Fault data driving analysis method and system for new energy commercial vehicle

Abstract

The invention relates to the technical field of vehicle intelligent management, in particular to a fault data driving analysis method and system for a new energy commercial vehicle, and the method comprises the steps: collecting vehicle operation data, cloud environment data and external working condition data to form multi-source time series data; preprocessing the data and generating a sliding time window data set; basic features, time-frequency domain features and environment coupling features are extracted from the data set; splicing the features into a sequence, inputting the sequence into a pre-trained anomaly detection model, outputting an anomaly score, and triggering early warning according to a preset rule; and analyzing the contribution degree of each feature to the exception score, and performing rule matching with a single-vehicle fault mode library in which a fault evolution path is stored to obtain a predicted exception type and generate a diagnosis report. According to the invention, accurate detection and real-time early warning of low-frequency gradual change faults under complex working conditions can be realized, an interpretable diagnosis report is generated, and the maintenance decision-making efficiency is improved.

Description

Technical Field

[0001] This application relates to the field of vehicle intelligent management technology, specifically to a fault data-driven analysis method and system for new energy commercial vehicles. Background Technology

[0002] With the large-scale application of new energy commercial vehicles, the operational reliability of their core components, such as power batteries, drive motors, and electronic control systems, is directly related to vehicle operational safety and total lifecycle costs. Therefore, achieving early and accurate warnings of potential vehicle faults, especially low-frequency gradual faults under long-distance heavy-load conditions, has become a key requirement for improving fleet operation and maintenance efficiency and safety levels.

[0003] Currently, common fault diagnosis methods for new energy commercial vehicles mainly rely on monitoring and analyzing vehicle bus data. For example, voltage and current data are obtained through the battery management system, or speed and torque data are obtained through the motor controller, and fixed thresholds are set for alarms. To further improve diagnostic capabilities, some solutions have introduced signal processing techniques, such as performing spectral analysis on vibration signals to identify mechanical component faults, or using historical data to train machine learning models to identify abnormal patterns.

[0004] However, existing fault diagnosis methods for new energy commercial vehicles still have significant limitations: the data sources of existing methods are usually limited to the vehicle's own sensors, failing to incorporate real-time changing cloud environmental data and external road condition data into the analysis framework, resulting in poor environmental adaptability of the diagnostic models and insufficient generalization ability under complex conditions across regions and seasons; the complexity of the models deployed in most solutions contradicts the limited computing power of edge computing units, making it difficult for complex models to meet the engineering requirements of millisecond-level real-time inference, while oversimplified models cannot effectively capture low-frequency, gradual fault characteristics; existing methods usually only output an abstract anomaly score or alarm signal, lacking interpretable output on "why the alarm occurred" and "what the root cause of the fault might be," making it difficult for maintenance personnel to quickly locate the problem, increasing decision-making costs and downtime. Summary of the Invention

[0005] To address the technical problems of existing fault diagnosis methods for new energy commercial vehicles, such as poor environmental adaptability due to single data sources, difficulty in balancing real-time performance and accuracy due to the contradiction between model complexity and edge computing power, and low maintenance decision efficiency due to the lack of interpretability of diagnostic results, this application provides a fault data-driven analysis method and system for new energy commercial vehicles. This method can accurately detect and provide real-time early warning of low-frequency gradual faults under complex operating conditions, and generate interpretable diagnostic reports, thereby improving the efficiency of maintenance decision-making.

[0006] Firstly, this application provides a fault data-driven analysis method for new energy commercial vehicles, comprising the following steps: S1. Collect multi-source time series data, including vehicle operating data, cloud environment data, and external operating condition data; The vehicle's own operating data includes battery management system data, motor controller data, and vibration data. The battery management system data includes at least the charging and discharging current, and the motor controller data includes at least the motor speed. S2. Preprocess the multi-source time series data to generate a sliding time window dataset divided according to a preset time length; S3. Extract data features based on the sliding time window dataset, including basic features, time-frequency domain features, and environmental coupling features; Among them, the basic features include statistical features and frequency domain features extracted from the vehicle's own operating data; S4. Concatenate the data features into a feature sequence and input it into the pre-trained anomaly detection model. The anomaly detection model encodes the feature sequence and evaluates anomalies, and outputs anomaly scores. The anomaly scores range from [0,1]. The anomaly detection model is built based on Transformer Encoder. S5. Based on the abnormal score, trigger an alert according to the preset alert rules; S6. After triggering the warning, analyze the contribution of each feature item in the feature sequence to the abnormal score, match the contribution with the preset single-vehicle fault mode library according to rules to obtain the predicted abnormal type, and generate a diagnostic report based on the predicted abnormal type. The single-vehicle fault mode library stores historical labeled data, abnormal sample clustering results, abnormal type labels determined based on the abnormal sample clustering results, feature thresholds corresponding to abnormal types, and fault evolution paths composed of abnormal type sequences. The feature thresholds are the thresholds of feature items in the feature sequence, the thresholds of the contribution of feature items to the abnormal score, or a combination thereof.

[0007] It should be further noted that in step S1, the battery management system data includes individual cell voltage, battery module temperature, charging and discharging current, and AC internal resistance. Motor controller data includes motor speed and motor torque; The vibration signal is vibration acceleration sensor data that reflects the mechanical state of the motor or transmission system, collected at a sampling rate of not less than 1 kHz. The cloud-based environmental data is weather forecast data for a future preset time period, including at least the air temperature, humidity, and precipitation intensity corresponding to the vehicle's location; External operating condition data includes: Road slope angle and vehicle turning radius obtained through lidar or inertial measurement unit; Road surface slippage information is obtained by fusing wheel speed sensor signals and capacitive ground humidity sensor signals.

[0008] It should be further noted that the vehicle's own operating data is collected through the vehicle bus and onboard sensors; Cloud-based environmental data is obtained by calling the meteorological service application interface and is then linked to vehicle location information. External operating condition data is collected through vehicle-mounted lidar or inertial measurement units and road surface sensing devices.

[0009] It should be further noted that the acquisition accuracy of the battery cell voltage is ±10mV, the acquisition accuracy of the battery module temperature is ±1℃, and the acquisition accuracy of the charging and discharging current is ±0.5% of full scale. The accuracy of motor speed acquisition is ±0.1 rpm, and the accuracy of motor torque acquisition is ±0.5% of full scale.

[0010] It should be further noted that step S2, the preprocessing of the multi-source time series data, includes: Data normalization is performed on multi-source time series data, including timestamp alignment, missing value completion, and outlier removal; After data normalization, the multi-source time series data is divided into sliding window segments according to a preset time length to generate a sliding time window dataset.

[0011] It should be further explained that the timestamp alignment is as follows: Using UTC timestamps, vehicle operating data, cloud environment data, and external operating condition data are synchronized and aligned to the millisecond level; Missing value imputation specifically involves using the Kalman filter algorithm to imput missing data points in each data sequence of multi-source time series data. Outlier removal specifically involves calculating the mean and standard deviation for each data series in the multi-source time series data, and identifying and removing data points that deviate from the mean by more than three times the standard deviation.

[0012] It should be further explained that the sliding window segmentation of the normalized multi-source time series data according to the preset time length is specifically as follows: The window length is set to 10s and the sliding step is 100ms. Multi-source time series data are continuously extracted along the time axis to generate a sliding time window dataset.

[0013] It should be further explained that in step S3, the basic features are statistical features and transformation features extracted from the vehicle's own operating data. Among them, statistical characteristics include mean, variance, and extreme values; The transformation features include the main frequency components in the range of 1Hz to 1000Hz extracted after performing fast Fourier transforms on the battery current signal and vibration signal, respectively.

[0014] It should be further noted that the method for extracting time-frequency domain features in step S3 is as follows: Morlet wavelet basis functions were used to perform three-level continuous wavelet transforms on the battery current signal and vibration signal, respectively. The three levels of transform correspond to the frequency band scales of 1-10Hz, 10-100Hz and 100-1000Hz. Energy entropy and kurtosis values ​​are extracted from wavelet transform results on a frequency band scale of 1-10 Hz. Extract the amplitude or energy concentration of the impulse component of the signal from the wavelet transform results on the 10-100Hz frequency band scale; Calculate the proportion of noise energy in a specified high-frequency sub-band from the wavelet transform results of the 100-1000Hz frequency band. Envelope demodulation analysis was performed on the vibration signal to extract fault characteristic frequencies. The calculation formula is as follows:

[0015] in, This indicates the number of rolling elements in the bearing corresponding to the vibration signal; The diameter of the rolling element of the bearing corresponding to the vibration signal; The pitch circle diameter of the bearing corresponding to the vibration signal; This refers to the motor speed; The time-frequency domain features include energy entropy, kurtosis, impulse component amplitude or energy concentration, and noise energy proportion extracted from the wavelet transform results at each scale, as well as fault characteristic frequencies extracted from envelope demodulation analysis. .

[0016] It should be further noted that the center frequency of the Morlet wavelet basis function is 1 Hz and the bandwidth is 0.5 Hz.

[0017] It should be further explained that the method for extracting environmental coupling features is as follows: A pre-defined multivariate correlation function is used to calculate parameters in vehicle operating data, cloud environment data, and external operating condition data to obtain environmental coupling characteristics.

[0018] It should be further noted that the environmental coupling characteristics include: Battery internal resistance correction value The corresponding multivariate correlation function is:

[0019] Motor torque compensation value The corresponding multivariate correlation function is:

[0020] in, This indicates the AC internal resistance in the vehicle's own operating data; This indicates the preset temperature influence coefficient; This indicates the air temperature corresponding to the vehicle's location in the cloud-based environmental data; Indicates the preset reference temperature; Motor torque in the vehicle's own operating data; This represents the road slope angle in the external working condition data.

[0021] It should be further noted that the temperature influence coefficient The data was obtained by performing a multiple linear regression fit on historical operating data, with units of ℃. -1 The value ranges from 0.005 to 0.02.

[0022] It should be further noted that in step S4, the anomaly detection model includes an input layer, four encoder layers, and an output layer. Each encoder layer sequentially executes a multi-head self-attention mechanism and a feedforward neural network operation. The multi-head self-attention mechanism uses a total of 4 independent attention heads; The hidden layer dimension of the feedforward neural network is 128, and the dimension of its internal feedforward network is 512. The model employs a sinusoidal positional encoding method to add temporal information to each position in the input feature sequence; The data processing procedure for the anomaly detection model is as follows: After the feature sequence is input through the input layer, the four encoder layers encode the input feature sequence in sequence. Then, the output layer maps the hidden state output by the last encoder layer to a scalar value with a range of [0,1], which is used as the anomaly score.

[0023] It should be further explained that the four attention heads of the multi-head attention mechanism include two general attention heads and two fault-sensitive attention heads that are more sensitive to low-frequency gradual fault patterns. The fault-sensitive attention heads learn weights during the training process to focus on the feature dimensions of the input feature sequence that are related to low-frequency gradual fault patterns.

[0024] It should be further noted that the anomaly detection model undergoes INT8 integer quantization, and its inference computing power requirement on the edge computing unit does not exceed 5 TOPS.

[0025] It should be further explained that the training method for the anomaly detection model is as follows: Acquire multi-source time series data collected within a historical time period, and after processing in the same way as steps S2 and S3, obtain multiple sets of data features to construct a training sample set; Based on the corresponding actual working conditions, each set of data features in the training sample set is assigned a working condition status label. The working condition status label corresponding to normal working conditions is 0, and the working condition status label corresponding to faulty working conditions is 1. Using the feature sequences in the training sample set as input and the corresponding working condition labels as supervision signals, the anomaly detection model is trained in a supervised manner using the binary cross-entropy loss function, so that the anomaly score output by the model approximates the working condition label until the model converges.

[0026] It should be further explained that the warning rules in step S5 are as follows: When the abnormal score exceeds the preset warning threshold, a level 2 warning is triggered; A Level 1 warning is triggered when the abnormal score exceeds the first threshold for three consecutive detection periods.

[0027] It should be further noted that when a Level 2 warning is triggered, a message will be sent to the user suggesting that they check within 24 hours; When a Level 1 warning is triggered, a message will be sent to the user suggesting that they stop the car immediately for inspection.

[0028] It should be further noted that the warning threshold is 0.8.

[0029] It should be further noted that, in step S6, the contribution of each feature item in the parsed feature sequence to the anomaly score specifically includes: Deploy a lightweight SHAP interpreter; The TreeSHAP algorithm in the lightweight SHAP interpreter is used to calculate the feature sequence of the input anomaly detection model. The Shapley value of each feature item in the feature sequence is quantified as the contribution of that feature item to the anomaly score.

[0030] It should be further explained that step S6, which involves matching the contribution score with the single-vehicle fault mode library according to rules, specifically includes: Read the feature thresholds corresponding to each anomaly type stored in the single-vehicle fault mode library; The feature values ​​of each feature item in the feature sequence and the contribution of each feature item to the anomaly score are logically compared with the feature thresholds. If a certain feature threshold is met, the current warning is determined to match the anomaly type corresponding to that anomaly threshold.

[0031] It should be further explained that the rule matching logic includes: if the contribution of the feature item "cell voltage deviation" to the anomaly score is greater than 0.5, and the value of the feature item "cell internal resistance correction value" is greater than the preset threshold, then the matched anomaly type is battery consistency deterioration.

[0032] It should be further noted that the single-vehicle fault mode library also stores fault repair knowledge corresponding to the fault types; Troubleshooting knowledge includes recommended inspection parts, repair procedures, or a list of replacement parts.

[0033] It should be further noted that the diagnostic report in step S6 includes: Predicted anomaly types; The contribution and value of each feature term; Trend charts of each feature item within a specified time period before and after the warning; Fault repair knowledge retrieved from the fault mode library that corresponds to the predicted anomaly type.

[0034] It should be further noted that this also includes the steps for updating the single-vehicle fault mode library, specifically: When the warning meets the preset update conditions, a data segment within a specified time period before and after the warning is triggered is generated and uploaded to the cloud. The data segment includes the original signal, data characteristics, anomaly score and timestamp of multi-source time series data. Based on data fragments, a pre-built library of single-vehicle fault modes in the cloud is updated using a semi-supervised learning method.

[0035] It should be further noted that the preset update condition is: triggering a Level 1 warning; When a Level 1 warning is triggered, data segments within a specified time period before and after the trigger time are generated and uploaded to the cloud.

[0036] It should be further noted that the specified time period before and after the warning is from 100 seconds before the Level 1 warning to 200 seconds after the Level 1 warning.

[0037] It should be further explained that in step S5, updating the pre-built single-vehicle fault mode library in the cloud based on the data fragment using a semi-supervised learning method specifically includes: Data corresponding to normal operating conditions in historical labeled data are used as labeled normal samples; For the data corresponding to the time when the warning was triggered in the data segment, the DBSCAN clustering algorithm was used for cluster analysis to form several unlabeled abnormal sample clusters; By combining labeled normal samples with aberrant sample clusters, new aberration types can be identified and defined, or the sample feature distribution of existing aberration types can be expanded, through label propagation or inductive learning. The newly identified or expanded anomaly types, their corresponding feature thresholds, and the correlation between fault evolution paths are updated to the single-vehicle fault mode library.

[0038] Secondly, this application provides a fault data-driven analysis system for new energy commercial vehicles, used to implement the above-mentioned fault data-driven analysis method, including: The data acquisition module is used to collect multi-source time-series data, including vehicle operating data, cloud environment data, and external operating condition data. The data preprocessing module is used to preprocess multi-source time series data and generate a sliding time window dataset divided according to a preset time length. The feature extraction module is used to extract data features based on the sliding time window dataset, including basic features, time-frequency domain features, and environmental coupling features; The anomaly score generation module concatenates data features into a feature sequence, which is then input into a pre-trained anomaly detection model. The anomaly detection model encodes the feature sequence and evaluates for anomalies, outputting anomaly scores. The early warning triggering module is used to trigger early warnings based on the abnormal score and according to preset early warning rules; The diagnostic report generation module is used to analyze the contribution of each feature item in the feature sequence to the abnormal score after the warning is triggered, match the contribution with the preset single-vehicle fault mode library according to rules to obtain the predicted abnormal type, and generate a diagnostic report based on the predicted abnormal type.

[0039] Thirdly, this application provides an electronic 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 steps of the above-described fault data-driven analysis method.

[0040] Fourthly, this application provides a storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described fault data-driven analysis method.

[0041] As can be seen from the above technical solutions, this application has the following advantages: 1. This application systematically solves the comprehensive problems of isolated data, poor model adaptability, and uninterpretable results in the prior art by sequentially performing multi-source time series data acquisition, preprocessing, multi-dimensional feature extraction, anomaly detection and hierarchical early warning based on Transformer Encoder, and finally generating an interpretable diagnostic report. It realizes accurate, real-time and actionable early warning and diagnosis of faults in new energy commercial vehicles, especially low-frequency gradual faults under complex working conditions.

[0042] 2. This application solves the problems of single data dimension and limited analysis perspective in the existing technology by synchronously collecting and integrating vehicle's own operating data, cloud environment data and external working condition data. It provides a more comprehensive information foundation for fault analysis that is closer to the actual operating scenario, and significantly improves the richness of fault feature extraction and the reliability of diagnostic conclusions.

[0043] 3. This application solves the problems of asynchronous, messy, and unsuitable raw data for direct model analysis by performing timestamp alignment, missing value completion, outlier removal, and generating a sliding time window dataset from multi-source time series data. This ensures the quality and consistency of the input data and provides the necessary conditions for subsequent stable and accurate model inference.

[0044] 4. This application solves the problems of traditional methods in constructing features in a one-sided manner and failing to effectively separate noise and fault information by systematically extracting basic features, time-frequency domain features and environmental coupling features from sliding time window datasets. In particular, it enhances the ability to capture low-frequency gradual fault modes through time-frequency domain analysis and eliminates external interference through environmental coupling correction, thereby improving the feature representation ability.

[0045] 5. This application solves the problem that simple thresholding methods or traditional machine learning models are insufficient in capturing complex temporal dependencies and deep anomaly patterns by inputting the concatenated feature sequence into a pre-trained anomaly detection model built on Transformer Encoder. It utilizes its powerful sequence encoding capability to achieve efficient evaluation of complex correlations between multidimensional features, thereby outputting more accurate anomaly scores.

[0046] 6. This application solves the problem that a single alarm threshold cannot distinguish the urgency of a fault by setting a hierarchical early warning rule based on anomaly scores, and realizes a tiered management from "potential risk warning" to "immediate handling requirements", making operation and maintenance response more targeted and efficient.

[0047] 7. This application generates a diagnostic report by analyzing the contribution of feature terms to the anomaly score and matching it with a single-vehicle fault mode library containing fault evolution paths. This solves the problems of ambiguous output results and poor maintenance guidance in the prior art, transforming the abstract anomaly score into a specific fault type prediction, which greatly improves the operability and efficiency of fault handling. Attached Figure Description

[0048] To more clearly illustrate the technical solution of this application, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying 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.

[0049] Figure 1 This is a flowchart of a fault data-driven analysis method for new energy commercial vehicles in one embodiment of this application.

[0050] Figure 2 This is a schematic block diagram of a fault data-driven analysis system for new energy commercial vehicles in one embodiment of this application.

[0051] Figure 3 This is a schematic diagram of the hardware structure of an electronic device in one embodiment of this application. Detailed Implementation

[0052] To make the purpose, features, and advantages of this application more apparent and understandable, specific embodiments and accompanying drawings will be used to clearly and completely describe the technical solution protected by this application. Obviously, the embodiments described below are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0053] The fault data-driven analysis method of this application will be described in detail below. Specific details such as particular system structures and technologies are presented for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application can also be implemented in other embodiments without these specific details.

[0054] In the fault data-driven analysis method involved in this application, the term "comprising" indicates the presence of the described feature, whole, step, operation, element, and / or component, but does not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components, and / or sets thereof. The terms "comprising," "including," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0055] To facilitate a clear description of the technical solutions of this application, the terms "first" and "second" are used to distinguish identical or similar items with essentially the same function and effect. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and that the terms "first" and "second" do not necessarily imply that they are different.

[0056] The terms "one embodiment" or "some embodiments" used in this application mean that one or more embodiments of this application include the specific features, structures, or characteristics described in that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this application do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized.

[0057] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings.

[0058] The fault data-driven analysis method provided in this application embodiment is executed by a computer device. Correspondingly, the fault data-driven analysis system for new energy commercial vehicles runs in a computer device.

[0059] Figure 1 This is a flowchart of a fault data-driven analysis method for new energy commercial vehicles according to an embodiment of this application. Figure 1 The executing entity can be a fault data-driven analysis system. Depending on different requirements, the order of the steps in this flowchart can be changed, and some can be omitted.

[0060] like Figure 1 As shown, the fault data-driven analysis method for this new energy commercial vehicle includes: Step S1: Collect multi-source time series data, including vehicle operating data, cloud environment data, and external operating condition data; The vehicle's own operating data includes battery management system data, motor controller data, and vibration data. The battery management system data includes at least the charging and discharging current, and the motor controller data includes at least the motor speed.

[0061] By synchronously collecting multi-source time-series data, including vehicle operating data, cloud environment data, and external operating condition data, fault analysis can integrate internal mechanical and electrical signals affecting vehicle status, external real-time environmental factors, and specific driving condition information. This provides a complete data foundation covering both internal and external factors for subsequently constructing a comprehensive, high-dimensional feature space to accurately characterize complex fault modes, especially low-frequency gradual faults affected by the coupling of environment and operating conditions.

[0062] In some specific embodiments, in step S1, the battery management system data includes individual cell voltage, battery module temperature, charging and discharging current, and AC internal resistance. Motor controller data includes motor speed and motor torque; The vibration signal is vibration acceleration sensor data that reflects the mechanical state of the motor or transmission system, collected at a sampling rate of not less than 1 kHz. The cloud-based environmental data is weather forecast data for a future preset time period, including at least the air temperature, humidity, and precipitation intensity corresponding to the vehicle's location; External operating condition data includes: Road slope angle and vehicle turning radius obtained through lidar or inertial measurement unit; Road surface slippage information is obtained by fusing wheel speed sensor signals and capacitive ground humidity sensor signals.

[0063] By specifically defining the battery management system data (including cell voltage, module temperature, current, and AC internal resistance), motor controller data (including speed and torque), vibration signals (acceleration data of at least 1kHz), and cloud-based environmental data (future weather forecasts), as well as external operating condition data (including slope, turning radius, and road surface slippage), the data collection content becomes more specific and refined, ensuring that subsequent feature extraction can cover key dimensions such as electrochemical state, thermal state, mechanical state, and environmental pre-adaptation.

[0064] In some specific embodiments, the vehicle's own operating data is collected via the vehicle bus and onboard sensors; Cloud-based environmental data is obtained by calling the meteorological service application interface and is then linked to vehicle location information. External operating condition data is collected through vehicle-mounted lidar or inertial measurement units and road surface sensing devices.

[0065] By specifically defining the vehicle's own operating data as collected through the vehicle bus and onboard sensors, cloud-based environmental data as obtained by calling the meteorological API and bound to the positioning, and external operating condition data as collected through onboard radar / IMU and road surface sensing devices, the reliable acquisition path and fusion method of multi-source data are clarified, ensuring the real-time performance, accuracy, and spatiotemporal consistency of the data, and providing technical support for building a reliable analytical foundation.

[0066] In some specific embodiments, the acquisition accuracy of the battery cell voltage is ±10mV, the acquisition accuracy of the battery module temperature is ±1℃, and the acquisition accuracy of the charge / discharge current is ±0.5% of full scale. The accuracy of motor speed acquisition is ±0.1 rpm, and the accuracy of motor torque acquisition is ±0.5% of full scale.

[0067] By specifically limiting the acquisition accuracy of battery cell voltage, module temperature, charging and discharging current, motor speed, and motor torque, a clear accuracy threshold is set for the measurement of key physical quantities, ensuring the quality of the raw data, reducing feature distortion and model misjudgment caused by measurement errors, and thus improving the credibility of the starting point data of the entire analysis chain.

[0068] Step S2: Preprocess the multi-source time series data to generate a sliding time window dataset divided according to a preset time length.

[0069] By preprocessing multi-source time series data and dividing it into sliding time window datasets according to preset time lengths, spatiotemporal alignment and normalization of data from different sources and with different sampling rates are achieved, transforming continuous running data into standardized time segments and providing a standard data structure that can be directly processed.

[0070] In some specific embodiments, preprocessing of multi-source time series data includes: Data normalization is performed on multi-source time series data, including timestamp alignment, missing value completion, and outlier removal; After data normalization, the multi-source time series data is divided into sliding window segments according to a preset time length to generate a sliding time window dataset.

[0071] By specifying the preprocessing of multi-source time series data into two sub-steps—data regularization (timestamp alignment, missing value completion, and outlier removal) and sliding window segmentation—it provides a clear and operable data cleaning and organization method. This method can effectively handle common problems in actual data collection, such as time asynchrony, missing data, and noise, and generate clean and regular time window data.

[0072] In some specific embodiments, timestamp alignment is specifically as follows: Using UTC timestamps, vehicle operating data, cloud environment data, and external operating condition data are synchronized and aligned to the millisecond level; Missing value imputation specifically involves using the Kalman filter algorithm to imput missing data points in each data sequence of multi-source time series data. Outlier removal specifically involves calculating the mean and standard deviation for each data series in the multi-source time series data, and identifying and removing data points that deviate from the mean by more than three times the standard deviation.

[0073] By specifying data normalization as millisecond-level alignment using UTC timestamps, filling in missing values ​​using Kalman filtering, and removing outliers using a three-standard-deviation criterion, this paper provides an efficient processing algorithm tailored to the characteristics of time-series data. It can achieve accurate spatiotemporal synchronization, smooth interpolation, and noise filtering while maintaining the dynamic characteristics of the signal, significantly improving the data quality used for subsequent analysis.

[0074] In some specific embodiments, the sliding window segmentation of the normalized multi-source time series data according to a preset time length is specifically as follows: The window length is set to 10s and the sliding step is 100ms. Multi-source time series data are continuously extracted along the time axis to generate a sliding time window dataset.

[0075] By specifically setting the sliding window length to 10 seconds and the sliding step size to 100 milliseconds, the specific spatiotemporal parameters for continuously extracting and generating datasets are defined. This allows the analysis to cover a sufficiently long period of time to capture low-frequency fault characteristics, while also tracking rapid changes in state with a high temporal resolution, thus balancing analytical granularity and computational efficiency.

[0076] Step S3: Extract data features based on the sliding time window dataset, including basic features, time-frequency domain features, and environmental coupling features; The basic features include statistical features and frequency domain features extracted from the vehicle's own operating data.

[0077] By extracting multi-dimensional data features, including basic features, time-frequency domain features, and environmental coupling features, from the sliding time window dataset, information can be deeply mined from multiple perspectives, such as the statistical regularity of the original signal, the frequency domain characteristics that change over time, and the correlation between environmental conditions and component states. This allows for the construction of feature representations with high discriminative power for various fault modes (especially mechanical faults with complex time-frequency characteristics).

[0078] In some specific embodiments, in step S3, the basic features are statistical features and transformation features extracted from the vehicle's own operating data; Among them, statistical characteristics include mean, variance, and extreme values; The transformation features include the main frequency components in the range of 1Hz to 1000Hz extracted after performing fast Fourier transforms on the battery current signal and vibration signal, respectively.

[0079] By defining the basic features specifically as statistical features (mean, variance, extreme values) and transformation features (1Hz to 1000Hz main frequency components extracted from the vehicle's own operating data after performing FFT on battery current and vibration signals), a standardized method is provided to describe signal characteristics from two basic perspectives: time domain and frequency domain. This method can effectively characterize the steady-state level and basic periodic characteristics of component operation.

[0080] In some specific embodiments, the method for extracting time-frequency domain features in step S3 is as follows: Morlet wavelet basis functions were used to perform three-level continuous wavelet transforms on the battery current signal and vibration signal, respectively. The three levels of transform correspond to the frequency band scales of 1-10Hz, 10-100Hz and 100-1000Hz. Energy entropy and kurtosis values ​​are extracted from wavelet transform results on a frequency band scale of 1-10 Hz. Extract the amplitude or energy concentration of the impulse component of the signal from the wavelet transform results on the 10-100Hz frequency band scale; Calculate the proportion of noise energy in a specified high-frequency sub-band from the wavelet transform results of the 100-1000Hz frequency band. Envelope demodulation analysis was performed on the vibration signal to extract fault characteristic frequencies. The calculation formula is as follows:

[0081] in, This indicates the number of rolling elements in the bearing corresponding to the vibration signal; The diameter of the rolling element of the bearing corresponding to the vibration signal; The pitch circle diameter of the bearing corresponding to the vibration signal; This refers to the motor speed; The time-frequency domain features include energy entropy, kurtosis, impulse component amplitude or energy concentration, and noise energy proportion extracted from the wavelet transform results at each scale, as well as fault characteristic frequencies extracted from envelope demodulation analysis. .

[0082] By employing Morlet wavelet basis functions to perform three-level continuous wavelet transform to extract time-frequency domain features, and defining the extraction of features such as energy entropy, kurtosis, impact component amplitude, and noise energy proportion from different frequency band scales (1-10Hz, 10-100Hz, 100-1000Hz), as well as extracting fault feature frequencies by envelope demodulation of vibration signals, multi-scale and localized time-frequency analysis of non-stationary signals can be performed to accurately capture specific frequency components and impact characteristics that characterize mechanical faults (such as bearing wear).

[0083] In some specific embodiments, the center frequency of the Morlet wavelet basis function is 1 Hz and the bandwidth is 0.5 Hz.

[0084] By specifically limiting the center frequency of the Morlet wavelet basis function to 1Hz and the bandwidth to 0.5Hz, the wavelet transform parameters used to analyze vehicle current and vibration signals were optimized, making them more suitable for analyzing signal details in the target frequency band (especially the low-frequency part) in terms of time-frequency focusing characteristics, thereby improving the sensitivity of the extracted time-frequency domain features to faults.

[0085] In some specific embodiments, the method for extracting environmental coupling features is as follows: A pre-defined multivariate correlation function is used to calculate parameters in vehicle operating data, cloud environment data, and external operating condition data to obtain environmental coupling characteristics.

[0086] By using a pre-defined multivariate correlation function to calculate environmental coupling characteristics, a mathematical framework is provided that actively associates and integrates vehicle operating data with cloud-based environmental data and external operating condition data, enabling the characteristics to directly reflect the quantitative impact of environmental factors (such as temperature and slope) on component states (such as internal resistance and torque).

[0087] In some specific embodiments, the environmental coupling features include: Battery internal resistance correction value The corresponding multivariate correlation function is:

[0088] Motor torque compensation value The corresponding multivariate correlation function is:

[0089] in, This indicates the AC internal resistance in the vehicle's own operating data; This indicates the preset temperature influence coefficient; This indicates the air temperature corresponding to the vehicle's location in the cloud-based environmental data; Indicates the preset reference temperature; Motor torque in the vehicle's own operating data; This represents the road slope angle in the external working condition data.

[0090] By providing specific calculation formulas for the two environmental coupling characteristics—battery internal resistance correction value and motor torque compensation value—the abstract multivariate correlation is concretized into a calculable physical quantity correction model. This model can dynamically eliminate or compensate for the interference of environmental and operating condition changes on key state parameters, allowing subsequent anomaly detection to focus more on the actual degradation of the components themselves.

[0091] In some specific embodiments, the temperature influence coefficient The data was obtained by performing a multiple linear regression fit on historical operating data, with units of ℃. -1 The value ranges from 0.005 to 0.02.

[0092] The temperature influence coefficient k is obtained by fitting historical data through multiple linear regression, with a value range of 0.005-0.02 degrees Celsius. This provides a data-driven parameter determination method and reasonable range that conforms to physical experience for the battery internal resistance correction model, ensuring the accuracy and engineering applicability of the correction formula.

[0093] Step S4: Concatenate the data features into a feature sequence and input it into the pre-trained anomaly detection model. The anomaly detection model encodes the feature sequence and evaluates anomalies, outputting anomaly scores. The anomaly scores range from [0,1]. The anomaly detection model is built based on Transformer Encoder.

[0094] By inputting the concatenated feature sequence into a pre-trained anomaly detection model built on Transformer Encoder for encoding and evaluation, and outputting anomaly scores with continuous values, the attention mechanism can effectively capture long-distance temporal dependencies and complex interactions between features in the feature sequence, thereby achieving a refined and quantitative evaluation of the boundary between normal and abnormal states, thus replacing simple threshold judgment.

[0095] In some specific embodiments, in step S4, the anomaly detection model includes an input layer, four encoder layers, and an output layer, with each encoder layer sequentially performing a multi-head self-attention mechanism and a feedforward neural network operation; The multi-head self-attention mechanism uses a total of 4 independent attention heads; The hidden layer dimension of the feedforward neural network is 128, and the dimension of its internal feedforward network is 512. The model employs a sinusoidal positional encoding method to add temporal information to each position in the input feature sequence; The data processing procedure for the anomaly detection model is as follows: After the feature sequence is input through the input layer, the four encoder layers encode the input feature sequence in sequence. Then, the output layer maps the hidden state output by the last encoder layer to a scalar value with a range of [0,1], which is used as the anomaly score.

[0096] By specifically defining the anomaly detection model as a Transformer structure containing four encoder layers, each layer performs multi-head self-attention and feedforward network operations and adopts sinusoidal positional encoding, a powerful encoder architecture that can efficiently process long sequences and capture complex temporal dependencies is provided, laying the model foundation for learning normal behavior patterns and identifying deviations from high-dimensional feature sequences.

[0097] In some specific embodiments, the four attention heads of the multi-head attention mechanism include two general attention heads and two fault-sensitive attention heads that are more sensitive to low-frequency gradual fault patterns. The fault-sensitive attention heads learn weights during the training process to focus on feature dimensions in the input feature sequence that are related to low-frequency gradual fault patterns.

[0098] By setting two fault-sensitive attention heads in the multi-head attention mechanism that are more sensitive to low-frequency gradual fault patterns, the model's attention mechanism was specifically optimized, enabling it to adaptively pay more attention to feature dimensions related to slowly developing faults, thereby improving the detection capability of this type of highly concealed fault.

[0099] In some specific embodiments, the anomaly detection model undergoes INT8 integer quantization, and its inference computing power requirement on the edge computing unit does not exceed 5 TOPS.

[0100] By performing INT8 integer quantization on the anomaly detection model and limiting its inference computing power requirement on the edge computing unit to no more than 5 TOPS, the model size and computational complexity are greatly compressed, enabling the model to be deployed on vehicle edge devices with limited computing power and meeting the stringent latency requirements for real-time fault detection.

[0101] In some specific embodiments, the training method for the anomaly detection model is as follows: Acquire multi-source time series data collected within a historical time period, and after processing in the same way as steps S2 and S3, obtain multiple sets of data features to construct a training sample set; Based on the corresponding actual working conditions, each set of data features in the training sample set is assigned a working condition status label. The working condition status label corresponding to normal working conditions is 0, and the working condition status label corresponding to faulty working conditions is 1. Using the feature sequences in the training sample set as input and the corresponding working condition labels as supervision signals, the anomaly detection model is trained in a supervised manner using the binary cross-entropy loss function, so that the anomaly score output by the model approximates the working condition label until the model converges.

[0102] By describing in detail a model training method that uses historical data to construct a training sample set, assigns operating condition labels, and employs a binary cross-entropy loss function for supervised training, a complete and repeatable model building process is provided. This ensures that the model can effectively learn the boundary between normal and fault modes from labeled data and output reliable anomaly scores.

[0103] Step S5: Based on the abnormal score, trigger an early warning according to the preset warning rules.

[0104] By triggering warnings based on continuous anomaly scores from model output and according to preset warning rules, fault warning decisions can be made dynamically based on a comprehensive quantitative index calculated by the model, rather than relying on a fixed threshold of a single sensor signal. This reduces false alarms while improving the sensitivity and scientific nature of warnings for potential and complex faults.

[0105] In some specific embodiments, the warning rule in step S5 is specifically as follows: When the abnormal score exceeds the preset warning threshold, a level 2 warning is triggered; A Level 1 warning is triggered when the abnormal score exceeds the first threshold for three consecutive detection periods.

[0106] By specifying the early warning rules as triggering a secondary early warning when the abnormal score exceeds a preset threshold and triggering a primary early warning when the score exceeds the first threshold for three consecutive periods, a hierarchical and cumulatively triggered early warning logic is established. This logic can provide a prompt when there is a single minor abnormality and upgrade the warning level when there is a continuous abnormality, thus balancing the timeliness and reliability of early warnings and reducing false alarms caused by single fluctuations.

[0107] In some specific embodiments, when a level-two warning is triggered, a prompt message is sent to the user suggesting that they check within 24 hours; When a Level 1 warning is triggered, a message will be sent to the user suggesting that they stop the car immediately for inspection.

[0108] By specifying that a "recommend inspection within 24 hours" message is sent when a Level 2 warning is triggered and a "recommend immediate shutdown and inspection" message is sent when a Level 1 warning is triggered, the warning scores of different levels are transformed into clear and actionable maintenance instructions. This allows the warning output to directly serve the operation and maintenance decision-making process, thereby improving the system's practicality and security.

[0109] In some specific embodiments, the warning threshold is 0.8.

[0110] By setting the warning threshold to 0.8, a clear quantitative benchmark is provided for warning decisions based on anomaly scores, making the warning triggering conditions uniform and adjustable, and facilitating calibration according to the safety margin requirements of actual application scenarios.

[0111] Step S6: After triggering the warning, analyze the contribution of each feature item in the feature sequence to the abnormal score, match the contribution with the preset single-vehicle fault mode library according to rules, obtain the predicted abnormal type, and generate a diagnostic report based on the predicted abnormal type. The single-vehicle fault mode library stores historical labeled data, abnormal sample clustering results, abnormal type labels determined based on the abnormal sample clustering results, feature thresholds corresponding to abnormal types, and fault evolution paths composed of abnormal type sequences. The feature thresholds are the thresholds of feature items in the feature sequence, the thresholds of the contribution of feature items to the abnormal score, or a combination thereof.

[0112] By analyzing the contribution of each feature to the anomaly score and matching it with a pre-defined single-vehicle fault mode library to obtain the predicted anomaly type and generate a diagnostic report, the black-box output of the model can be transformed into interpretable diagnostic conclusions that point to specific fault causes and components. Combined with a historical fault knowledge base, it provides maintenance personnel with direct and targeted maintenance decision support.

[0113] In some specific embodiments, step S6, analyzing the contribution of each feature item in the feature sequence to the anomaly score, specifically includes: Deploy a lightweight SHAP interpreter; The TreeSHAP algorithm in the lightweight SHAP interpreter is used to calculate the feature sequence of the input anomaly detection model. The Shapley value of each feature item in the feature sequence is quantified as the contribution of that feature item to the anomaly score.

[0114] By specifically deploying a lightweight SHAP interpreter and utilizing its TreeSHAP algorithm to quantify the Shapley value of each feature item in the feature sequence as its contribution to the anomaly score, a model interpretability technology solution that is applicable to edge devices and computationally efficient is provided, which can decompose the complex decision of the model into the quantified influence of each input feature.

[0115] In some specific embodiments, step S6, which involves matching the contribution level with the single-vehicle fault mode library according to rules, specifically includes: Read the feature thresholds corresponding to each anomaly type stored in the single-vehicle fault mode library; The feature values ​​of each feature item in the feature sequence and the contribution of each feature item to the anomaly score are logically compared with the feature thresholds. If a certain feature threshold is met, the current warning is determined to match the anomaly type corresponding to that anomaly threshold.

[0116] By specifying rule matching as reading feature thresholds from the fault mode library and logically comparing feature values ​​and contribution with the thresholds, a specific matching mechanism is provided that combines interpretability analysis results (contribution) with traditional rule diagnosis (feature thresholds). This makes the generation of diagnostic conclusions both data-driven and constrained by knowledge rules.

[0117] In some specific embodiments, the rule matching logic includes: if the contribution of the feature item "cell voltage deviation" to the anomaly score is greater than 0.5, and the value of the feature item "cell internal resistance correction value" is greater than a preset threshold, then the matched anomaly type is battery consistency deterioration.

[0118] By providing a specific rule example, "If the contribution of feature item 'cell voltage deviation' to the anomaly score is greater than 0.5, and the value of feature item 'cell internal resistance correction value' is greater than a preset threshold, then the matched anomaly type is battery consistency deterioration," the logic of joint determination of contribution and feature value is vividly illustrated, demonstrating how interpretable output can be used to achieve refined and conditional fault type identification.

[0119] In some specific embodiments, the single-vehicle fault mode library also stores fault repair knowledge corresponding to the fault type; Troubleshooting knowledge includes recommended inspection parts, repair procedures, or a list of replacement parts.

[0120] By storing fault repair knowledge (such as recommended inspection parts and repair steps) corresponding to the fault types in the single-vehicle fault mode library, the fault diagnosis system can not only identify the fault type, but also directly associate and provide repair solutions, extending the diagnostic closed loop to repair guidance, and greatly improving the system's auxiliary value to on-site repair personnel.

[0121] In some specific embodiments, in step S6, the content of the diagnostic report includes: Predicted anomaly types; The contribution and value of each feature term; Trend charts of each feature item within a specified time period before and after the warning; Fault repair knowledge retrieved from the fault mode library that corresponds to the predicted anomaly type.

[0122] By specifying that the content of the diagnostic report includes the predicted anomaly type, the contribution and value of each feature item, the feature item trend chart, and the corresponding maintenance knowledge retrieved from the database, a complete and information-rich diagnostic report template is defined, which can comprehensively support maintenance decisions from multiple dimensions such as conclusions, basis, historical trends, and solutions.

[0123] In some specific embodiments, a single-vehicle fault mode library update step is also included, specifically including: When the warning meets the preset update conditions, a data segment within a specified time period before and after the warning is triggered is generated and uploaded to the cloud. The data segment includes the original signal, data characteristics, anomaly score and timestamp of multi-source time series data. Based on data fragments, a pre-built library of single-vehicle fault modes in the cloud is updated using a semi-supervised learning method.

[0124] By adding a step to update the single-vehicle fault mode library, and stipulating that relevant data fragments be uploaded to the cloud and the mode library be updated through a semi-supervised learning method when preset update conditions are met, the system has the ability to continuously learn from actual operation, and can continuously accumulate new fault cases, optimize feature thresholds, discover unknown fault modes, and realize the evolution of diagnostic knowledge.

[0125] In some specific embodiments, the preset update condition is: triggering a level one warning; When a Level 1 warning is triggered, data segments within a specified time period before and after the trigger time are generated and uploaded to the cloud.

[0126] By specifying the pre-defined update conditions as updating when a Level 1 warning is triggered, the update of the pattern library is ensured to be driven by the most urgent and specific fault events. This allows limited cloud computing and storage resources to focus on high-value fault samples, improving the efficiency and relevance of knowledge updates.

[0127] In some specific embodiments, the specified time period before and after the warning is: 100 seconds before the Level 1 warning time to 200 seconds after the Level 1 warning time.

[0128] By defining the time range of the data segments uploaded during the update as 100 seconds before and 200 seconds after the warning time, key data periods containing the stages of fault incubation, occurrence, and duration are identified, providing a data sample with sufficient time window and concentrated information for in-depth cloud analysis of the fault evolution process.

[0129] In some specific embodiments, step S5, updating the pre-built single-vehicle fault mode library in the cloud based on the data fragment using a semi-supervised learning method, specifically includes: Data corresponding to normal operating conditions in historical labeled data are used as labeled normal samples; For the data corresponding to the time when the warning was triggered in the data segment, the DBSCAN clustering algorithm was used for cluster analysis to form several unlabeled abnormal sample clusters; By combining labeled normal samples with aberrant sample clusters, new aberration types can be identified and defined, or the sample feature distribution of existing aberration types can be expanded, through label propagation or inductive learning. The newly identified or expanded anomaly types, their corresponding feature thresholds, and the correlation between fault evolution paths are updated to the single-vehicle fault mode library.

[0130] By describing the specific method of updating the cloud-based pattern library as combining labeled normal samples, using DBSCAN clustering to form abnormal sample clusters from early warning data, and then identifying new abnormal types or expanding existing types through label propagation or inductive learning, a feasible semi-supervised learning technology path is provided, which can optimize the fault pattern library using a large amount of unlabeled early warning data with only a small number of labels.

[0131] The following are embodiments of the fault data-driven analysis system for new energy commercial vehicles provided in this application. This fault data-driven analysis system for new energy commercial vehicles belongs to the same inventive concept as the fault data-driven analysis methods in the above embodiments. For details not described in detail in the embodiments of the fault data-driven analysis system, please refer to the embodiments of the fault data-driven analysis methods for new energy commercial vehicles described above.

[0132] like Figure 2 As shown, the fault data-driven analysis system for new energy commercial vehicles includes: The data acquisition module is used to collect multi-source time-series data, including vehicle operating data, cloud environment data, and external operating condition data. The data preprocessing module is used to preprocess multi-source time series data and generate a sliding time window dataset divided according to a preset time length. The feature extraction module is used to extract data features based on the sliding time window dataset, including basic features, time-frequency domain features, and environmental coupling features; The anomaly score generation module concatenates data features into a feature sequence, which is then input into a pre-trained anomaly detection model. The anomaly detection model encodes the feature sequence and evaluates for anomalies, outputting anomaly scores. The early warning triggering module is used to trigger early warnings based on the abnormal score and according to preset early warning rules; The diagnostic report generation module is used to analyze the contribution of each feature item in the feature sequence to the abnormal score after the warning is triggered, match the contribution with the preset single-vehicle fault mode library according to rules to obtain the predicted abnormal type, and generate a diagnostic report based on the predicted abnormal type.

[0133] The fault data-driven analysis system in this embodiment is used to implement a fault data-driven analysis method for new energy commercial vehicles.

[0134] This application also provides an electronic device for implementing the various embodiments of this application. Figure 3 To illustrate the hardware structure of an electronic device according to various embodiments of this application, as shown in the following diagram... Figure 3As shown, the electronic device includes a memory, a processor, and a computer program stored in the memory and capable of running on the processor.

[0135] Those skilled in the art will understand that the electronic device structure involved in the embodiments of this application does not constitute a limitation on the electronic device. The electronic device may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0136] In embodiments of this application, electronic devices include, but are not limited to, laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. Electronic devices may also represent various forms of mobile devices and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the embodiments of this application described and / or claimed herein.

[0137] In this application embodiment, the processor can be implemented using at least one of an Application-Specific Integrated Circuit (ASIC), a Digital Signal Processor (DSP), a Digital Signal Processing Device (DSPD), a processor, a controller, a microcontroller, a microprocessor, or an electronic unit designed to perform the functions described herein. In some cases, such implementations can be implemented within a controller. For software implementations, implementations such as processes or functions can be implemented with separate software modules that allow the performance of at least one function or operation. The software code can be implemented by a software application (or program) written in any suitable programming language, and the software code can be stored in memory and executed by the controller.

[0138] In addition, the electronic device includes some functional modules not shown, which will not be described in detail here.

[0139] Those skilled in the art will understand that the various aspects of the electronic device provided in this application can be implemented as a system, method, or program product. Therefore, the various aspects of this application can be specifically implemented in the following forms: a completely hardware implementation, a completely software implementation (including firmware, microcode, etc.), or a combination of hardware and software aspects, collectively referred to herein as a "circuit," "module," or "system."

[0140] This application also provides a storage medium storing a program product capable of implementing a fault data-driven analysis method for new energy commercial vehicles. In some possible implementations, various aspects of this application can also be implemented as a program product comprising program code, which, when run on a terminal device, causes the terminal device to perform the steps described in the "Exemplary Methods" section of this specification according to various exemplary embodiments of this application.

[0141] The storage medium may be any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example,, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of readable storage media include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0142] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A fault data-driven analysis method for new energy commercial vehicles, characterized in that, include: S1. Collect multi-source time series data, including vehicle operating data, cloud environment data, and external operating condition data; The vehicle's own operating data includes battery management system data, motor controller data, and vibration data. The battery management system data includes at least the charging and discharging current, and the motor controller data includes at least the motor speed. S2. Preprocess the multi-source time series data to generate a sliding time window dataset divided according to a preset time length; S3. Extract data features based on the sliding time window dataset, including basic features, time-frequency domain features, and environmental coupling features; Among them, the basic features include statistical features and frequency domain features extracted from the vehicle's own operating data; S4. Concatenate the data features into a feature sequence and input it into the pre-trained anomaly detection model. The anomaly detection model encodes the feature sequence and evaluates anomalies, and outputs anomaly scores. The anomaly scores range from [0,1]. The anomaly detection model is built based on Transformer Encoder. S5. Based on the abnormal score, trigger an alert according to the preset alert rules; S6. After triggering the warning, analyze the contribution of each feature item in the feature sequence to the abnormal score, match the contribution with the preset single-vehicle fault mode library according to rules to obtain the predicted abnormal type, and generate a diagnostic report based on the predicted abnormal type. The single-vehicle fault mode library stores historical labeled data, abnormal sample clustering results, abnormal type labels determined based on the abnormal sample clustering results, feature thresholds corresponding to abnormal types, and fault evolution paths composed of abnormal type sequences. The feature thresholds are the thresholds of feature items in the feature sequence, the thresholds of the contribution of feature items to the abnormal score, or a combination thereof.

2. The fault data-driven analysis method as described in claim 1, characterized in that, In step S1, the battery management system data includes individual cell voltage, battery module temperature, charging and discharging current, and AC internal resistance. Motor controller data includes motor speed and motor torque; The vibration signal is vibration acceleration sensor data that reflects the mechanical state of the motor or transmission system, collected at a sampling rate of not less than 1 kHz. The cloud-based environmental data is weather forecast data for a future preset time period, including at least the air temperature, humidity, and precipitation intensity corresponding to the vehicle's location; External operating condition data includes: Road slope angle and vehicle turning radius obtained through lidar or inertial measurement unit; Road surface slippage information is obtained by fusing wheel speed sensor signals and capacitive ground humidity sensor signals.

3. The fault data-driven analysis method as described in claim 1, characterized in that, In step S3, the method for extracting time-frequency domain features is as follows: Morlet wavelet basis functions were used to perform three-level continuous wavelet transforms on the battery current signal and vibration signal, respectively. The three levels of transform correspond to the frequency band scales of 1-10Hz, 10-100Hz and 100-1000Hz. Energy entropy and kurtosis values ​​are extracted from wavelet transform results on a frequency band scale of 1-10 Hz. Extract the amplitude or energy concentration of the impulse component of the signal from the wavelet transform results on the 10-100Hz frequency band scale; Calculate the proportion of noise energy in a specified high-frequency sub-band from the wavelet transform results of the 100-1000Hz frequency band. Envelope demodulation analysis was performed on the vibration signal to extract fault characteristic frequencies. The calculation formula is as follows: in, This indicates the number of rolling elements in the bearing corresponding to the vibration signal; The diameter of the rolling element of the bearing corresponding to the vibration signal; The pitch circle diameter of the bearing corresponding to the vibration signal; This refers to the motor speed; The time-frequency domain features include energy entropy, kurtosis, impulse component amplitude or energy concentration, and noise energy proportion extracted from the wavelet transform results at each scale, as well as fault characteristic frequencies extracted from envelope demodulation analysis. .

4. The fault data-driven analysis method as described in claim 1, characterized in that, The method for extracting environmental coupling features is as follows: A pre-defined multivariate correlation function is used to calculate parameters in vehicle operating data, cloud environment data, and external operating condition data to obtain environmental coupling characteristics.

5. The fault data-driven analysis method as described in claim 4, characterized in that, Environmental coupling characteristics include: Battery internal resistance correction value The corresponding multivariate correlation function is: Motor torque compensation value The corresponding multivariate correlation function is: in, This indicates the AC internal resistance in the vehicle's own operating data; This indicates the preset temperature influence coefficient; This indicates the air temperature corresponding to the vehicle's location in the cloud-based environmental data; Indicates the preset reference temperature; Motor torque in the vehicle's own operating data; This represents the road slope angle in the external working condition data.

6. The fault data-driven analysis method as described in claim 1, characterized in that, In step S4, the anomaly detection model includes an input layer, four encoder layers, and an output layer. Each encoder layer sequentially executes a multi-head self-attention mechanism and a feedforward neural network operation. The multi-head self-attention mechanism uses a total of 4 independent attention heads; The hidden layer dimension of the feedforward neural network is 128, and the dimension of its internal feedforward network is 512. The model employs a sinusoidal positional encoding method to add temporal information to each position in the input feature sequence; The data processing procedure for the anomaly detection model is as follows: After the feature sequence is input through the input layer, the four encoder layers encode the input feature sequence in sequence. Then, the output layer maps the hidden state output by the last encoder layer to a scalar value with a range of [0,1], which is used as the anomaly score.

7. The fault data-driven analysis method as described in claim 1, characterized in that, In step S5, the warning rules are as follows: When the abnormal score exceeds the preset warning threshold, a level 2 warning is triggered; A Level 1 warning is triggered when the abnormal score exceeds the first threshold for three consecutive detection periods.

8. The fault data-driven analysis method as described in claim 1, characterized in that, In step S6, the specific steps of matching the contribution score with the single-vehicle fault mode library include: Read the feature thresholds corresponding to each anomaly type stored in the single-vehicle fault mode library; The feature values ​​of each feature item in the feature sequence and the contribution of each feature item to the anomaly score are logically compared with the feature thresholds. If a certain feature threshold is met, the current warning is determined to match the anomaly type corresponding to that anomaly threshold.

9. The fault data-driven analysis method as described in claim 1, characterized in that, It also includes the steps for updating the single-vehicle fault mode library, specifically: When the warning meets the preset update conditions, a data segment within a specified time period before and after the warning is triggered is generated and uploaded to the cloud. The data segment includes the original signal, data characteristics, anomaly score and timestamp of multi-source time series data. Based on data fragments, a pre-built library of single-vehicle fault modes in the cloud is updated using a semi-supervised learning method.

10. A fault data-driven analysis system for new energy commercial vehicles, characterized in that, To implement the fault data-driven analysis method as described in any one of claims 1-9, the method includes: The data acquisition module is used to collect multi-source time-series data, including vehicle operating data, cloud environment data, and external operating condition data. The data preprocessing module is used to preprocess multi-source time series data and generate a sliding time window dataset divided according to a preset time length. The feature extraction module is used to extract data features based on the sliding time window dataset, including basic features, time-frequency domain features, and environmental coupling features; The anomaly score generation module concatenates data features into a feature sequence, which is then input into a pre-trained anomaly detection model. The anomaly detection model encodes the feature sequence and evaluates for anomalies, outputting anomaly scores. The early warning triggering module is used to trigger early warnings based on the abnormal score and according to preset early warning rules; The diagnostic report generation module is used to analyze the contribution of each feature item in the feature sequence to the abnormal score after the warning is triggered, match the contribution with the preset single-vehicle fault mode library according to rules to obtain the predicted abnormal type, and generate a diagnostic report based on the predicted abnormal type.

Images