Intelligent diagnosis and early warning system and method for power supply fault of electric vehicle

Abstract

The invention discloses an intelligent diagnosis and early warning system and method for a power supply fault of an electric vehicle. The system comprises a data acquisition module for acquiring running data of the current vehicle; the data transmission module performs data transmission; the data analysis module analyzes and compares the preprocessed current operation data with historical operation data sets of other similar vehicles, and obtains a user working condition related data set and a fault diagnosis related data set based on an analysis result; the user working condition learning module performs model training based on a space-time attention hybrid neural network algorithm, a user working condition related data set and operation data to obtain an updated user working condition learning model; and the fault prediction and warning module is used for training the updated user working condition learning model and the fault diagnosis model through a multi-source domain adaptive transfer learning algorithm based on the fault diagnosis related data to obtain an updated fault diagnosis model, inputting the processed fault diagnosis related data of the current vehicle into the updated fault diagnosis model, and carrying out fault diagnosis on the current vehicle. And obtaining a fault prediction result.

Description

Technical Field

[0001] This application relates to the field of electric vehicle diagnostic technology, and in particular to an intelligent diagnostic and early warning system and method for electric vehicle power supply faults. Background Technology

[0002] Currently, electric vehicle diagnostics refers to the use of various technologies to detect, analyze, and locate faults in key components and systems of electric vehicles, such as the power system, battery management system, charging system, and vehicle control system. This allows for the timely identification and resolution of potential problems, ensuring the safe and efficient operation of the vehicle. The power system of an electric vehicle consists of numerous electrical components and control modules. A failure in the power system can lead to serious consequences such as the vehicle failing to start, abnormal charging, and drive system failure, and may even cause safety accidents such as fires or explosions. Therefore, timely resolution of power system faults is crucial for electric vehicles.

[0003] However, existing electric vehicle power supply fault diagnosis systems typically require connecting to a diagnostic tool via an OBD-II interface to read vehicle fault codes and related data, lacking real-time capability. Therefore, it is necessary to propose an intelligent diagnostic and early warning system for electric vehicle power supply faults to at least address some of the aforementioned issues. Summary of the Invention

[0004] This invention provides an intelligent diagnostic and early warning system and method for electric vehicle power supply faults to solve the above-mentioned technical problems.

[0005] In a first aspect, embodiments of this application provide an intelligent diagnostic and early warning system for electric vehicle power supply faults, the system comprising: The data acquisition module is used to collect the current vehicle's operating data. The operating data includes user-related operating data of the current vehicle. If the vehicle has malfunctioned, the operating data also includes fault diagnosis-related data. The data transmission module is used to transmit the current vehicle operation data to the cloud server and the user operating condition learning module. The data transmission module transmits the collected operation data to the cloud server through 5G slicing technology. During the transmission process, key fault feature data is transmitted first, and its transmission priority is 3-5 times higher than that of conventional data transmission. The data transmission module utilizes 5G slicing technology to transmit collected operational data to the cloud server, employing a hybrid technology adapted to the high-security scenarios of autonomous driving. This hybrid technology includes quantum key distribution, AES-256 encryption symmetric algorithm, quantum random number generator, and 5G hard slicing isolation technology. It also applies a quantum-enhanced hybrid data transmission encryption algorithm, whose encryption algorithm layers include: a key negotiation layer, a data encryption layer, and a security enhancement layer. The key negotiation layer includes a QKD quantum key wireless distribution layer to achieve wide area network wireless transmission and local area network wired negotiation, and to achieve quantum security in the hybrid network. The data encryption layer includes an AES-256 stream encryption layer to adapt to the massive sensor data transmission in autonomous driving. The security enhancement layer includes a quantum random number generation key layer, with an entropy value ≥256 bits during the generation process to resist quantum computing attacks. The data analysis module is used to acquire the current operating data of the electric vehicle and the historical operating datasets of other similar vehicles (including relevant operating condition datasets and fault diagnosis datasets of other similar vehicles) stored in the cloud server. The data analysis module is also used to preprocess the current operating data of the electric vehicle to obtain processed current vehicle operating data. Further, it analyzes and compares the historical operating datasets of other similar vehicles in the cloud server and the processed current electric vehicle operating data to obtain analysis results. Based on the analysis results, it obtains the current vehicle user operating condition related dataset and fault diagnosis related dataset, and updates the similar vehicle related operating condition dataset and fault diagnosis dataset stored in the cloud server in real time. At the same time, it sends the current vehicle user operating condition related dataset to the user operating condition learning module and the current electric vehicle fault diagnosis related dataset to the fault prediction and warning module. The user operating condition related dataset includes the current electric vehicle user operating condition learning model and the processed user operating condition related operating data. The fault diagnosis related dataset includes the current electric vehicle fault diagnosis model and the processed fault diagnosis related data. The user operating condition learning module is used to train the user operating condition learning model of the current vehicle based on the spatiotemporal attention hybrid neural network algorithm, the user operating condition related dataset of the current vehicle, and the processed operating data of the current vehicle, so as to obtain an updated user operating condition learning model of the current vehicle; and to send the updated user operating condition learning model of the current vehicle and the processed user operating condition related operating data of the current vehicle back to the data analysis module for updating the cloud server database. If a data communication transmission failure prevents the preprocessing of the current operating data of the electric vehicle in the data analysis module and the cloud server, the preprocessing of the current user operating data of the electric vehicle will be completed in the user operating condition learning module of the vehicle, resulting in the processed user operating condition-related operating data of the current vehicle. This preprocessing process also includes the preprocessing of the fault diagnosis-related data of the current vehicle, resulting in the processed fault diagnosis-related data of the current vehicle. Simultaneously, since the user operating condition-related dataset of the current electric vehicle transmitted from the data analysis module of the cloud server has not been successfully obtained, the user operating condition learning model of the current vehicle transmitted from the data analysis module of the cloud server used for model training will be automatically replaced with the most recent user operating condition learning model of the current electric vehicle before the data communication transmission failure, until the data communication transmission failure is successfully resolved.

[0006] The fault prediction and warning module is used to train the updated user operating condition learning model and the fault diagnosis model based on the fault diagnosis-related data using a transfer learning algorithm, update the parameters of the fault diagnosis model to obtain an updated current vehicle fault diagnosis model, input the processed current vehicle fault diagnosis-related data into the updated current vehicle fault diagnosis model to obtain a fault prediction result, issue a warning if the fault prediction result meets preset warning conditions, and send the updated current vehicle fault diagnosis model, the processed current vehicle fault diagnosis-related data, and the fault prediction result back to the data analysis module for updating the cloud server database.

[0007] If the fault prediction and warning module cannot obtain the current vehicle's fault diagnosis-related dataset from the data analysis module due to a data communication transmission failure, then the fault prediction and warning module will obtain the preprocessed current vehicle fault diagnosis-related data from the user operating condition learning module. At the same time, since the fault diagnosis model of the current vehicle from the data analysis module in the cloud server was not successfully obtained, the fault diagnosis model of the current vehicle from the data analysis module in the cloud server used for model training will be automatically replaced with the most recent fault diagnosis model of the current vehicle before the data communication transmission failure, until the data communication transmission failure is successfully resolved.

[0008] The notification feedback module is used to provide notification feedback to the user based on the warning information; The human-machine interface module is used to display the warning information on the human-machine interface and to display the vehicle-related parameters of the current vehicle. The cloud server is used to store the updated similar vehicle related operating condition dataset and fault diagnosis dataset; The cloud server employs a composite secure storage method based on layered data encryption, distributed storage, end-to-end auditing, and dynamic access control mechanisms. Specifically, the layered data encryption uses the AES-256-GCM algorithm to encrypt the original dataset, encapsulates the AES dynamic key using the SM2 elliptic curve algorithm, and combines a multi-attribute layered attribute-based encryption mechanism, dividing user attributes into three layers: basic attributes, data attributes, and environmental attributes. A quantum random number generator provides the encryption entropy source. The basic attributes include identity and role; the data attributes include data sensitivity level; and the environmental attributes include access location and device security level. The distributed storage uses a distributed hash table to divide the encrypted data into data blocks, encodes each data block into a data fragment using erasure coding technology, and stores them distributed across multiple remote cloud nodes. The end-to-end auditing includes audit traceability, verifies data integrity through a provable data holding protocol, and records all data operation logs based on blockchain technology. The dynamic access control constructs a four-dimensional permission model based on user identity, role, data sensitivity, and operation time window, and achieves real-time verification and traceability of access requests through an API security gateway combined with blockchain identity authentication.

[0009] In one embodiment of the present invention, a data analysis module is deployed on the cloud server. The data analysis module is used to clean and process the current operating data of the electric vehicle to obtain initial operating data; and to extract features from the initial operating data to obtain processed operating data. The data analysis module is also used to analyze and compare the historical operating data and the processed operating data based on the vehicle's model, years of use, and / or degree of aging to obtain analysis results.

[0010] In one embodiment of the present invention, the user work condition learning module includes: The data receiving module is used to acquire the user's working condition related dataset and the current vehicle's operating data; The data encoding module processes the user operating condition related dataset and the current electric vehicle's operating data based on a dynamic hybrid standardization method to obtain the initial user operating condition related dataset and the processed user operating condition related operating data of the current vehicle. The data encoding module processes the user working condition related dataset based on a dynamic hybrid standardization method, including a time-series segmentation step, a distribution adaptation step, and a periodic enhancement step. The time-series segmentation step includes: dividing each data segment according to a preset time window and calculating the local statistics of each data segment; The distribution adaptation step includes: determining the data distribution type using the Shapiro-Wilk test, wherein if the data distribution type is a normal distribution, improved Z-Score standardization is performed using the following formula:

[0011] in, This is the median of the data segment. This represents the absolute deviation of the median of the data segment. This is a minimum value, used to reduce the impact of outliers; If the data distribution type is skewed, logarithmic transformation and max-min normalization are used for processing; The periodic enhancement step includes: performing a Fourier transform on the periodic data (such as daily mileage), extracting the periodic components, and fusing them into the original data; Based on the processing of the time-series segmentation step, the distribution adaptation step, and the periodic enhancement step, an initial user working condition related dataset is obtained. The model training output module is used to train the user condition learning model based on the spatiotemporal attention hybrid neural network algorithm, the initial user condition related dataset, and the processed user condition related operating data of the current vehicle. When the accuracy of the validation set of the user condition learning model improves less than a first preset threshold for five consecutive rounds, the learning rate of the user condition learning model is automatically adjusted and training continues until the user condition learning model converges, thus obtaining the updated user condition learning model.

[0012] In one embodiment of the present invention, the model training output module uses a spatiotemporal attention hybrid neural network algorithm to train the user working condition learning model. The spatiotemporal attention hybrid neural network structure used includes a bottom convolutional layer, a middle bidirectional long short-term memory network layer, and a top dual-channel attention layer. The bottom convolutional layer is a 1D-CNN (one-dimensional convolutional neural network), which includes three parallel convolutional kernels with sizes of 3, 5, and 7 respectively. These kernels are used to extract local spatial features at different time scales, including instantaneous power supply voltage fluctuations and short-term power supply current trends. The middle layer of the bidirectional long short-term memory network (LSTM) is used to capture long-term temporal dependencies (such as the battery temperature change pattern over 30 consecutive minutes) and dynamically adjust the weights using the hidden state output at each time step through a gating mechanism. The top-level dual-channel attention layer includes a spatial attention layer and a temporal attention layer; The spatial attention layer calculates sensor weights through a fully connected layer activated by the Sigmoid function (e.g., the weight of the battery sensor is dynamically higher than that of the light sensor, with a weight value range of 0.8-0.95). The temporal attention layer is used to weight the hidden states of the output of the Long Short-Term Memory (LSTM) network layer; its weighting process features include that the temporal weight of the charging phase is 1.5-2 times that of the idling phase.

[0013] The loss function for the spatiotemporal attention hybrid neural network algorithm is obtained by weighting the focal loss function and the mean squared error using the following formula:

[0014] Where Focal Loss is the focus loss function, and MSE is the mean squared error. This is the balance coefficient, with a value of 0.7. During training, a dynamic early stopping mechanism is adopted. When the accuracy of the validation set increases by less than the first preset threshold (0.1%) for five consecutive batches, the learning rate is automatically multiplied by 0.5 until the model converges, thus obtaining the updated user working condition learning model. The first preset threshold is 0.1%.

[0015] In one embodiment of the present invention, the fault prediction and warning module includes: The data source receiving module is used to acquire the fault diagnosis-related dataset and the updated user working condition learning model. The transfer learning module is used to construct a transfer learning model based on the updated user working condition learning model and the fault diagnosis model contained in the fault diagnosis related dataset. The model training module is used to train the transfer learning model using the backpropagation algorithm and the fault diagnosis related dataset, and to obtain the updated fault diagnosis model when the loss function of the fault diagnosis model is less than a second preset threshold. The fault prediction module is used to input the processed current vehicle fault diagnosis data into the updated fault diagnosis model to obtain the fault prediction result, and to issue a warning when the fault prediction result meets the dynamic warning conditions. The transfer learning module uses the fault diagnosis model and the updated user working condition model as source tasks to construct the transfer learning model. The construction process of the transfer learning is as follows: The convolutional layer of the source task is used as a feature extractor, and features are extracted using the following formula:

[0016] in, For the target task characteristics, Features of the source task; Let the source task model be , The parameters of the source task model are used for feature fine-tuning using the following formula;

[0017] in, For the target task model, These are the fine-tuned parameters. Update the values ​​for the parameters to be fine-tuned; Based on the gradient descent algorithm, the parameters are fine-tuned using the following formula:

[0018] in, The loss function used to train the fault diagnosis model is... For learning rate, The gradient of the parameters of the source task model; Based on a weighted average method, the features of the source task and the target task are fused using the following formula:

[0019] in, The characteristics after fusion The fusion coefficient; In joint transfer learning, the overall loss function can be expressed as a weighted combination of the target task loss and the source task loss using the following formula:

[0020] in, For the overall loss function, The source task loss function, Let the target task loss function be... The weights are the source task loss function weights. Weights for the target task loss function; The target task generates the updated fault diagnosis model through the feature extractor and fine-tuning model of the source task; In the joint transfer learning, the overall loss function adopts a staged dynamic adjustment strategy, which includes an initial stage dynamic adjustment strategy, a mid-stage dynamic adjustment strategy, and a convergence stage dynamic adjustment strategy. The initial dynamic adjustment strategy corresponds to the first 20% of iterations, and the overall loss function for the initial stage is: (Focusing on source domain knowledge transfer); The dynamic adjustment strategy for the intermediate stage corresponds to the middle 20%~80% of the iterations, and the overall loss function corresponding to the intermediate stage is: (Focusing on target domain adaptation); The dynamic adjustment strategy for the convergence phase corresponds to the remaining iterations (the last 20% of iterations), and the overall loss function for the convergence phase is: ; in, For source domain loss, For the target domain loss, Consistency constraints for tasks; To avoid knowledge conflicts.

[0021] In one embodiment of the present invention, if the transfer learning model is a multi-source domain adaptive transfer learning model, the transfer learning module constructs the multi-source domain adaptive transfer learning model based on the fault diagnosis model and the user working condition model as source tasks. The construction steps include a multi-source domain selection step, a dynamic domain weight adjustment step, an adversarial feature alignment step, and a feature fusion step. The multi-source domain selection step includes: selecting the three source domains with the highest cosine similarity to the vehicle model features of the current electric vehicle from the cloud server, and the similarity calculation formula is:

[0022] Where A is the target domain feature vector and B is the source domain feature vector; the three source domains with the highest similarity can include electric cars of the same brand and models with the same battery type.

[0023] The dynamic domain weight adjustment step includes: adjusting the source domain weights using the following formula. Updated dynamically during the training process:

[0024] in, The feature distribution distance between the source and target domains is updated every 100 batches of data processed. To and Orthogonal distribution distance; The adversarial feature alignment step includes: introducing a domain discriminator with a convolutional neural network structure, and aligning the feature distributions of the source and target domains by minimizing the following loss function:

[0025] in, The domain classification probability output by the discriminator. Here, N represents the domain label, and N represents the number of iterations. The feature fusion step includes: fusing the features of the multi-source domain and the target domain by weighted average using the following formula:

[0026] in, The characteristics after fusion For the i-th source domain feature, For target domain features, The source domain weights.

[0027] In one embodiment of the present invention, the model training module is used to train the transfer learning model using the backpropagation algorithm and the fault diagnosis-related data, calculate the gradient and update the model parameters, minimize the loss function, and obtain the updated fault diagnosis model when the loss function is less than a second preset threshold. During the training process, the target task parameters are calculated using the following formula:

[0028] in, These are the target task parameters for the next batch. The target task parameters for the current batch. For learning rate, This represents the gradient of the corrected overall loss function with respect to the target task parameters; If the backpropagation algorithm is a hierarchical adaptive backpropagation algorithm, when the model training module trains the transfer model using the hierarchical adaptive backpropagation algorithm, its training steps include a hierarchical learning rate optimization step, a momentum dynamic adjustment step, and a gradient clipping step. When employing the hierarchical adaptive backpropagation algorithm, the formula for calculating the target task parameters is updated as follows:

[0029] in, For tiered learning rates, Momentum factor These are the target task parameters from the previous batch; The hierarchical learning rate optimization step includes setting the learning rate of the Long Short-Term Memory (LSTM) layer to 0.001, the learning rate of the attention layer to 0.0005, and the learning rate of the discriminator to 0.002. The momentum dynamic adjustment step includes: adjusting the momentum factor using the following formula strategy. Adaptive adjustment as loss changes:

[0030] in, This represents the difference in loss between the current batch and the previous batch; a faster decrease in loss indicates... Increase the speed of convergence; The gradient clipping step includes: proportionally clipping parameters with a gradient norm exceeding 10.0 using the following formula:

[0031] Where grad is the gradient and threshold is the threshold value (which better protects the gradient of key features than traditional truncation).

[0032] In one embodiment of the present invention, the fault prediction module is used to input the processed fault diagnosis-related data of the current vehicle into the updated fault diagnosis model, and output a fault prediction result containing fault type, fault confidence and hazard level. When the product of fault confidence and hazard level in the fault prediction result is greater than a dynamic threshold, a warning message is generated and a warning is issued. The dynamic threshold is dynamically adjusted according to the current electric vehicle model, service life and / or aging degree.

[0033] In one embodiment of the present invention, the notification feedback module sends the warning information through multiple channels, including pushing real-time data visualization charts through an application installed on the user terminal, and sending content containing the warning information to the user terminal via SMS and / or email; the user can provide feedback on the processing results through the App, and the system will send the feedback data back to the cloud for model optimization.

[0034] For example, the notification feedback module 701 sends warning information through multiple channels. App push notifications include real-time data visualization charts (such as battery voltage fluctuation curves), while SMS messages contain a concise warning message (limited to 70 characters). Users can provide feedback on the processing results through the App, and the system will send the feedback data back to the cloud for model optimization.

[0035] The human-computer interaction interface module is used to display a battery health trend chart (based on the trend predicted by the improved model for the next 30 days), a fault risk heat map (marking high-risk components), and system settings on the visualized human-computer interaction interface.

[0036] Secondly, this application proposes an intelligent diagnosis and early warning method for power supply faults in electric vehicles, the method comprising: The system collects current vehicle operation data, obtains historical operation data of the current vehicle and other similar vehicles stored in the cloud server, preprocesses the operation data to obtain processed operation data, and analyzes and compares the historical operation data and the processed operation data to obtain analysis results. Based on the analysis results, user operating condition related datasets and fault diagnosis related datasets are obtained and the data on the cloud server is updated. The user operating condition related dataset includes a user operating condition learning model and processed user operating condition related data. The fault diagnosis related dataset includes a fault diagnosis model and processed fault diagnosis related data. The user condition learning model of the current vehicle is trained based on the spatiotemporal attention hybrid neural network algorithm, the user condition related dataset, and the processed user condition related operating data of the current vehicle to obtain the updated user condition learning model of the current vehicle. Based on the updated user working condition learning model and the fault diagnosis model, a transfer learning model is constructed using a multi-source domain adaptive transfer learning algorithm. The transfer learning model is trained using a hierarchical adaptive backpropagation algorithm and the fault diagnosis-related dataset. The gradient is calculated and the model parameters are updated to minimize the loss function. When the loss function is less than a second preset threshold, it indicates that the model can accurately diagnose faults to a certain extent. At this point, the updated fault diagnosis model is obtained.

[0037] The processed current vehicle fault diagnosis data is input into the updated fault diagnosis model to obtain fault prediction results that include fault confidence and hazard level. If the product of the fault confidence level and the hazard level in the fault prediction result is greater than the dynamic threshold, an early warning message is generated and an early warning is issued. Based on the early warning information, users are notified and feedback is provided. The warning information is displayed on the human-computer interaction interface, along with the vehicle-related parameters of the current vehicle.

[0038] Thirdly, this application also proposes a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the steps of the intelligent diagnostic and early warning system for electric vehicle power supply faults of the second aspect.

[0039] In summary, the intelligent diagnostic and early warning system for electric vehicle power supply faults according to embodiments of this application, through a trained fault prediction model combined with the electric vehicle's operating data and historical operating data, can predict fault prediction results. When the fault prediction results meet preset early warning conditions, an early warning is issued, enabling analysis of the real-time operating status of the electric vehicle and providing timeliness. Furthermore, the trained fault prediction model improves the accuracy of fault prediction results, thereby enhancing the safety of electric vehicle use. Other advantages, objectives, and features of the intelligent diagnostic and early warning system for electric vehicle power supply faults proposed in this application will be partly apparent from the following description, and partly understood by those skilled in the art through research and practice of this application. Attached Figure Description

[0040] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit this specification. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This is a schematic diagram of the structure of an intelligent diagnostic and early warning system for electric vehicle power supply faults provided in an embodiment of this application; Figure 2 This is a flowchart illustrating an intelligent diagnosis and early warning method for power supply faults in electric vehicles, provided in an embodiment of this application. Detailed Implementation

[0041] To better understand the technical solutions provided in the embodiments of this specification, the technical solutions of the embodiments of this specification will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of this specification and the specific features in the embodiments are detailed descriptions of the technical solutions of the embodiments of this specification, rather than limitations on the technical solutions of this specification. In the absence of conflict, the embodiments of this specification and the technical features in the embodiments can be combined with each other.

[0042] In this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, without necessarily requiring or implying any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. The term "two or more" includes two or more cases.

[0043] Please see Figure 1 This is a flowchart illustrating an intelligent diagnosis and early warning method for electric vehicle power supply faults provided in an embodiment of this application. Specifically, it may include: a data acquisition module 101, a cloud server 201, a data analysis module 2011, a user operating condition learning module 301, a fault prediction and warning module 401, a data transmission module 501, a notification feedback module 701, and a human-machine interaction interface module 601. The specific functions implemented by each module will be described in detail below.

[0044] The data acquisition module 101 is used to collect the current vehicle's operating data. The operating data includes user-related operating data of the current vehicle. If the vehicle has malfunctioned, the operating data also includes fault diagnosis-related data. The data transmission module 501 is used to transmit the current vehicle operation data to the cloud server 201 and the user operating condition learning module 301. The data analysis module 2011 is used to acquire the current operating data of the current vehicle and the historical operating datasets of other similar vehicles stored in the cloud server 201. The historical operating datasets of other similar vehicles include relevant operating condition datasets and fault diagnosis datasets of other similar vehicles. The data analysis module is also used to preprocess the current operating data of the current vehicle to obtain processed current operating data, analyze and compare the historical operating datasets of other similar vehicles in the cloud server with the processed current operating data to obtain analysis results, obtain the current vehicle user operating condition related dataset and fault diagnosis related dataset based on the analysis results, and update the relevant operating condition datasets and fault diagnosis datasets of other similar vehicles stored in the cloud server in real time. Simultaneously, the current vehicle user operating condition related dataset is sent to the user operating condition learning module, and the current vehicle fault diagnosis related dataset is sent to the fault prediction and warning module 401. The user operating condition related dataset includes the current vehicle user operating condition learning model and the processed user operating condition related operating data, and the fault diagnosis related dataset includes the current vehicle fault diagnosis model and the processed fault diagnosis related data. The user operating condition learning module 301 is used to train the user operating condition learning model of the current vehicle based on the spatiotemporal attention hybrid neural network algorithm, the user operating condition related dataset of the current vehicle, and the processed operating data of the current vehicle, so as to obtain an updated user operating condition learning model of the current vehicle; and to send the updated user operating condition learning model of the current vehicle and the processed user operating condition related operating data back to the data analysis module for updating the cloud server database. If a data communication transmission failure prevents the data analysis module and the cloud server from preprocessing the current operating data of the current vehicle, then the user operating condition learning module of this vehicle is used to preprocess the current operating data and fault diagnosis related data of the current vehicle. It can also preprocess the user operating condition related operating data to obtain the processed user operating condition related operating data and fault diagnosis related data of the current vehicle. The most recent user operating condition learning model of the current vehicle transmitted by the data analysis module before the data communication transmission failure is used for model training until the data communication transmission failure is successfully eliminated. The fault prediction and warning module 401 is used to train the updated user operating condition learning model and the fault diagnosis model based on the fault diagnosis-related data using a transfer learning algorithm, update the parameters of the fault diagnosis model to obtain an updated current vehicle fault diagnosis model, input the processed current vehicle fault diagnosis-related data into the updated current vehicle fault diagnosis model to obtain a fault prediction result, issue a warning if the fault prediction result meets preset warning conditions, and send the updated current vehicle fault diagnosis model, current vehicle fault data, and fault prediction result back to the data analysis module for updating the cloud server database. If a data communication transmission failure prevents the fault prediction and warning module from obtaining the current vehicle's fault diagnosis-related dataset from the data analysis module, the fault prediction and warning module will obtain the preprocessed current vehicle fault diagnosis-related data from the user working condition learning module, and use the most recent fault diagnosis model of the current vehicle from the data analysis module before the data communication transmission failure to train the model until the data communication transmission failure is successfully eliminated. The notification feedback module 701 is used to provide notification feedback to users based on early warning information; The human-machine interface module 601 is used to display the warning information on the human-machine interface and to display the vehicle-related parameters of the current vehicle. The cloud server 201 is also used to store updated datasets of similar vehicle operating conditions and fault diagnosis datasets.

[0045] Cloud server 201 uses a composite secure storage method based on layered encryption, distributed storage, process auditing, and dynamic access control mechanisms to store data, specifically including: 1) Layered Data Encryption: The original dataset is encrypted using the AES-256-GCM (Advanced Encryption Standard-256-bit-Galois / Counter Mode. AES is a symmetric encryption algorithm using a 256-bit key length) algorithm. The AES dynamic key is encapsulated using the SM2 elliptic curve cryptography algorithm (SM2 is an elliptic curve public-key cryptography algorithm independently designed in my country, belonging to the asymmetric encryption system). Combined with the Multi-Hierarchy Attribute-Based Encryption (MH-ABE) mechanism, user attributes are divided into three layers: basic attributes, data attributes, and environmental attributes. A quantum random number generator (QRNG) is superimposed to provide an encryption entropy source. The basic attributes include identity and role, the data attributes include data sensitivity level, and the environmental attributes include access location and device security level. 2) Distributed storage: The encrypted data is divided into data blocks using a distributed hash table (DHT), and each data block is encoded into a data fragment using erasure coding technology, which is then distributed and stored on multiple remote cloud nodes; 3) Audit traceability: Data integrity verification is achieved through the Provable Data Possession (PDP) protocol, and a full data operation log is recorded based on blockchain technology; 4) Dynamic access control: A four-dimensional permission model is built based on user identity, role, data sensitivity, and operation time window. Real-time verification and traceability of access requests are achieved through an API (Application Programming Interface) security gateway combined with blockchain identity authentication.

[0046] To verify the core advantages of the composite secure storage method of this invention, traditional single-encryption storage (AES + centralized storage) and industry-standard secure storage (AES + basic permissions + distributed storage) were selected as comparison groups. Experiments were conducted focusing on three core indicators: security protection, storage performance, and access control. The results are as follows: Table 1 Comparison of Core Indicators

[0047] The analytical conclusions are as follows: 1. Security Protection: This application achieves a data leakage prevention rate and unauthorized access blocking rate of nearly 100% through layered encryption and dynamic access control, significantly reducing the risk of data leakage compared to conventional industry solutions; 2. Storage efficiency: By adopting erasure coding technology instead of traditional multiple copies, storage overhead is reduced by 14.3% while ensuring data reliability, thus balancing security and cost; 3. Access Control: The four-dimensional permission model enables fine-grained control, improves permission matching accuracy to 99%, avoids unauthorized access, and reduces access latency by 15%, without affecting the user experience.

[0048] Experiments show that this composite secure storage method is superior to traditional solutions in terms of protection effectiveness, storage efficiency, and access control, and can effectively prevent customer data leakage.

[0049] The current vehicle is an electric vehicle. For example, the data acquisition module 101 collects operational data from the electric vehicle via sensors. This data can be used for vehicle status monitoring, fault diagnosis, performance optimization, and range prediction. The sensors include at least one of the following: temperature sensor, pressure sensor, light sensor, angular velocity sensor, distance sensor, tilt sensor, current sensor, oxygen sensor, driver facial recognition sensor, speed sensor, acceleration sensor, torque sensor, battery voltage sensor, battery internal resistance sensor, tire pressure monitoring sensor, vision sensor, ultrasonic sensor, and lidar sensor. The operational data includes at least one of the following: motor parameters, battery usage data, vehicle driving data, and user driving data. Motor parameters include motor speed, motor torque, motor temperature, and motor power. Battery usage data includes battery charge, battery voltage, battery current, battery temperature, and charging status. Vehicle driving data includes vehicle speed, mileage, acceleration, braking status, and steering angle. User driving data includes driving habits, driving routes, charging behavior, and driving mode selection. Among them, the data acquisition module 101 adds an edge computing node, which completes real-time feature extraction (such as battery SOC change rate and motor torque fluctuation coefficient) locally in the vehicle, reducing the amount of data transmitted to the cloud.

[0050] The cloud server 201 plays a crucial role in data storage, preserving various data generated during past operations. This historical operational data forms a vital foundation for subsequent analysis and learning. Preprocessing the operational dataset yields processed operational data. The aim is to make the data more targeted and meet the specific needs of different modules. Preprocessing removes noisy data, fills in missing values, and standardizes the data, improving data quality and usability. The preprocessed operational data is then analyzed and compared with existing historical operational data to obtain analytical results. Based on these results, the historical and processed operational data are integrated to create user-specific operational condition datasets and fault diagnosis datasets. The fault diagnosis datasets include fault diagnosis models and related data.

[0051] The user work condition learning module 301 uses the user work condition-related dataset integrated by the cloud server 201 to train the user work condition learning model. By continuously adjusting the model's parameters, the model can accurately describe and predict user work condition behaviors. The trained data is then fed back to the cloud server 201. By incorporating new knowledge into historical data, historical data can more accurately reflect current and future user work conditions, providing a more reliable basis for subsequent analysis and decision-making.

[0052] The fault prediction and warning module 401 uses a multi-source domain adaptive transfer learning algorithm to train the updated user operating condition learning model and the fault diagnosis model based on a fault diagnosis-related dataset. This algorithm integrates knowledge from multiple related source tasks, improving the model's adaptability to the target task (fault diagnosis for a specific vehicle model) through dynamic weight adjustment and adversarial feature alignment, thus overcoming the limitations of single-source domain transfer. During training, the parameters of the fault diagnosis model are adjusted and optimized to obtain an updated model. Fault data is input into the updated model, which can be user operating condition related datasets, fault diagnosis related data, or both. The updated model analyzes and calculates based on this input data to predict whether a fault will occur in the equipment or system and the possible scenarios, ultimately obtaining the fault prediction result. When the fault prediction result meets pre-set warning conditions, the system triggers a warning mechanism. The warning conditions are dynamically set by combining the fault confidence level (the fault probability output by the model) and the fault severity level (e.g., a battery short circuit severity level of 5, and abnormal motor noise of 2). The higher the severity level, the lower the warning threshold. Once the warning conditions are met, the system will promptly issue an alarm to notify relevant personnel to take measures to prevent or respond to potential faults and reduce the losses and impacts caused by the faults.

[0053] In summary, the intelligent diagnostic and early warning system for electric vehicle power supply faults proposed in this application, through a trained fault prediction model combined with the electric vehicle's operating data and historical operating data, can predict fault prediction results. When the fault prediction results meet preset early warning conditions, an early warning is issued, enabling analysis of the electric vehicle's real-time operating status and providing timeliness. Furthermore, the trained fault prediction model improves the accuracy of fault prediction results, thereby enhancing the safety of electric vehicle use.

[0054] In some examples, the data analysis module 2011 is deployed on the cloud server 201. The data analysis module 2011 is used to clean and process the current operating data of the current vehicle to obtain initial operating data; to extract features from the initial operating data to obtain processed current operating data; the data analysis module 2011 is also used to analyze and compare the historical operating dataset and the processed current operating data based on the vehicle model, years of use and / or degree of aging of the current vehicle to obtain analysis results.

[0055] For example, the running dataset may contain noise, outliers, and missing values, requiring cleaning and processing. Noise can be removed using methods such as wavelet filtering (more suitable for non-stationary signals than traditional filtering). Outliers can be detected or processed using statistical analysis or machine learning algorithms, such as the Isolation Forest algorithm (more suitable for high-dimensional data than statistical analysis). Missing values ​​can be supplemented using methods such as interpolation and imputation, for example, attention-based interpolation methods (prioritizing the preservation of the temporal correlation of key features). Feature extraction is then performed on the cleaned and processed initial running dataset to obtain the processed running data.

[0056] In some examples, the cloud server stores historical operating datasets of other similar vehicles. These other similar vehicles can be vehicles of the same model as the current vehicle, or vehicles using essentially the same components as the current vehicle; no limitation is made here. After collecting relevant vehicle data, other similar vehicles can also upload the data to the cloud server for storage. In the embodiments of this specification, the historical operating datasets of other similar vehicles include relevant operating condition datasets and fault diagnosis datasets. The relevant operating condition datasets may include various sensor data, vehicle type data, battery usage data, vehicle driving data, etc., of the corresponding vehicle. The fault diagnosis datasets may include diagnostic data generated when the corresponding vehicle experiences a fault.

[0057] Since the cloud server stores a large number of historical operation datasets of other similar vehicles, in order to obtain a more accurate model during the subsequent model training process, in this embodiment of the specification, the processed current operation data of the current vehicle can be analyzed and compared with the historical operation datasets of other similar vehicles to filter out vehicles with similar vehicle states, and the current vehicle user operating condition related dataset and fault diagnosis related dataset can be constructed based on the historical operation datasets corresponding to these vehicles.

[0058] In some embodiments, the processed current operating data of the current vehicle can be analyzed and compared with the historical operating datasets of other similar vehicles by comprehensively considering the vehicle model, years of use, and / or degree of aging. For example, the similarity between the current vehicle and other vehicles in terms of model, years of use, and / or degree of aging can be calculated, and vehicles with a similarity greater than a preset similarity can be filtered out. Of course, other methods can also be used to filter other similar vehicles, which are not limited here. In some embodiments, the current vehicle operating condition related dataset and the fault diagnosis related dataset may include data from the current vehicle and data from other filtered vehicles to enrich the data volume and improve the accuracy of subsequent model training.

[0059] In some embodiments, the user working condition learning module 301 includes a data receiving module, a data encoding module, and a model training output module.

[0060] The data receiving module is used to acquire the user's working condition related dataset and the current vehicle's operating data; The data encoding module processes the user operating condition related dataset and the current vehicle's operating data based on a dynamic hybrid normalization method to obtain an initial user operating condition related dataset and processed current vehicle user operating condition related operating data. The data encoding module processes the user operating condition related dataset based on a dynamic hybrid normalization method, including a time-series segmentation step, a distribution adaptation step, and a periodic enhancement step.

[0061] The time-series segmentation step includes: dividing each data segment according to a preset time window and calculating the local statistics of each data segment.

[0062] The distribution adaptation step includes: determining the data distribution type using the Shapiro-Wilk test, wherein if the data distribution type is a normal distribution, improved Z-Score standardization is performed using the following formula:

[0063] in, This is the median of the data segment. This represents the absolute deviation of the median of the data segment. This is a minimum value, used to reduce the impact of outliers; If the data distribution is skewed, logarithmic transformation and max-min normalization are used for processing.

[0064] The periodic enhancement step includes: performing a Fourier transform on the periodic data, extracting the periodic components and fusing them into the original data, wherein the periodic data includes, for example, daily mileage.

[0065] Based on the processing of the time-series segmentation step, the distribution adaptation step, and the periodic enhancement step, an initial user working condition related dataset is obtained. The model training output module is used to train the user condition learning model based on the spatiotemporal attention hybrid neural network algorithm, the initial user condition related dataset, and the processed user condition related operating data of the current vehicle. When the accuracy of the validation set of the user condition learning model improves less than a first preset threshold for five consecutive rounds, the learning rate of the user condition learning model is automatically adjusted and training continues until the user condition learning model converges, thus obtaining the updated user condition learning model.

[0066] Specifically, in the process of training the user condition learning model, the loss function in the model training output module is used to measure the difference between the model's predictions and the actual results. During training, the model's goal is to minimize the value of the loss function. When the loss function of the user condition learning model is less than a threshold, it indicates that the model has been able to fit the data well and has achieved the expected training effect.

[0067] For example, in a neural network, an epoch refers to the process of performing a forward and backward propagation through the entire training dataset to update all parameters. In other words, an epoch means that all samples in the training dataset have undergone one training process. In deep learning, the epoch is an important concept. It represents the process of completing a full traversal of the training data. In other words, each epoch is an iteration over the entire training set. During this process, the deep learning model uses all the training data to update its internal parameters. The main role of the epoch is to help the model understand the trends and patterns of the entire training set. Through multiple iterations (i.e., multiple epochs), the model can gradually optimize its parameters to more accurately predict the output results. The first preset threshold corresponding to a continuous improvement in validation set accuracy over five epochs can be set according to actual needs. For example, the first preset threshold can be 0.1%. When the validation set accuracy of the user condition learning model improves by less than 0.1% over five consecutive epochs, the learning rate is automatically adjusted and training continues until the model converges, resulting in an updated user condition learning model.

[0068] In the data encoding module, which processes user-related datasets using a dynamic hybrid standardization method, the time-series segmentation step involves dividing the data into segments according to a preset time window. This preset time window can be set according to actual needs, for example, 10 seconds. Calculating local statistics for each data segment allows users to characterize the local distribution, fluctuation patterns, and trends of the data within that event window.

[0069] In some examples, the model training output module uses a spatiotemporal attention hybrid neural network algorithm (ST-Attention Net) to train the user condition learning model. The spatiotemporal attention hybrid neural network structure used includes a bottom convolutional layer, a middle bidirectional long short-term memory network layer, and a top dual-channel attention layer. The bottom convolutional layer includes three parallel convolutional kernels with sizes of 3, 5, and 7 respectively, which are used to extract local spatial features at different time scales. The local spatial features include instantaneous power supply voltage fluctuations and short-term power supply current trends. The bottom convolutional layer is a 1D-CNN. The middle layer of the bidirectional long short-term memory (LSTM) network is used to capture long temporal dependencies and dynamically adjust the weights through a gating mechanism using the hidden state output at each time step. The long temporal dependencies include the battery temperature change pattern over a continuous 30-minute period. The top-level dual-channel attention layer includes a spatial attention layer and a temporal attention layer; The spatial attention layer calculates sensor weights through a fully connected layer activated by a Sigmoid function. The sensor weights include those of a battery sensor and a light sensor, with the battery sensor weight dynamically higher than that of the light sensor, and the weight value range being 0.8-0.95. The temporal attention layer is used to weight the hidden states output by the Long Short-Term Memory (LSTM) network layer (in some embodiments, the weight during the charging phase is 1.5-2 times that during the idling phase). The loss function for the spatiotemporal attention hybrid neural network algorithm is obtained by weighting the focal loss function and the mean squared error using the following formula:

[0070] Where Focal Loss is the focus loss function, and MSE is the mean squared error. The balance coefficient is set to 0.7 to address the imbalance of fault samples. During training, a dynamic early stopping mechanism is adopted. When the accuracy of the validation set increases by less than a preset value for five consecutive batches, the learning rate is automatically multiplied by 0.5 until the model converges, thus obtaining the updated user working condition learning model. The preset value is 0.1%.

[0071] Once the user work condition learning model's training meets the conditions (loss function less than a threshold), the model training output module updates the user work condition learning model. This updated model training output module then sends this data back to the cloud server 201, which uses this new information to update the historical operational data. Updating the historical operational data makes the data more accurate and complete, reflecting the latest user work conditions and providing a more reliable basis for subsequent data analysis, prediction, and decision-making. In this way, the system can continuously adapt to and learn from changes in user behavior, improving its performance and service quality.

[0072] We selected 100 electric vehicles of the same type and collected 3 months of operating data (including 2000+ operating condition samples) to compare the performance of the spatiotemporal attention hybrid neural network model in this application with that of the traditional LSTM model: Table 2: Experimental Verification: Performance Comparison of User Working Condition Learning Models

[0073] Note: The model in this application significantly improves the recognition accuracy of complex operating conditions such as charging and idling through multi-scale convolution and attention mechanisms, and improves the preprocessing efficiency by 51%.

[0074] In some embodiments, the fault prediction and warning module 401 includes a data source receiving module, a transfer learning module, a model training module, and a fault prediction module.

[0075] The data source receiving module is used to acquire the fault diagnosis-related dataset and the updated user working condition learning model. The transfer learning module is used to construct a transfer learning model based on the updated user working condition learning model and the fault diagnosis model contained in the fault diagnosis related dataset. The model training module is used to train the transfer learning model using the backpropagation algorithm and the fault diagnosis related dataset. When the loss function of the fault diagnosis model is less than a second preset threshold, an updated fault diagnosis model is obtained. The fault prediction module is used to input the processed current vehicle fault diagnosis data into the updated fault diagnosis model to obtain the fault prediction result, and to issue an early warning when the fault prediction result meets the dynamic early warning conditions. The transfer learning module uses the fault diagnosis model and the updated user working condition model as source tasks to construct the transfer learning model. The transfer learning model can be a multi-source domain adaptive transfer learning model. The construction process of the transfer learning is as follows: The convolutional layer of the source task is used as a feature extractor, and features are extracted using the following formula:

[0076] in, For the target task characteristics, Features of the source task; Let the source task model be , The parameters of the source task model are used for feature fine-tuning using the following formula;

[0077] in, For the target task model, These are the fine-tuned parameters. Update the values ​​for the parameters to be fine-tuned; Based on the gradient descent algorithm, the parameters are fine-tuned using the following formula:

[0078] in, The loss function used to train the fault diagnosis model is... For learning rate, The gradient of the parameters of the source task model; Based on a weighted average method, the features of the source task and the target task are fused using the following formula:

[0079] in, The characteristics after fusion The fusion coefficient; In joint transfer learning, the overall loss function can be expressed as a weighted combination of the target task loss and the source task loss using the following formula:

[0080] in, For the overall loss function, The source task loss function, Let the target task loss function be... The weights are the source task loss function weights. Weights for the target task loss function; The target task generates the updated fault diagnosis model through the feature extractor and fine-tuning model of the source task; In the joint transfer learning, the overall loss function adopts a staged dynamic adjustment strategy, which includes an initial stage dynamic adjustment strategy, a mid-stage dynamic adjustment strategy, and a convergence stage dynamic adjustment strategy. The initial dynamic adjustment strategy corresponds to the first 20% of iterations, and the overall loss function for the initial stage is: The initial stage dynamic adjustment strategy focuses on source domain knowledge transfer. The dynamic adjustment strategy for the intermediate stage corresponds to the middle 20%~80% of the iterations, and the overall loss function corresponding to the intermediate stage is: The mid-term dynamic adjustment strategy focuses on target domain adaptation. The dynamic adjustment strategy for the convergence phase corresponds to the remaining iterations, and the overall loss function for the convergence phase is: The remaining iterations are the last 20% of iterations; in, For source domain loss, For the target domain loss, Consistency constraints for tasks; , Avoid knowledge conflicts.

[0081] In some embodiments, the dynamic adjustment strategy in the initial stage can correspond to the first 20% of iterations, the dynamic adjustment strategy in the middle stage can correspond to the next 20%-80% of iterations, and the dynamic adjustment strategy in the convergence stage can correspond to the last 20% of iterations.

[0082] In some examples, if the transfer learning model is a multi-source domain adaptive transfer learning model, the transfer learning module constructs the multi-source domain adaptive transfer learning model based on the fault diagnosis model and the user working condition model as source tasks. The construction steps include a multi-source domain selection step, a dynamic domain weight adjustment step, an adversarial feature alignment step, and a feature fusion step. The multi-source domain selection step includes: filtering from the cloud server the three source domains with the highest cosine similarity to the vehicle model features of the current vehicle. In some embodiments, the three source domains with the highest cosine similarity to the vehicle model (e.g., an electric SUV) features of the current vehicle (such as electric sedans of the same brand or other vehicle models with the same battery type) can be filtered from the cloud server. The similarity calculation formula is:

[0083] Where A is the target domain feature vector and B is the source domain feature vector. The three source domains with the highest similarity can include electric cars of the same brand and models with the same battery type. The dynamic domain weight adjustment step includes: adjusting the source domain weights using the following formula. Updated dynamically during the training process:

[0084] in, The feature distribution distance between the source and target domains is updated every 100 batches of data processed. To and Orthogonal distribution distance. It should be noted that in transfer learning, "batch" refers to the subset of data used each time the model is trained, consistent with the definition in conventional deep learning.

[0085] The adversarial feature alignment step includes: introducing a domain discriminator with a convolutional neural network (CNN) structure, and aligning the feature distributions of the source and target domains by minimizing the following loss function:

[0086] in, The domain classification probability output by the discriminator. Here, N represents the domain label, and N represents the number of iterations. The feature fusion step includes: fusing the features of the multi-source domain and the target domain by weighted average using the following formula:

[0087] in, The characteristics after fusion For the i-th source domain feature, For target domain features, The source domain weights.

[0088] Fifty electric vehicles of different service life (including 20 vehicles older than 3 years) were selected, and 10 common power supply faults (battery short circuit, cell aging, etc.) were simulated. The diagnostic performance of the multi-source domain adaptive transfer learning model of this application was compared with that of the traditional single model. Table 3: Experimental Verification: Comparison of Fault Diagnosis Model Performance

[0089] Note: The accuracy of this model in diagnosing faults in older vehicles is significantly improved (from 68.3% to 91.5%), and the false alarm rate is reduced by 74%, allowing users ample time to process the faults.

[0090] In some examples, the model training module is used to train the transfer learning model using the backpropagation algorithm and the fault diagnosis-related data, calculate gradients and update model parameters, minimize the loss function, and obtain the updated fault diagnosis model when the loss function is less than a second preset threshold. It should be noted that when the loss function is less than the second preset threshold, it indicates that the model can accurately diagnose faults to a certain extent. At this point, the updated fault diagnosis model is obtained, which has higher accuracy and reliability in handling fault diagnosis tasks compared to the initial fault diagnosis model.

[0091] During training, the target task parameters are calculated using the following formula:

[0092] in, These are the target task parameters for the next batch. The target task parameters for the current batch. For learning rate, This represents the gradient of the corrected overall loss function with respect to the target task parameters; If the backpropagation algorithm is a hierarchical adaptive backpropagation algorithm, when the model training module trains the transfer model using the hierarchical adaptive backpropagation algorithm, its training steps include a hierarchical learning rate optimization step, a momentum dynamic adjustment step, and a gradient clipping step.

[0093] When employing the hierarchical adaptive backpropagation algorithm, the formula for calculating the target task parameters is updated as follows:

[0094] in, For tiered learning rates, Momentum factor These are the target task parameters from the previous batch; The hierarchical learning rate optimization step includes setting the learning rate of the Long Short-Term Memory (LSTM) layer to 0.001, the learning rate of the attention layer to 0.0005, and the learning rate of the discriminator to 0.002. The momentum dynamic adjustment step includes: adjusting the momentum factor using the following formula strategy. Adaptive adjustment as loss changes:

[0095] in, This represents the difference in loss between the current batch and the previous batch; a faster decrease in loss indicates... Increase the speed of convergence; The gradient clipping step includes: proportionally clipping parameters with a gradient norm exceeding 10.0 using the following formula:

[0096] Where grad is the gradient and threshold is the threshold value.

[0097] When the loss function is less than the second preset threshold, an updated fault diagnosis model is obtained. This updated fault diagnosis model improves the accuracy of cross-vehicle fault diagnosis by 15%-20% compared with the initial model.

[0098] In some examples, the fault prediction module is used to input the processed current vehicle fault diagnosis-related data into the updated fault diagnosis model and output a fault prediction result containing fault type, fault confidence level, and hazard level. When the product of the fault confidence level and hazard level in the fault prediction result is greater than a dynamic threshold, a warning message is generated and a warning is issued. The dynamic threshold is dynamically adjusted according to the current vehicle model, service life, and / or aging degree.

[0099] It should be noted that the fault data can be user operating condition related datasets, fault diagnosis related datasets, or a combination of both. The updated fault diagnosis model analyzes and calculates based on this input fault data to derive fault prediction results. If the fault prediction results meet preset warning conditions, the fault prediction module will issue a warning.

[0100] The fault prediction module inputs fault data into the updated fault diagnosis model and outputs prediction results including fault type, fault confidence (0-100%), and hazard level (1-5). When "fault confidence × hazard level" is greater than a dynamic threshold (e.g., the threshold is 300 for hazard level 5 and 150 for hazard level 2), an early warning is triggered.

[0101] In some examples, the data transmission module 501 is used to transmit the operational data to the cloud server and the user operating condition learning module. For example, the data transmission module 501 transmits the collected operational data to the cloud server 201 and the user operating condition learning module 301 via wireless communication methods such as 4G / 5G, Wi-Fi, and Bluetooth. In some embodiments, the data transmission module 501 transmits the collected operational data to the cloud server 201 using 5G slicing technology (which improves transmission stability and real-time performance compared to 4G / Wi-Fi), with key fault characteristic data being transmitted first (transmission priority is 3-5 times higher than regular data).

[0102] The data transmission module, through 5G slicing technology, transmits the collected operational data to the cloud server using a hybrid technology adapted to the high-safety scenarios of autonomous driving. This technology combines Quantum Key Distribution (QKD) with AES-256 (Advanced Encryption Standard with 256-bit key) symmetric encryption algorithm, quantum random number generator technology, and 5G hard slicing isolation technology. It also employs a quantum-enhanced hybrid data transmission encryption algorithm, whose encryption algorithm layer includes: Key negotiation layer: QKD quantum key distribution layer, which enables wireless transmission over wide area networks and wired negotiation over local area networks, and can achieve quantum security in hybrid networks.

[0103] Data encryption layer: AES-256 stream encryption layer, which can be adapted to the massive data transmission of sensors in autonomous driving. Security enhancement layer: Quantum random number generation key layer, with an entropy value of ≥256 bits in the generation process, used to resist quantum computing cracking.

[0104] The performance of the quantum-enhanced hybrid encryption scheme and the traditional AES encryption scheme were tested under high-speed (80km / h) and complex electromagnetic environments: Table 4: Experimental Verification: Data Transmission Encryption Performance Test

[0105] Note: This application's solution reduces transmission latency by 66% through 5G hard slicing isolation and quantum key distribution. It achieves near-perfect data transmission success rate in high-speed mobile scenarios and can resist quantum computing attacks.

[0106] In some examples, issuing an early warning when the fault prediction result meets preset early warning conditions includes: If the product of the fault confidence level and the hazard level in the fault prediction result is greater than a third preset threshold, an early warning message is generated and an early warning is issued. The third preset threshold can be a fixed value or a dynamic threshold.

[0107] For example, when the fault prediction result exceeds the third preset threshold, it indicates that the probability of equipment or system failure has reached or exceeded the pre-set acceptable risk level. At this time, the system will automatically generate a warning message. The warning message usually includes the fault type (such as "battery cell aging"), the expected fault time (such as "may occur within 7 days"), affected components, and corresponding response suggestions (such as "it is recommended to replace the battery cells first"). The dynamic threshold is dynamically adjusted according to the vehicle's age (the threshold is reduced by 20% for vehicles used for more than 3 years to adapt to the sensitivity changes of aging components) so that relevant personnel can quickly understand the situation.

[0108] In some examples, the notification feedback module 701 sends the warning information through multiple channels, including pushing real-time data visualization charts through an application installed on the user terminal, and sending content containing the warning information to the user terminal via SMS and / or email; wherein, when pushing through the application, the processing results fed back by the user through the application are transmitted back to the cloud for model optimization; the visualization chart includes a battery voltage fluctuation curve, and the SMS contains a simplified version of the warning content; the human-computer interaction interface module is used to display a battery health trend chart, a fault risk heat map, and system settings on the visualized human-computer interaction interface, wherein the battery health trend chart includes the battery health change trend predicted for the next 30 days based on an improved model, and the fault risk heat map marks high-risk components.

[0109] For example, the notification feedback module 701 sends warning information through multiple channels. App push notifications include real-time data visualization charts (such as battery voltage fluctuation curves), and SMS messages contain a concise warning message (limited to 70 characters). Users can provide feedback on the processing results through the App, and the system sends the feedback data back to the cloud for model optimization. The human-computer interaction interface module 601 allows users to view battery health trend charts (predicted trends for the next 30 days based on the improved model), fault risk heatmaps (marking high-risk components), and system settings through a visual interface.

[0110] like Figure 2 As shown, this application proposes an intelligent diagnostic and early warning method for electric vehicle power supply faults, the method comprising: Step S101: Collect the current vehicle's operating data, obtain the historical operating data of the current vehicle and other similar vehicles stored in the cloud server, preprocess the operating data to obtain processed operating data, and analyze and compare the historical operating data and the processed operating data to obtain analysis results; Step S102: Based on the analysis results, obtain the user operating condition related dataset and the fault diagnosis related dataset, and update the data on the cloud server. The user operating condition related dataset includes the user operating condition learning model and the processed user operating condition related data. The fault diagnosis related dataset includes the fault diagnosis model and the processed fault diagnosis related data. Step S103: Train the current vehicle's user condition learning model based on the spatiotemporal attention hybrid neural network algorithm, the user condition related dataset, and the processed current vehicle's user condition related operating data to obtain an updated current vehicle's user condition learning model. Step S104: Based on the updated user working condition learning model and the fault diagnosis model, construct a transfer learning model using a multi-source domain adaptive transfer learning algorithm; Step S105: Train the transfer learning model using the hierarchical adaptive backpropagation algorithm and the fault diagnosis related dataset, calculate the gradient and update the model parameters to minimize the loss function, and obtain the updated fault diagnosis model when the loss function is less than the second preset threshold. Step S106: Input the processed current vehicle fault diagnosis related data into the updated fault diagnosis model to obtain fault prediction results containing fault confidence and hazard level; Step S107: If the product of the fault confidence level and the hazard level in the fault prediction result is greater than the dynamic threshold, generate early warning information and issue an early warning. Step S108: Based on the warning information, notify and provide feedback to the user; Step S109: Display the warning information on the human-computer interaction interface, and display the vehicle-related parameters of the current vehicle.

[0111] The effects of the above method when applied to the aforementioned system can be found in the description of the aforementioned system embodiments, and will not be repeated here.

[0112] Based on the same inventive concept, embodiments of this specification provide a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the above-described intelligent diagnosis and early warning method for electric vehicle power supply faults.

[0113] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-readable program code.

[0114] In the several embodiments provided in this application, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between devices or units, and may be electrical, mechanical, or other forms.

[0115] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0116] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0117] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of the claims of the present invention.

Claims

1. An intelligent diagnostic and early warning system for electric vehicle power supply faults, characterized in that, include: The data acquisition module is used to collect the current vehicle's operating data. The operating data includes user-related operating data of the current vehicle. If the vehicle has malfunctioned, the operating data also includes fault diagnosis-related data. The data transmission module is used to transmit the current vehicle operation data to the cloud server and the user operating condition learning module. The data transmission module transmits the collected operation data to the cloud server through 5G slicing technology. During the transmission process, key fault feature data is transmitted first, and its transmission priority is 3-5 times higher than that of conventional data transmission. The data transmission module uses 5G slicing technology to transmit the collected running data to the cloud server. This process employs a hybrid technology adapted to the high-safety scenario of autonomous driving. The hybrid technology includes quantum key distribution technology, AES-256 encryption symmetric algorithm technology, quantum random number generator technology, and 5G hard slicing isolation technology. It also applies a quantum-enhanced hybrid data transmission encryption algorithm, whose encryption algorithm layer includes: a key negotiation layer, a data encryption layer, and a security enhancement layer. The key negotiation layer includes a QKD quantum key distribution layer to enable wireless transmission over wide area networks and wired negotiation over local area networks, and to achieve quantum security in hybrid networks. The data encryption layer includes an AES-256 stream encryption layer to adapt to the massive data transmission of sensors in autonomous driving. The security enhancement layer includes a quantum random number generation key layer, with an entropy value of ≥256 bits in the generation process, used to resist quantum computing cracking; The data analysis module is used to acquire the current operating data of the current vehicle and the historical operating datasets of other similar vehicles stored in the cloud server. The historical operating datasets of other similar vehicles include relevant operating condition datasets and fault diagnosis datasets of other similar vehicles. The data analysis module is also used to preprocess the current operating data of the current vehicle to obtain processed current operating data, analyze and compare the historical operating datasets of other similar vehicles in the cloud server with the processed current operating data to obtain analysis results, obtain the current vehicle's user operating condition related dataset and fault diagnosis related dataset based on the analysis results, and update the relevant operating condition datasets and fault diagnosis datasets of other similar vehicles stored in the cloud server in real time. Simultaneously, the current vehicle's user operating condition related dataset is sent to the user operating condition learning module, and the current vehicle's fault diagnosis related dataset is sent to the fault prediction and warning module. The user operating condition related dataset includes the current vehicle's user operating condition learning model and the processed user operating condition related operating data, and the fault diagnosis related dataset includes the current vehicle's fault diagnosis model and the processed fault diagnosis related data. The user operating condition learning module is used to train the user operating condition learning model of the current vehicle based on the spatiotemporal attention hybrid neural network algorithm, the user operating condition related dataset of the current vehicle, and the processed operating data of the current vehicle, so as to obtain an updated user operating condition learning model of the current vehicle; and to transmit the updated user operating condition learning model of the current vehicle and the processed user operating condition related operating data of the current vehicle to the data analysis module for updating the cloud server database. If a data communication transmission failure prevents the data analysis module and the cloud server from preprocessing the current operating data of the current vehicle, then the current operating data and fault diagnosis-related data of the current vehicle are preprocessed based on the user operating condition learning module of the vehicle to obtain the processed user operating condition-related operating data and fault diagnosis-related data of the current vehicle; and the most recent user operating condition learning model of the current vehicle transmitted by the data analysis module before the data communication transmission failure is used for model training until the data communication transmission failure is successfully eliminated. The fault prediction and warning module is used to train the updated user operating condition learning model and the fault diagnosis model based on the fault diagnosis-related data using a transfer learning algorithm, update the parameters of the fault diagnosis model to obtain an updated current vehicle fault diagnosis model, input the processed current vehicle fault diagnosis-related data into the updated current vehicle fault diagnosis model to obtain a fault prediction result, and issue a warning if the fault prediction result meets the preset warning conditions; and send the updated current vehicle fault diagnosis model, the processed current vehicle fault diagnosis-related data, and the fault prediction result back to the data analysis module for updating the cloud server database; If a data communication transmission failure prevents the fault prediction and warning module from obtaining the current vehicle's fault diagnosis-related dataset from the data analysis module, the fault prediction and warning module will obtain the preprocessed current vehicle fault diagnosis-related data from the user working condition learning module, and train the model using the most recent fault diagnosis model of the current vehicle from the data analysis module before the data communication transmission failure, until the data communication transmission failure is successfully eliminated. The notification feedback module is used to provide notification feedback to users based on early warning information; The human-machine interface module is used to display the warning information on the human-machine interface and to display the vehicle-related parameters of the current vehicle. The cloud server is also used to store the updated similar vehicle-related operating condition dataset and fault diagnosis dataset; The cloud server uses a composite security storage method based on data layered encryption, distributed storage, end-to-end auditing, and dynamic access control mechanisms to store data, specifically including: The data layered encryption uses the AES-256-GCM algorithm to encrypt the original dataset, encapsulates the AES dynamic key through the SM2 elliptic curve algorithm, and combines a multi-attribute layered attribute-based encryption mechanism, dividing user attributes into three layers: basic attributes, data attributes, and environmental attributes. A quantum random number generator is superimposed to provide an encryption entropy source. The basic attributes include identity and role, the data attributes include data sensitivity level, and the environmental attributes include access location and device security level. The distributed storage uses a distributed hash table to divide the encrypted data into data blocks, and uses erasure coding technology to encode each data block into a data fragment, which is then distributed and stored on multiple remote cloud nodes. The full-process audit includes audit traceability, data integrity verification through a provable data holding protocol, and recording of full data operation logs based on blockchain technology. The dynamic access control constructs a four-dimensional permission model based on user identity, role, data sensitivity, and operation time window, and realizes real-time verification and traceability of access requests through API security gateway combined with blockchain identity authentication.

2. The system as described in claim 1, characterized in that, The data analysis module is deployed on the cloud server. The data analysis module is used to clean and process the current operating data of the current vehicle to obtain initial operating data; and to extract features from the initial operating data to obtain processed current operating data. The data analysis module is also used to analyze and compare the historical operating dataset and the processed current operating data based on the current vehicle model, years of use and / or degree of aging, and obtain analysis results.

3. The system as described in claim 1, characterized in that, The user work condition learning module includes: The data receiving module is used to acquire the user's working condition related dataset and the current vehicle's operating data; The data encoding module processes the user operating condition related dataset and the current vehicle's operating data based on a dynamic hybrid standardization method to obtain the initial user operating condition related dataset and the processed user operating condition related operating data of the current vehicle. The data encoding module processes the user working condition related dataset based on a dynamic hybrid standardization method, including a time-series segmentation step, a distribution adaptation step, and a periodic enhancement step. The time-series segmentation step includes: dividing each data segment according to a preset time window, and calculating the local statistics of each data segment; The distribution adaptation step includes: determining the data distribution type using the Shapiro-Wilk test, wherein if the data distribution type is a normal distribution, improved Z-Score standardization is performed using the following formula: in, This is the median of the data segment. This represents the absolute deviation of the median of the data segment. This is a minimum value, used to reduce the impact of outliers; If the data distribution type is skewed, logarithmic transformation and max-min normalization are used for processing; The periodic enhancement step includes: performing a Fourier transform on the periodic data, extracting the periodic components and fusing them into the original data, wherein the periodic data includes daily mileage; Based on the processing of the time-series segmentation step, the distribution adaptation step, and the periodic enhancement step, an initial user working condition related dataset is obtained. The model training output module is used to train the user condition learning model based on the spatiotemporal attention hybrid neural network algorithm, the initial user condition related dataset, and the processed user condition related operating data of the current vehicle. When the accuracy of the validation set of the user condition learning model improves less than a first preset threshold for five consecutive rounds, the learning rate of the user condition learning model is automatically adjusted and training continues until the user condition learning model converges, thus obtaining the updated user condition learning model.

4. The system as described in claim 3, characterized in that, The model training output module uses a spatiotemporal attention hybrid neural network algorithm to train the user working condition learning model. The spatiotemporal attention hybrid neural network structure used includes a bottom convolutional layer, a middle bidirectional long short-term memory network layer, and a top dual-channel attention layer. The bottom convolutional layer includes three parallel convolutional kernels with sizes of 3, 5, and 7 respectively, which are used to extract local spatial features at different time scales. The local spatial features include instantaneous power supply voltage fluctuations and short-term power supply current trends. The bottom convolutional layer is a 1D-CNN. The middle-layer bidirectional long short-term memory network is used to capture long-term dependencies and dynamically adjust weights through a gating mechanism using the hidden state output at each time step. The long-term dependencies include the battery temperature change pattern over a continuous 30-minute period. The top-level dual-channel attention layer includes a spatial attention layer and a temporal attention layer; The spatial attention layer calculates sensor weights through a fully connected layer activated by a Sigmoid function. The sensor weights include battery sensor weights and light sensor weights. The battery sensor weight is dynamically higher than the light sensor weight, and the weight value ranges from 0.8 to 0.

95. The temporal attention layer is used to weight the hidden states of the output of the Long Short-Term Memory (LSTM) network layer; its weighting process features include that the temporal weight of the charging phase is 1.5-2 times that of the idling phase. The loss function for the spatiotemporal attention hybrid neural network algorithm is obtained by weighting the focus loss function and the mean squared error using the following formula: Where Focal Loss is the focus loss function, and MSE is the mean squared error. This is the balance coefficient, with a value of 0.

7. During training, a dynamic early stopping mechanism is adopted. When the accuracy of the validation set increases by less than a first preset threshold for five consecutive batches, the learning rate is automatically multiplied by 0.5 until the model converges, thus obtaining the updated user working condition learning model. The first preset threshold is 0.1%.

5. The system as described in claim 1, characterized in that, The fault prediction and warning module includes: The data source receiving module is used to acquire the fault diagnosis-related dataset and the updated user working condition learning model. The transfer learning module is used to construct a transfer learning model based on the updated user working condition learning model and the fault diagnosis model contained in the fault diagnosis related dataset. The model training module is used to train the transfer learning model using the backpropagation algorithm and the fault diagnosis related dataset. When the loss function of the fault diagnosis model is less than a second preset threshold, an updated fault diagnosis model is obtained. The fault prediction module is used to input the processed current vehicle fault diagnosis data into the updated fault diagnosis model to obtain the fault prediction result, and to issue an early warning when the fault prediction result meets the dynamic early warning conditions. The transfer learning module uses the fault diagnosis model and the updated user working condition model as source tasks to construct the transfer learning model. The construction process of the transfer learning is as follows: The convolutional layer of the source task is used as a feature extractor, and features are extracted using the following formula: in, For the target task characteristics, Features of the source task; Let the source task model be , The parameters of the source task model are used for feature fine-tuning using the following formula; in, For the target task model, These are the fine-tuned parameters. Update the values ​​for the parameters to be fine-tuned; Based on the gradient descent algorithm, the parameters are fine-tuned using the following formula: in, The loss function used to train the fault diagnosis model is... For learning rate, The gradient of the parameters of the source task model; Based on a weighted average method, the features of the source task and the target task are fused using the following formula: in, The characteristics after fusion The fusion coefficient; In joint transfer learning, the overall loss function can be expressed as a weighted combination of the target task loss and the source task loss using the following formula: in, For the overall loss function, The source task loss function, Let the target task loss function be... The weights are the source task loss function weights. Weights for the target task loss function; The target task generates the updated fault diagnosis model through the feature extractor and fine-tuning model of the source task; In the joint transfer learning, the overall loss function adopts a staged dynamic adjustment strategy, which includes an initial stage dynamic adjustment strategy, a mid-stage dynamic adjustment strategy, and a convergence stage dynamic adjustment strategy. The initial dynamic adjustment strategy corresponds to the first 20% of iterations, and the overall loss function for the initial stage is: The initial stage dynamic adjustment strategy focuses on source domain knowledge transfer. The dynamic adjustment strategy for the intermediate stage corresponds to the middle 20%~80% of the iterations, and the overall loss function corresponding to the intermediate stage is: The mid-term dynamic adjustment strategy focuses on target domain adaptation. The dynamic adjustment strategy for the convergence phase corresponds to the remaining iterations, and the overall loss function for the convergence phase is: The remaining iterations are the last 20% of iterations; in, For source domain loss, For the target domain loss, Consistency constraints for tasks; , Avoid knowledge conflicts.

6. The system as described in claim 5, characterized in that, If the transfer learning model is a multi-source domain adaptive transfer learning model, the transfer learning module constructs the multi-source domain adaptive transfer learning model based on the fault diagnosis model and the user working condition model as source tasks. Its construction steps include a multi-source domain selection step, a dynamic domain weight adjustment step, an adversarial feature alignment step, and a feature fusion step. The multi-source domain selection step includes: selecting the three source domains with the highest cosine similarity to the vehicle model features of the current vehicle from the cloud server, wherein the similarity calculation formula is: Where A is the target domain feature vector and B is the source domain feature vector. The three source domains with the highest similarity can include electric cars of the same brand and models with the same battery type. The dynamic domain weight adjustment step includes: adjusting the source domain weights using the following formula. Updated dynamically during the training process: in, The feature distribution distance between the source and target domains is updated every 100 batches of data processed. To and Orthogonal distribution distance; The adversarial feature alignment step includes: introducing a domain discriminator with a convolutional neural network structure, and aligning the feature distributions of the source and target domains by minimizing the following loss function: in, The domain classification probability output by the discriminator. Here, N represents the domain label, and N represents the number of iterations. The feature fusion step includes: fusing the features of the multi-source domain and the target domain by weighted average using the following formula: in, The characteristics after fusion For the i-th source domain feature, For target domain features, The source domain weights.

7. The system as described in claim 5, characterized in that, The model training module is used to train the transfer learning model using the backpropagation algorithm and the fault diagnosis-related data, calculate the gradient and update the model parameters, minimize the loss function, and obtain the updated fault diagnosis model when the loss function is less than a second preset threshold. During the training process, the target task parameters are calculated using the following formula: in, These are the target task parameters for the next batch. The target task parameters for the current batch. For learning rate, This represents the gradient of the corrected overall loss function with respect to the target task parameters; If the backpropagation algorithm is a hierarchical adaptive backpropagation algorithm, when the model training module trains the transfer learning model using the hierarchical adaptive backpropagation algorithm, its training steps include a hierarchical learning rate optimization step, a momentum dynamic adjustment step, and a gradient clipping step. When employing the hierarchical adaptive backpropagation algorithm, the formula for calculating the target task parameters is updated as follows: in, For tiered learning rates, Momentum factor These are the target task parameters from the previous batch; The hierarchical learning rate optimization step includes: setting the learning rate of the long short-term memory network layer to 0.001, the learning rate of the attention layer to 0.0005, and the learning rate of the discriminator to 0.

002. The momentum dynamic adjustment step includes: adjusting the momentum factor using the following formula strategy. Adaptive adjustment as loss changes: in, This represents the difference in loss between the current batch and the previous batch; a faster decrease in loss indicates... Increase the speed of convergence; The gradient clipping step includes: proportionally clipping parameters with a gradient norm exceeding 10.0 using the following formula: Where grad is the gradient and threshold is the threshold value.

8. The system as described in claim 1, characterized in that, The fault prediction module is used to input the processed current vehicle fault diagnosis data into the updated fault diagnosis model and output a fault prediction result containing fault type, fault confidence level and hazard level. When the product of the fault confidence level and hazard level in the fault prediction result is greater than a dynamic threshold, a warning message is generated and a warning is issued. The dynamic threshold is dynamically adjusted according to the current vehicle model, service life and / or aging degree.

9. The system as described in claim 1, characterized in that, The notification feedback module sends the warning information through multiple channels, including pushing real-time data visualization charts through an application installed on the user terminal, and sending content containing the warning information to the user terminal via SMS and / or email; wherein, when pushing through the application, the processing results fed back by the user through the application are sent back to the cloud for model optimization; the visualization chart includes a battery voltage fluctuation curve, and the SMS contains a simplified version of the warning content; The human-computer interaction interface module is used to display a battery health trend chart, a fault risk heat map, and system settings on the visualized human-computer interaction interface. The battery health trend chart includes the battery health change trend predicted by an improved model for the next 30 days, and the fault risk heat map marks high-risk components.

10. A method for intelligent diagnosis and early warning of power supply faults in electric vehicles, characterized in that, include: The system collects current vehicle operation data, obtains historical operation data of the current vehicle and other similar vehicles stored in the cloud server, preprocesses the operation data to obtain processed operation data, and analyzes and compares the historical operation data and the processed operation data to obtain analysis results. Based on the analysis results, user operating condition related datasets and fault diagnosis related datasets are obtained and the data on the cloud server is updated. The user operating condition related dataset includes a user operating condition learning model and processed user operating condition related data. The fault diagnosis related dataset includes a fault diagnosis model and processed fault diagnosis related data. The user condition learning model of the current vehicle is trained based on the spatiotemporal attention hybrid neural network algorithm, the user condition related dataset, and the processed user condition related operating data of the current vehicle to obtain the updated user condition learning model of the current vehicle. Based on the updated user working condition learning model and the fault diagnosis model, a transfer learning model is constructed using a multi-source domain adaptive transfer learning algorithm. The transfer learning model is trained using a hierarchical adaptive backpropagation algorithm and the fault diagnosis-related dataset. The gradient is calculated and the model parameters are updated to minimize the loss function. When the loss function is less than a second preset threshold, the updated fault diagnosis model is obtained. The processed current vehicle fault diagnosis data is input into the updated fault diagnosis model to obtain fault prediction results that include fault confidence and hazard level. If the product of the fault confidence level and the hazard level in the fault prediction result is greater than the dynamic threshold, an early warning message is generated and an early warning is issued. Based on the early warning information, users are notified and feedback is provided. The warning information is displayed on the human-computer interaction interface, along with the vehicle-related parameters of the current vehicle.

Images