aspect of this application provides a computer program product, including a computer program that, when executed by a processor, implements the fault diagnosis method for the electric drive system provided in the first aspect.

[0040] Compared with the prior art, this application has the following beneficial effects:

[0041] This application includes acquiring multi-source data of an electric drive system; preprocessing the multi-source data to obtain processed multi-source data; inputting the multi-source data into a diagnostic model to obtain a fault diagnosis result of the electric drive system output by the diagnostic model; the diagnostic model includes an image branch, a sequence branch, a structured branch, and a fusion layer; the image branch is used to extract image features from the processed multi-source data; the sequence branch is used to extract sequence features from the processed multi-source data; the structured branch is used to extract structural features from the processed multi-source data; and the fusion layer is used to obtain the fault diagnosis result of the electric drive system based on image features, sequence features, and structural parameters.

[0042] This application constructs a diagnostic model comprising image branch, sequence branch, structured branch, and a fusion layer: the image branch extracts image features from the processed multi-source data; the sequence branch extracts sequence features from the processed multi-source data; and the structured branch extracts structural features from the processed multi-source data. The fusion layer further adaptively integrates the above three types of features, thereby accurately outputting fault diagnosis results. This not only significantly improves the ability to identify complex coupled faults but also effectively suppresses misjudgments caused by interference from a single data source, enhancing the robustness of the diagnosis. Compared to traditional manual analysis relying on expert experience or rule-driven expert systems, this achieves automated fault diagnosis of electric drive systems, improving the accuracy of fault diagnosis. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 A flowchart illustrating a fault diagnosis method for an electric drive system provided in this application embodiment;

[0045] Figure 2 This is a structural diagram of a fault diagnosis device for an electric drive system provided in an embodiment of this application. Detailed Implementation

[0046] As described above, the working principle of an existing technical solution is as follows: experts predefine a series of diagnostic rules in the form of "if-then" (e.g., "if the amplitude of the gear meshing frequency exceeds threshold A and its sideband amplitude exceeds threshold B, then it is determined to be gear wear"). The system preprocesses the input NVH data (such as the spectrum) and extracts manual features (such as the amplitude of specific frequency components). Then, the extraction results are matched with the thresholds and logical conditions in the rule base, and the corresponding fault codes and suggestions are output.

[0047] However, such rule bases struggle to cover all potential faults, especially unknown faults or complex fault modes arising from the coupling of multiple physical fields such as electromagnetics, mechanics, and heat, resulting in low diagnostic coverage and weak generalization ability. Furthermore, relying on manually designed features (such as single-frequency amplitude) cannot effectively capture deep fault patterns in the time-frequency domain (such as texture structures and spatiotemporal evolution patterns in color images), leading to insufficient feature representation and limiting diagnostic accuracy. In addition, the system lacks the ability to learn autonomously and iteratively optimize from new data; rule updates require manual intervention, resulting in high maintenance costs, slow response times, and difficulty adapting to the rapidly evolving NVH challenges of electric drive systems. Overall, this type of technology has limited intelligence and is insufficient to meet the needs of modern high-performance electric drive systems in terms of diagnostic accuracy, adaptability, and efficiency.

[0048] In view of the above problems, this application provides a fault diagnosis generation method and related products for electric drive systems. The method includes: acquiring multi-source data of the electric drive system; preprocessing the multi-source data to obtain processed multi-source data; inputting the multi-source data into a diagnostic model to obtain the fault diagnosis result of the electric drive system output by the diagnostic model; the diagnostic model includes an image branch, a sequence branch, a structured branch, and a fusion layer; the image branch is used to extract image features from the processed multi-source data; the sequence branch is used to extract sequence features from the processed multi-source data; the structured branch is used to extract structural features from the processed multi-source data; and the fusion layer is used to obtain the fault diagnosis result of the electric drive system based on image features, sequence features, and structural parameters. This application extracts image features, temporal features, and structural features from the processed multi-source data through the image branch, sequence branch, and structured branch, respectively, and the fusion layer adaptively integrates these features to achieve high-precision and automated fault diagnosis. Compared with traditional methods that rely on human experience or rule-based systems, this significantly improves the ability to identify complex coupled faults and the robustness of diagnosis.

[0049] Explanation of relevant professional terms:

[0050] NVH: Noise, Vibration, and Harshness, used to evaluate the comfort and noise and vibration performance of a vehicle or mechanical system.

[0051] Electric drive system: This includes the electric vehicle powertrain system consisting of a motor, reducer, inverter, drive shaft, etc.

[0052] Intelligent diagnosis: A method that uses artificial intelligence technology to automatically identify the cause and location of equipment failures.

[0053] Feature extraction: Extracting discriminative statistical or spectral features from the raw data for model training.

[0054] Order analysis: A frequency analysis method based on rotational speed, used to identify vibration components related to rotational speed.

[0055] Deep learning: A machine learning method based on deep neural networks that can automatically learn the hierarchical structure of features from raw data.

[0056] NVH (Noise, Vibration, and Harshness) Color Maps: A form of 3D data visualization, typically using time and frequency as coordinates, with color depth representing signal amplitude (such as sound pressure level or vibration acceleration level). Examples include Colormaps and Campbell Diagrams. **Sound Pressure / Vibration Level Color Map (Colormap):** 1. A 3D graph with time as the Y-axis and frequency as the X-axis, where color depth represents the magnitude of sound pressure or vibration acceleration level. It visually displays the energy distribution at different times and frequencies. 2. A 3D graph with rotational speed as the Y-axis and frequency as the X-axis, where color depth represents the magnitude of sound pressure or vibration acceleration level. It visually displays the energy distribution at different rotational speeds and frequencies. **Order Cut:** A 3D graph with rotational speed as the X-axis and order amplitude as the Y-axis, displaying vibration noise components strictly proportional to rotational speed. This is crucial for identifying faults caused by rotating components such as gears, bearings, and electromagnetic waves. **Campbell Diagram:** A 3D graph of frequency, rotational speed, and amplitude, clearly showing which frequency components change with rotational speed, used to identify structural resonance.

[0057] Multimodal fusion refers to the technique of jointly analyzing data of different types (such as images, sequences, and numerical values).

[0058] Convolutional Neural Networks (CNNs): A deep learning architecture specifically designed for processing grid-like data (such as images and one-dimensional time series). It automatically extracts local features (such as edges and textures) by sliding convolutional kernels across the input data. In this invention, 2D-CNNs are used to identify fault-related spatial patterns (such as vertical bright bands representing transient resonance, diagonal lines representing rotational order, and scattered points representing random impacts) from NVH color images; 1D-CNNs are used to capture features such as local impacts and peaks from vibration waveforms or spectral curves.

[0059] Long Short-Term Memory (LSTM) networks are a special type of recurrent neural network (RNN) that effectively learns long-term dependencies through internal gating mechanisms (input gate, forget gate, output gate), making them particularly adept at processing and predicting dynamic patterns in time series. In this invention, LSTM is used to analyze the evolution of NVH signals over time, such as capturing vibration and noise response patterns under transient conditions like torque switching and temperature changes.

[0060] Fully connected networks: a basic neural network structure in which each neuron in each layer is connected to all neurons in the next layer. The core function of this network is to learn complex, nonlinear combinations of relationships between input features. In this invention, it is used to process structured feature vectors (such as traditional statistical features, operating parameters, and theoretical fault frequencies). Through multi-layer nonlinear transformations, it learns the hidden, high-dimensional mapping relationship between these quantitative indicators and fault modes, acting as a "comprehensive decision-maker." For example, "a significant increase in the amplitude of gear meshing frequency under high load" corresponds to gear squealing, and "the appearance of bearing characteristic frequencies accompanied by high temperature" corresponds to poor lubrication. These abstract numerical correlations are highly relevant to the fault mechanism, requiring no manually defined rules.

[0061] Fusion Layer: Located in a deep learning model, this specific network layer is responsible for integrating feature information from different network branches. In the multimodal model of this invention, the fusion layer receives feature vectors output from the image branch (2D-CNN), sequence branch (1D-CNN / LSTM), and structured branch (fully connected network), and combines them into a unified, more comprehensive joint feature vector through specific operations (such as concatenation, weighted summation, or attention mechanisms). This layer is key to achieving complementary multi-source information and collaborative decision-making, enabling the model to comprehensively consider "what it sees," "what it hears," and "what it infers" to make a final judgment.

[0062] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0063] Figure 1 A flowchart illustrating a fault diagnosis method for an electric drive system provided in this application embodiment is shown below. Figure 1 As shown, a fault diagnosis method for an electric drive system includes:

[0064] S101: Acquire multi-source data from the electric drive system.

[0065] This application does not limit the type of multi-source data. For example, it may include multi-source NVH data, such as time-domain waveforms, frequency-domain spectra, sound pressure or vibration time-frequency color graphs, order slices, etc., covering a variety of signal forms and representation dimensions to comprehensively reflect the operating status of the electric drive system.

[0066] S102: Preprocess the multi-source data to obtain processed multi-source data.

[0067] This application does not limit the preprocessing methods for multi-source data. Various preprocessing techniques can be flexibly adopted according to actual needs. For example, data can be standardized or normalized to eliminate differences in dimensions, or filtering, wavelet denoising, empirical mode decomposition and other methods can be applied to suppress noise interference. Operations such as resampling, alignment, missing value imputation, and time-frequency transformation can also be included to improve the consistency between data quality and model input, and lay a reliable foundation for subsequent feature extraction and fault diagnosis.

[0068] S103: Input the multi-source data into the diagnostic model to obtain the fault diagnosis results of the electric drive system output by the diagnostic model.

[0069] This application does not limit the diagnostic model, but considering that the multi-source data in this application may include various types of data, the diagnostic model includes an image branch, a sequence branch, a structured branch, and a fusion layer; the image branch is used to extract image features from the processed multi-source data; the sequence branch is used to extract sequence features from the processed multi-source data; the structured branch is used to extract structural features from the processed multi-source data; and the fusion layer is used to obtain the fault diagnosis result of the electric drive system based on image features, sequence features, and structural parameters.

[0070] This application aims to address common industry pain points in the diagnosis of NVH (Noise, Vibration, and Harshness) problems in current electric drive systems, such as heavy reliance on expert subjective experience, long diagnostic cycles, low efficiency of manual analysis, and the difficulty of accurately identifying and locating complex faults caused by the coupling of multiple physical fields such as electromagnetics, mechanics, and heat using traditional methods. To address this, this application deeply integrates artificial intelligence technology with professional knowledge in the field of NVH for electric drive systems to construct an end-to-end intelligent diagnostic model. This model can directly use raw or pre-processed NVH test data (such as acoustic vibration color images, spectrum curves, order diagrams, time-domain waveforms, etc.) as input to automatically complete feature extraction, fault identification, and root cause analysis, and can further generate targeted optimization suggestions. By automating, intelligentizing, and standardizing the diagnostic process, it significantly improves the accuracy, response speed, and engineering practicality of fault identification, effectively supporting efficient decision-making and rapid iteration in the closed loop of product development, verification testing, and after-sales quality issues for electric drive systems.

[0071] The above describes the main technical solution of this application. Further implementations of the main technical solution are now introduced. Details are as follows:

[0072] Regarding the acquisition of multi-source data from the electric drive system in step S101, this application provides an optional embodiment:

[0073] Collect basic data on the motor, reducer, and housing in the electric drive system; the basic data includes vibration acceleration data and voltage signals.

[0074] This application collects fundamental physical signals during the operation of an electric drive system by arranging vibration acceleration sensors and microphones on key components (including the motor, reducer, and housing). The fundamental data mainly includes two categories: first, vibration acceleration signals from the vibration acceleration sensors, used to characterize the structural dynamic response; and second, acoustic or electrical raw voltage signals from the microphone or voltage measurement unit, reflecting noise radiation characteristics or the electromagnetic behavior of the motor. These raw physical signals serve as input sources for subsequent NVH analysis and intelligent diagnostics, comprehensively capturing the multi-dimensional dynamic characteristics of the system under various operating conditions, providing high-fidelity data support for fault identification and mechanism analysis.

[0075] Collect data on motor speed, output torque, DC bus voltage, three-phase current, temperature, and operating mode flags in the electric drive system.

[0076] This application further collects key control and status signals during the operation of the electric drive system, including motor speed, output torque, DC bus voltage, three-phase current, component temperatures, and operating mode flags (such as drive, energy recovery, and idling). This data is typically acquired in real-time via the vehicle's CAN bus, providing a complete operating context of the system during testing or operation in standard message format. As indispensable auxiliary information for NVH phenomenon analysis, these control and status signals not only help to accurately define the operating boundary conditions for fault occurrence but also effectively distinguish NVH response differences caused by factors such as operating mode switching, load changes, or thermal evolution, thereby significantly improving the accuracy and interpretability of fault diagnosis.

[0077] The system acquires speed phase data from a speed sensor or motor resolver in an electric drive system; the speed phase data includes instantaneous speed pulses and phase signals.

[0078] This application also collects speed phase data from high-precision speed sensors or motor resolvers in the electric drive system, including instantaneous speed pulse signals and high-resolution phase signals. These speed phase signals accurately reflect the instantaneous angular velocity and rotational position of the motor shaft during operation, providing a high-precision time reference for key NVH processing steps such as order analysis, time-frequency synchronization, and phase alignment of vibration / noise events. Especially under variable speed conditions, this signal plays a crucial role in achieving accurate order slicing, fault feature extraction, and spatiotemporal alignment of multi-source data, serving as vital supporting data for improving the accuracy of coupled fault diagnosis.

[0079] The basic data of the motor, the basic data of the reducer, the basic data of the housing, the motor speed, the output torque, the DC bus voltage, the three-phase current, the temperature data, the operating mode flag, and the speed phase data are used as the multi-source data.

[0080] The multi-source data collected in this application are updated synchronously in real time during system operation, ensuring strict alignment of various signals such as vibration, noise, torque, speed, current, and temperature under a unified time reference. Through high-precision clock synchronization or hardware triggering mechanisms, time drift and sampling deviations between multiple sensors are effectively eliminated, guaranteeing the precise consistency of all physical quantities on the time axis. This high-fidelity spatiotemporal synchronization capability provides a reliable data foundation for subsequent multimodal feature fusion, order analysis, fault event localization, and causal correlation mining, significantly improving the accuracy and reliability of NVH diagnosis.

[0081] Regarding the preprocessing of the multi-source data in S102 to obtain processed multi-source data, this application provides an optional embodiment:

[0082] The image modal data in the multi-source data are standardized to obtain standardized multi-source data.

[0083] This application standardizes image modal data (such as time-frequency color images, Campbell diagrams, and order slices of sound pressure level or vibration level) from multi-source data sources. This standardization process includes, but is not limited to, pixel value normalization, size unification, color mapping consistency correction, and background noise suppression. The result is standardized image data with consistent format, numerical range, and clear visual and semantic features. This standardization process not only improves the comparability of image data across different operating conditions and test batches but also provides a reliable input foundation for stable feature extraction and cross-sample generalization capabilities of image branches in subsequent diagnostic models.

[0084] The sequence modal data in the multi-source data are denoised and standardized to obtain denoised multi-source data.

[0085] This application performs denoising and standardization on sequence modal data from multi-source datasets to improve signal quality and enhance the consistency of model inputs. Specifically, for one-dimensional time-series signals, methods such as filtering, wavelet denoising, and empirical mode decomposition are used to effectively suppress environmental noise and measurement interference. Normalization or standardization strategies (such as Z-score and Min-Max scaling) are combined to unify the amplitude scale under different channels or operating conditions. The processed sequence data includes: time-domain vibration or sound pressure waveforms at key measurement points, frequency-domain spectrum (FFT) curves under steady-state operating conditions, and envelope spectrum curves for fault diagnosis of rotating components such as bearings. These high-quality, structurally consistent time-series features not only preserve the fault-sensitive information in the original signal but also significantly improve the accuracy and robustness of subsequent sequence branches in dynamic pattern recognition and fault feature extraction.

[0086] Structured modal data is obtained by calculating the structural parameters in the multi-source data.

[0087] The structure parameters in the multi-source data are replaced with the structured modal data to obtain the processed multi-source data.

[0088] This application uses original structural parameters (such as the number of motor pole pairs, the number of gear teeth, and the geometric dimensions of bearings) from multi-source data and real-time acquired operating condition signals (such as speed and torque) to generate high-information-density structured modal data through built-in physical models and engineering formulas. The structured modal data is organized in the form of feature vectors and contains three types of key information: first, traditional time-frequency domain statistical features extracted from vibration or current signals, such as RMS value, peak value, kurtosis, and amplitude of the main peak in the spectrum; second, core operating parameters reflecting the test conditions, such as average speed, average output torque, and lubricating oil temperature; and third, theoretical fault characteristic frequencies dynamically calculated based on real-time speed and component geometric parameters, including electromagnetic force wave frequencies, gear meshing frequencies, and inner / outer ring / rolling element fault frequencies of rolling bearings.

[0089] The basic structural parameters in the original multi-source data are replaced with computationally generated structured feature vectors, resulting in processed multi-source data with richer semantics and higher diagnostic value. This process not only transforms static design parameters into dynamic diagnostic indicators that are strongly correlated with the current operating state, but also provides the structured branches of the diagnostic model with inputs that have clear physical meaning and fault orientation, significantly enhancing the model's ability to identify specific fault mechanisms and its interpretability.

[0090] Regarding S103, which inputs the multi-source data into the diagnostic model to obtain the fault diagnosis result of the electric drive system output by the diagnostic model, this application provides an optional embodiment:

[0091] The image modal data from the multi-source data is input into the image branch of the diagnostic model to obtain image features; the image branch includes a pre-trained 2D-CNN network (such as a variant of ResNet or VGG).

[0092] Image modal data (such as time-frequency color images, Campbell plots, and order slices of sound pressure level or vibration level) are input into the image branch. This branch is typically built based on a convolutional neural network (CNN), whose multi-layer convolutional structure can automatically learn and identify visual patterns in the image closely related to the fault: for example, vertical bright lines may represent resonance phenomena at specific frequencies, diagonal lines correspond to order components that change with rotational speed (such as gear meshing order or motor electromagnetic order), and local scattered or clustered bright areas may reflect random impacts or localized bearing damage. Through layer-by-layer abstraction and spatial feature extraction, the CNN encodes the original image into a high-dimensional, semantically rich image feature vector. This feature not only preserves the spatial distribution characteristics of the fault in the time-frequency domain but also implicitly integrates multi-dimensional information such as frequency, amplitude, and temporal evolution, providing a strong discriminative visual representation foundation for subsequent multimodal fusion.

[0093] The sequence modality data from the multi-source data is input into the sequence branch of the diagnostic model to obtain sequence features; the sequence branch includes a 1D-CNN network or a combined network; the combined network includes a 1D-CNN network and an LSTM network.

[0094] The sequence modal data (such as time-domain vibration / sound pressure waveforms, frequency-domain FFT spectra, envelope spectra, and other one-dimensional time-series or frequency-series signals at key measurement points) are input into the sequence branch. This branch is typically built based on architectures such as recurrent neural networks (RNNs), temporal convolutional networks (TCNs), or Transformers, and is specifically used to model local transient features in the signal (such as shocks, modulations, harmonic clusters, etc.) and dynamic dependencies in the time dimension (such as the hysteresis effect of vibration response to torque step changes, resonance evolution during speed ramp-up, etc.). Through deep learning, the sequence branch can adaptively extract time-frequency patterns and evolution laws highly correlated with the fault, thereby outputting a sequence feature vector rich in dynamic semantic information. This feature not only preserves the temporal continuity and causal structure of the original signal, but also significantly enhances the model's ability to perceive and recognize transient faults, variable operating condition responses, and coupled excitation mechanisms.

[0095] The structured modal data from the multi-source data are input into the structured branch of the diagnostic model to obtain structural features; the structured branch includes a fully connected network.

[0096] Structured modal data (including traditional signal statistical features such as RMS, peak value, kurtosis, and spectral peak, as well as key operating parameters such as average speed, output torque, and oil temperature) are input into a structured branch. This branch is typically composed of a multilayer perceptron (MLP) or other neural networks suitable for structured data, capable of deeply learning the complex nonlinear mapping relationship between these quantitative indicators and various NVH faults. Through end-to-end training, the structured branch not only captures the fault sensitivity of single parameters but also uncovers the interaction effects between multiple parameters (such as the gear meshing abnormality implied by a sudden increase in vibration within a specific torque range at high speeds), thereby outputting a structured feature vector with strong discriminative capabilities. This feature integrates prior engineering knowledge and data-driven principles, providing crucial information support with physical interpretability and fault orientation for subsequent multimodal fusion.

[0097] The image features, sequence features, and structural features are input into the fusion layer of the diagnostic model to obtain the fault diagnosis results of the electric drive system output by the diagnostic model.

[0098] The image feature vectors, sequence feature vectors, and structured feature vectors output by the three branches are effectively concatenated in the fusion layer to form a high-dimensional, information-complete, and multi-dimensional complementary joint feature vector. This joint feature not only integrates the spatial patterns of time-frequency images, the dynamic evolution laws of time-series signals, and the structured fault prior knowledge calculated based on the physical model, but also retains the sensitivity and discriminative power of each modality to different fault types. Based on this, the fusion layer, as the "final review" module of the diagnostic system, uses strategies such as fully connected networks, attention mechanisms, or lightweight classification heads to perform nonlinear interactions and adaptive weight adjustments on the joint features, comprehensively assessing the probability of various faults, outputting a normalized fault category probability distribution, and determining the highest probability item as the final diagnostic result. This design significantly enhances the model's comprehensive reasoning ability, robustness, and diagnostic accuracy in complex operating conditions and multi-fault coupled scenarios.

[0099] For example, the image branch tells the fusion layer, "I see a bright diagonal line in the image." The sequence branch tells the fusion layer, "I hear a roaring sound that accompanies the torque switching." The structured branch tells the fusion layer, "I deduce that the current speed and load are exactly what will excite the resonance of the transmission system, and the amplitude of the key frequency has exceeded the limit." The "opinions" of these three branches (i.e., their output feature vectors) are sent to the fusion layer for "final review" and a final diagnostic decision is made.

[0100] Regarding the training process of the diagnostic model, this application also provides an optional embodiment:

[0101] Obtain a training dataset; the training dataset includes historical multi-source data of the electric drive system and the corresponding fault diagnosis results (such as "radial electromagnetic force wave resonance" and "bearing electrical corrosion").

[0102] A diagnostic model to be trained is constructed based on deep learning.

[0103] The historical multi-source data in the training dataset is input into the diagnostic model to obtain the output of the diagnostic model to be trained.

[0104] The loss value is calculated based on the output of the diagnostic model to be trained and the fault diagnosis results corresponding to the historical multi-source data.

[0105] This application does not limit the method for calculating the loss value. For example, the loss value can be calculated using the cross-entropy loss function based on the output of the diagnostic model to be trained and the fault diagnosis results corresponding to the historical multi-source data.

[0106] With the goal of minimizing the loss value, the parameters of the diagnostic model to be trained are adjusted using the backpropagation algorithm. If the training termination condition is met, the training ends, and the diagnostic model is obtained.

[0107] If the training termination condition is not met, the process returns to the step of inputting historical multi-source data from the training dataset into the diagnostic model to obtain the output of the diagnostic model to be trained.

[0108] This application does not impose limitations on the termination conditions for model training, which can be flexibly set according to actual needs. For example, a fixed number of iterations (e.g., 800 training rounds), verification loss convergence, diagnostic accuracy plateau, or an early stopping mechanism can be used as termination criteria. The essence of model training lies in implicitly encoding the diagnostic experience accumulated by NVH experts over a long period of time, such as criteria like "the appearance of a stable high-energy band independent of rotational speed near 800 Hz usually indicates structural resonance" and "sideband modulation on the order line often indicates local gear damage," into the weight parameters of the neural network through a data-driven approach. This process not only realizes the digital accumulation and generalized reuse of expert knowledge, but also enables the model to automatically learn the deep mapping relationship between complex fault modes and multi-source signals from massive labeled samples, thereby possessing expert-level fault identification and reasoning capabilities without the need for explicit rules.

[0109] In practical applications, after determining the fault diagnosis results of the electric drive system, it is generally necessary to determine the repair plan for the electric drive system based on the fault diagnosis results. In view of this situation, this application provides an optional embodiment:

[0110] Based on the fault diagnosis results, a diagnostic report is obtained from the diagnostic report database to repair the electric drive system.

[0111] When the diagnostic model outputs fault cause A, it can automatically retrieve the corresponding diagnostic report (such as the identified fault mode and confidence level, the inferred root cause of the fault, and the recommended optimization scheme) from the diagnostic report library. Based on the diagnostic report, the electric drive system is repaired, and the final repair result (whether the problem was successfully resolved) and new data from the repair process are fed back to the diagnostic report library and the diagnostic model, which are then updated in real time. Through this closed loop, new fault modes can be continuously learned, diagnostic logic optimized, and continuous performance evolution and knowledge accumulation achieved.

[0112] This application does not limit the diagnostic report database, as shown in Tables 1 to 7.

[0113] Table 1. Diagnostic Report Database - 1

[0114]

[0115] Table 2. Diagnostic Report Database - 2

[0116]

[0117] Table 3. Diagnostic Report Database - 3

[0118]

[0119] Table 4. Diagnostic Report Database - 4

[0120]

[0121] Table 5. Diagnostic Report Database - 5

[0122]

[0123] Table 6. Diagnostic Report Database - 6

[0124]

[0125] Table 7. Diagnostic Report Database - 7

[0126]

[0127] As shown in Table 1, if the major fault category is electromagnetic excitation; the specific fault cause / phenomenon is torque pulsation (especially the 6th order and its harmonics) leading to transmission system vibration (low-speed rumbling); the diagnostic model identification mechanism is: image branch: identifying slant lines on the Campbell diagram that are strictly multiples of the motor speed (such as the 6th and 12th orders); structured branch: associating with low-speed, high-torque conditions to confirm that vibration energy is concentrated at low frequencies; then the immediate measures (symptomatic relief and verification) in the optimization scheme should include: 1. Injecting current harmonic compensation to actively cancel torque pulsation. 2. Checking and optimizing the stiffness and damping of the transmission system suspension. The fundamental measures (fundamental solution and design iteration) in the optimization scheme should include: 1. Adopting rotor skewed pole (V-type or segmented) technology to significantly reduce torque pulsation. 2. Optimizing the switching frequency and modulation strategy of the motor controller.

[0128] As shown in Table 2, if the major fault category is gears and transmissions; the specific fault cause / phenomenon is gear squealing (a mid-to-high frequency "hissing" sound, related to load and speed); the diagnostic model identification mechanism is: image branch: identifying bright diagonal lines on the Campbell's diagram corresponding to the gear order; sequence branch: detecting modulation sidebands at the gear meshing frequency; structured branch: confirming a positive correlation between fault amplitude and load height; then the immediate measures (symptomatic relief and verification) in the optimization scheme should include: 1. Within the limit tolerance range, fine-tune the gear meshing marks by selecting shims. 2. Replace with high-damping gear oil (compatibility with materials needs to be verified). The fundamental measures (fundamental solutions and design iterations) in the optimization scheme should include: 1. Perform micro-geometric modification (tooth profile / tooth direction modification) on the gear pair to compensate for load deformation and machining errors. 2. Adopt high-precision gear grinding technology to improve gear machining accuracy (e.g., reaching DIN 6 grade or above).

[0129] As shown in Table 3, if the major fault category is bearing-related; the specific fault cause / phenomenon is abnormal noise from the bearing cage (irregular "clattering" sound); the diagnostic model identification mechanism is sequence branching: detecting non-periodic, low-frequency impacts in the time-domain signal; and the LSTM network: capturing irregular temporal patterns; then the immediate measures (symptomatic relief and verification) in the optimization scheme should include: 1. Checking whether the grease is sufficient or deteriorated. The fundamental measures (fundamental solutions and design iterations) in the optimization scheme should include: 1. Selecting cages with better guiding methods and stronger materials (such as polyamide PA46+GF25, phenolic resin cages). 2. Optimizing the bearing housing design to ensure unobstructed lubrication.

[0130] As shown in Table 4, if the major fault category is structural integration and acoustic radiation; the specific fault cause / phenomenon is shell structure resonance (excited by electromagnetic or gear excitation, producing a "roaring" sound); the diagnostic model identification mechanism is image branch: identifying stable vertical bright bands in the color image, and structured branch: matching the resonance frequency with the pre-stored structural modal database to distinguish between electromagnetic resonance and structural resonance; then the immediate measures (symptomatic relief and verification) in the optimization scheme should include: 1. Applying constraint layer damping patches or installing dynamic vibration absorbers at the resonance panel location. The fundamental measures (fundamental solutions and design iterations) in the optimization scheme should include: 1. Replacing part of the aluminum shell with composite materials (such as long-fiber nylon), or performing topology optimization to improve stiffness and change modes while reducing weight. 2. Optimizing bolt connection strategies and flange stiffness to avoid local modes.

[0131] As shown in Table 5, if the major fault category is thermal management; the specific fault cause / phenomenon is abnormal noise caused by thermal deformation (inconsistent NVH performance in cold and hot conditions); the diagnostic model identification mechanism is time series analysis: establishing a temperature-time-NVH feature correlation model; LSTM network: identifying the dynamic process of NVH features "drifting" or changing with increasing temperature; image branch: identifying frequency components that change with temperature in a long-term color image; then the immediate measures (symptomatic relief and verification) in the optimization scheme should include: 1. Checking and adjusting gear meshing and bearing clearance under hot conditions. The fundamental measures (fundamental solutions and design iterations) in the optimization scheme should include: 1. Optimizing the cooling channel design (e.g., using spray + shaft hole cooling) to balance the internal temperature field of the electric drive system. 2. Selecting materials with low thermal expansion coefficients or implementing symmetrical structural designs for key components.

[0132] As shown in Table 6, if the fault category is control strategy type; the specific fault cause / phenomenon is mode switching jitter (such as switching from drive to recovery); the diagnostic model identification mechanism is sequence branch: LSTM network analysis of the entire time series of torque command and vibration response; image branch: identification of low-frequency vertical bright bands within the torque zero-crossing time window; structured branch: association with the working mode switching flag bit on the CAN bus; then the immediate measures (symptomatic relief and verification) in the optimization scheme should include: 1. Optimize the torque coordination control strategy of VCU / MCU to achieve smooth torque transition and accurate calibration. The fundamental measures (fundamental solutions and design iteration) in the optimization scheme should include: 1. Develop model-based predictive control algorithms to predict and smooth torque requests in advance.

[0133] As shown in Table 7, if the major fault category is assembly and process-related; the specific fault cause / phenomenon is foreign object intrusion (particulate matter during assembly); the diagnostic model identification mechanism is: image branch: identifying random bright spots (snowflake noise) across the entire frequency band in the color image; sequence branch: abnormally high time-domain signal impact index, detecting randomly distributed impact sequences; diagnostic logic: abnormal detection after excluding known periodic faults; then the immediate measures (symptomatic relief and verification) in the optimization plan should include: 1. Cleaning and oil changing the assembly. The fundamental measures (fundamental solutions and design iteration) in the optimization plan should include: 1. Upgrading cleanliness control standards (e.g., meeting ISO 16232 standards). 2. Adding an online vibration detection station at the end of the assembly line.

[0134] This application can automatically learn high-dimensional, nonlinear fault characterization directly from raw multi-source NVH data, without relying on manually defined features or rules. The model can not only accurately identify typical faults, but also effectively capture weak signals, implicit coupling patterns, and cross-operating condition evolution patterns that are difficult for the human eye to detect, significantly improving diagnostic accuracy and sensitivity. By unifying the modeling of three heterogeneous modes—images (such as acoustic vibration color images and Campbell diagrams), sequences (such as time-domain waveforms and spectral curves), and structured numerical values ​​(such as operating condition parameters and theoretical fault frequencies)—it achieves deep fusion of multi-dimensional information, simulates the cognitive logic of NVH experts' "comprehensive judgment," overcomes the one-sidedness of single data source analysis, and greatly enhances the robustness, comprehensiveness, and interpretability of diagnosis.

[0135] The entire diagnostic process is highly automated, from data input, feature extraction, fault identification to diagnostic report generation, all completed within minutes. This reduces the traditional reliance on days of manual spectrum analysis and experience-based reasoning by experts to just a few minutes, significantly improving engineering response efficiency. Simultaneously, the model possesses continuous learning and online iteration capabilities, enabling it to evolve dynamically by continuously incorporating new samples and feedback data. This allows it to adapt to new electric drive platforms, new fault modes, and complex operating conditions, gradually reducing long-term dependence on individual expert experience. Crucially, this application is not a "black box" AI application. The model architecture design and training process deeply embeds prior physical knowledge from the NVH field, such as order calculation principles, structural modal characteristics, and typical fault mechanisms like electromagnetic / gear / bearing faults, guiding the neural network's learning process with physical laws. For example, the model can automatically establish mapping relationships between visual patterns such as "vertical bright bands" (structural resonance), "diagonal lines" (order excitation), and "scattered points" (random impacts) in color images and specific fault root causes, surpassing traditional feature engineering based on fixed indicators.

[0136] In short, this application constructs a complete engineering closed loop of "data acquisition - intelligent diagnosis - optimization suggestions - effect verification". It not only outputs the fault category, but also recommends targeted symptomatic and fundamental solutions, and supports subsequent verification feedback. It truly realizes a value closed loop from algorithm innovation to engineering implementation, and provides intelligent support for the rapid location, efficient solution and positive design of NVH problems in electric drive systems.

[0137] Figure 2 A structural diagram of a fault diagnosis device for an electric drive system provided in an embodiment of this application is shown below. Figure 2 As shown, based on the fault diagnosis method for an electric drive system provided in the preceding embodiments, this application also provides a fault diagnosis device for an electric drive system, comprising:

[0138] The acquisition module is used to acquire multi-source data from the electric drive system.

[0139] The processing module is used to preprocess the multi-source data to obtain processed multi-source data.

[0140] A diagnostic module is used to input the multi-source data into a diagnostic model to obtain the fault diagnosis result of the electric drive system output by the diagnostic model. The diagnostic model includes an image branch, a sequence branch, a structured branch, and a fusion layer. The image branch is used to extract image features from the processed multi-source data. The sequence branch is used to extract sequence features from the processed multi-source data. The structured branch is used to extract structural features from the processed multi-source data. The fusion layer is used to obtain the fault diagnosis result of the electric drive system based on image features, sequence features, and structural parameters.

[0141] As an optional embodiment, the acquisition module specifically includes:

[0142] The basic data acquisition unit is used to acquire basic data of the motor, reducer and housing in the electric drive system; the basic data includes vibration acceleration data and voltage signals.

[0143] The control data acquisition unit is used to acquire data on motor speed, output torque, DC bus voltage, three-phase current, temperature, and operating mode flags in the electric drive system.

[0144] The speed phase data acquisition unit is used to acquire speed phase data emitted by the speed sensor or motor resolver in the electric drive system; the speed phase data includes instantaneous speed pulses and phase signals.

[0145] The acquisition unit is used to acquire the basic data of the motor, the basic data of the reducer, the basic data of the housing, the motor speed, the output torque, the DC bus voltage, the three-phase current, the temperature data, the operating mode flag, and the speed phase data as the multi-source data.

[0146] As an optional embodiment, the processing module is specifically used for:

[0147] The image modal data in the multi-source data is standardized to obtain standardized multi-source data; the sequence modal data in the multi-source data is denoised and standardized to obtain denoised multi-source data; the structure parameters in the multi-source data are calculated to obtain structured modal data; the structure parameters in the multi-source data are replaced with the structured modal data to obtain processed multi-source data.

[0148] As an optional embodiment, the diagnostic module specifically includes:

[0149] Image modal data from the multi-source data are input into the image branch of the diagnostic model to obtain image features; the image branch includes a pre-trained 2D-CNN network. Sequence modal data from the multi-source data are input into the sequence branch of the diagnostic model to obtain sequence features; the sequence branch includes a 1D-CNN network or a combined network; the combined network includes a 1D-CNN network and an LSTM network. Structured modal data from the multi-source data are input into the structured branch of the diagnostic model to obtain structure features; the structured branch includes a fully connected network. The image features, sequence features, and structure features are input into the fusion layer of the diagnostic model to obtain the fault diagnosis result of the electric drive system output by the diagnostic model.

[0150] As an optional embodiment, the apparatus further includes:

[0151] A training module is used to acquire a training dataset, which includes historical multi-source data of the electric drive system and corresponding fault diagnosis results. A diagnostic model to be trained is constructed based on deep learning. The historical multi-source data in the training dataset is input into the diagnostic model to obtain its output. A loss value is calculated based on the output of the diagnostic model and the corresponding fault diagnosis results. The parameters of the diagnostic model to be trained are adjusted using a backpropagation algorithm with the goal of minimizing the loss value. If the training termination condition is met, the training ends, and the diagnostic model is obtained. If the training termination condition is not met, the process returns to the step of inputting the historical multi-source data in the training dataset into the diagnostic model to obtain its output.

[0152] As an optional embodiment, the apparatus further includes:

[0153] The report generation module is used to query the diagnostic report library based on the fault diagnosis results to obtain a diagnostic report for the electric drive system; the diagnostic report is used to repair the electric drive system.

[0154] This application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a fault diagnosis method for an electric drive system.

[0155] This application provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a fault diagnosis method for an electric drive system.

[0156] This application provides a computer program product, including a computer program that, when executed by a processor, implements a fault diagnosis method for an electric drive system.

[0157] It should be noted that the various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for the device and equipment embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method embodiments. The device and equipment embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components indicated as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the solution in this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0158] The above description is merely one specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A fault diagnosis method for an electric drive system, characterized in that, The method includes: Acquire multi-source data from the electric drive system; The multi-source data is preprocessed to obtain processed multi-source data; The multi-source data is input into a diagnostic model to obtain the fault diagnosis result of the electric drive system output by the diagnostic model. The diagnostic model includes an image branch, a sequence branch, a structured branch, and a fusion layer. The image branch is used to extract image features from the processed multi-source data. The sequence branch is used to extract sequence features from the processed multi-source data. The structured branch is used to extract structural features from the processed multi-source data. The fusion layer is used to obtain the fault diagnosis result of the electric drive system based on image features, sequence features, and structural parameters.

2. The fault diagnosis method for an electric drive system according to claim 1, characterized in that, The acquisition of multi-source data from the electric drive system specifically includes: Collect basic data of the motor, reducer, and housing in the electric drive system; the basic data includes vibration acceleration data and voltage signals. Collect data on motor speed, output torque, DC bus voltage, three-phase current, temperature, and operating mode flags in the electric drive system; The system acquires speed phase data from a speed sensor or motor resolver in an electric drive system; the speed phase data includes instantaneous speed pulses and phase signals. The basic data of the motor, the basic data of the reducer, the basic data of the housing, the motor speed, the output torque, the DC bus voltage, the three-phase current, the temperature data, the operating mode flag, and the speed phase data are used as the multi-source data.

3. The fault diagnosis method for an electric drive system according to claim 1, characterized in that, The preprocessing of the multi-source data to obtain processed multi-source data specifically includes: The image modal data in the multi-source data is standardized to obtain standardized multi-source data; The sequence modal data in the multi-source data are denoised and standardized to obtain denoised multi-source data. Structured modal data is obtained by calculating the structural parameters in the multi-source data. The structure parameters in the multi-source data are replaced with the structured modal data to obtain the processed multi-source data.

4. The fault diagnosis method for an electric drive system according to claim 1, characterized in that, The step of inputting the multi-source data into the diagnostic model to obtain the fault diagnosis result of the electric drive system output by the diagnostic model specifically includes: The image modal data from the multi-source data is input into the image branch of the diagnostic model to obtain image features; the image branch includes a pre-trained 2D-CNN network. The sequence modality data from the multi-source data is input into the sequence branch of the diagnostic model to obtain sequence features; the sequence branch includes a 1D-CNN network or a combined network; the combined network includes a 1D-CNN network and an LSTM network. The structured modal data from the multi-source data are input into the structured branch of the diagnostic model to obtain structural features; the structured branch includes a fully connected network. The image features, sequence features, and structural features are input into the fusion layer of the diagnostic model to obtain the fault diagnosis results of the electric drive system output by the diagnostic model.

5. The fault diagnosis method for an electric drive system according to claim 1, characterized in that, The training process of the diagnostic model specifically includes: Obtain a training dataset; the training dataset includes historical multi-source data of the electric drive system and the corresponding fault diagnosis results of the historical multi-source data; Constructing a diagnostic model to be trained based on deep learning; The historical multi-source data in the training dataset is input into the diagnostic model to obtain the output of the diagnostic model to be trained; The loss value is calculated based on the output of the diagnostic model to be trained and the fault diagnosis results corresponding to the historical multi-source data. With the goal of minimizing the loss value, the parameters of the diagnostic model to be trained are adjusted using the backpropagation algorithm. If the training termination condition is met, the training ends, and the diagnostic model is obtained. If the training termination condition is not met, the process returns to the step of inputting historical multi-source data from the training dataset into the diagnostic model to obtain the output of the diagnostic model to be trained.

6. The fault diagnosis method for an electric drive system according to claim 1, characterized in that, The method further includes: Based on the fault diagnosis results, a diagnostic report is obtained from the diagnostic report database to repair the electric drive system.

7. A fault diagnosis device for an electric drive system, characterized in that, The fault diagnosis device for the electric drive system includes: The acquisition module is used to acquire multi-source data from the electric drive system; The processing module is used to preprocess the multi-source data to obtain processed multi-source data; A diagnostic module is used to input the multi-source data into a diagnostic model to obtain the fault diagnosis result of the electric drive system output by the diagnostic model. The diagnostic model includes an image branch, a sequence branch, a structured branch, and a fusion layer. The image branch is used to extract image features from the processed multi-source data. The sequence branch is used to extract sequence features from the processed multi-source data. The structured branch is used to extract structural features from the processed multi-source data. The fusion layer is used to obtain the fault diagnosis result of the electric drive system based on image features, sequence features, and structural parameters.

8. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the fault diagnosis method for the electric drive system according to any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the fault diagnosis method for the electric drive system as described in any one of claims 1-6.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the fault diagnosis method for the electric drive system as described in any one of claims 1-6.

Images