Power battery early warning method, electric vehicle and storage medium

Abstract

The invention discloses a power battery early warning method, an electric automobile and a storage medium, and belongs to the technical field of battery safety. The method comprises the steps that low-frequency operation data, uploaded by multiple electric automobiles, of power batteries is acquired; constructing a neural network model based on an encoder and decoder structure, training the neural network model by taking the low-frequency operation data as input and adopting an unsupervised learning mode, and performing model optimization by minimizing a mean square error of the low-frequency operation data input and decoder reconstruction output to obtain a pre-training model; the encoder of the pre-training model has the capability of extracting the general characteristics of the normal working condition of the power battery. According to the method, the high data utilization rate is achieved through unsupervised pre-training, the extracted cross-model universal features have robustness and mobility, and meanwhile, the modular architecture design provides an extensible technical basis for subsequent supervised learning and vehicle end lightweight deployment.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of battery safety, in particular to a power battery early warning method, an electric vehicle and a storage medium. BACKGROUND

[0002] The prior art mainly includes pure vehicle end scheme and pure cloud end scheme: the pure vehicle end scheme (such as threshold method, equivalent circuit model or lightweight machine learning) can only process shallow features and relies on isolated data due to limited vehicle-mounted computing power, resulting in insufficient early abnormality detection capability; the pure cloud end scheme can train complex models using multi-vehicle data, but is limited by communication bandwidth, and the vehicle end must be reduced to seconds, losing high-frequency features and reducing the accuracy of early warning, and the cloud end result return has a delay. In addition, the existing preliminary vehicle-cloud collaborative scheme (such as CN115384351A patent document) only realizes simple interaction of vehicle end and cloud end data, resulting in that the vehicle end is still limited by computing power, the cloud end is still limited by low-frequency data, and lacks closed-loop optimization capability, forming a mechanical superposition of "vehicle end rule detection + cloud end offline analysis", which cannot balance real-time and generalization. SUMMARY

[0003] The application provides a power battery early warning method, an electric vehicle and a storage medium, aiming to establish a general feature representation capability of a power battery under normal working conditions through cloud end unsupervised pre-training, and solve the problem of insufficient feature extraction and insufficient model generalization performance caused by the cloud end only being able to obtain low-frequency data in the prior art.

[0004] In a first aspect, a power battery early warning method is provided, comprising: obtaining low-frequency running data of a power battery uploaded by a plurality of electric vehicles; constructing a neural network model based on an encoder and a decoder structure, taking the low-frequency running data as input, training the neural network model in an unsupervised learning manner, optimizing the model by minimizing the mean square error of the low-frequency running data input and the decoder reconstruction output, obtaining a pre-training model, and the encoder of the pre-training model has the capability of extracting general features of the power battery under normal working conditions.

[0005] In a second aspect, a power battery early warning method is also provided, comprising: deploying the encoder model and the early warning model processed by the power battery early warning method according to any one of the first aspect to the vehicle end chip; obtaining high-frequency running data of the power battery, and inputting the high-frequency running data into the encoder model to generate a high-frequency feature vector; inputting the high-frequency feature vector and the obtained low-frequency running data of the power battery into the early warning model, and outputting an abnormal probability of the power battery through the early warning model.

[0006] In a third aspect, the application provides an electric vehicle, comprising: a vehicle-mounted BMS system configured to acquire high-frequency operation data and low-frequency operation data of the power battery; a vehicle-end chip loaded with a lightweight encoder model and a warning model; a vehicle networking communication module configured to upload the high-frequency feature vector and the associated low-frequency operation data to a cloud server and receive an optimized warning model issued by the cloud server; a memory configured to store the encoder model, the warning model, and a computer program; a processor connected to the vehicle-mounted BMS system, the vehicle networking communication module, and the memory, wherein the computer program is executed by the processor to implement the power battery warning method according to any one of the second aspect.

[0007] In a fourth aspect, the application provides a computer readable storage medium having a computer program stored thereon, wherein the computer program is loaded by a processor to execute the steps of the power battery warning method according to any one of the second aspect.

[0008] Advantages: Based on the cloud unsupervised pre-training method, the general feature representation ability of the power battery is established and the existing problems are solved through the following technical path: first, the low-frequency operation data uploaded by multiple electric vehicles is aggregated to construct a deep neural network with encoder and decoder structure; second, the unsupervised learning method is adopted, the self-reconstruction of the input data is taken as the target, and the mean square error between the original low-frequency data and the decoder output is minimized to force the encoder to extract the essential features representing the normal working condition of the battery in the data compression process. The application solves two technical bottlenecks: first, through the unsupervised fusion training of massive vehicle data, the encoder can learn the general feature expression across models and battery systems, overcoming the generalization defect caused by insufficient labeled data in traditional supervised learning; second, the feature extraction prior knowledge of the encoder provides an optimization starting point for subsequent high-frequency data fine-tuning, effectively alleviating the feature distortion problem caused by pure low-frequency data training.

[0009] The application has the advantages of: the application realizes high data utilization through unsupervised pre-training, extracts cross-model general features with robustness and portability, and provides an extensible technical foundation for subsequent supervised learning and vehicle-end lightweight deployment through modular architecture design. BRIEF DESCRIPTION OF DRAWINGS

[0010] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings needed in the embodiment description. Obviously, the drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained from these drawings without creative labor.

[0011] Figure 1 is one of the flow diagrams of the power battery early warning method provided by the embodiments of the present application; Figure 2 is a flow diagram of the model construction and training process provided by the embodiments of the present application; Figure 3 is a schematic diagram of the model architecture provided by the embodiments of the present application; Figure 4 is a schematic diagram of the model update provided by the embodiments of the present application; Figure 5 is another flow diagram of the power battery early warning method provided by the embodiments of the present application; Figure 6 is a third flow diagram of the power battery early warning method provided by the embodiments of the present application. DETAILED DESCRIPTION

[0012] The technical solutions in the embodiments of the present application will be described clearly and completely in the following with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0013] In the description of the present application, it should be understood that the terms "first", "second" are used only for the purpose of description, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features limited by "first", "second" can explicitly or implicitly include one or more of the features. In the description of the present application, the meaning of "multiple" is two or more, unless otherwise specifically limited.

[0014] "A and / or B" includes the following three combinations: only A, only B, and the combination of A and B.

[0015] The use of "applicable to" or "configured to" in the present application means open and inclusive language, which does not exclude devices applicable to or configured to perform additional tasks or steps. In addition, the use of "based on" means open and inclusive, because the process, step, calculation or other action "based on" one or more stated conditions or values can be based on additional conditions or values beyond those stated in practice.

[0016] In this application, the word "exemplary" is used to mean "serving as an example, instance, or illustration." Any implementation described as "exemplary" in this application is not necessarily to be construed as preferred or advantageous over other implementations. The following description is presented to enable any person skilled in the art to make and use the application. In the following description, for purposes of explanation, numerous details are set forth. It should be appreciated, however, that the present application can be practiced in a variety of contexts without some of these specific details. In other instances, well known structures and processes are not elaborated in order not to obscure the description of the application. Thus, the present application is not intended to be limited to the embodiments shown, but is to be accorded the widest scope consistent with the principles and features presented herein.

[0017] The present application is aimed at the problems existing in the existing power battery anomaly detection technology, that is, the early warning lag and the lack of generalization of the vehicle end scheme due to limited computing power and isolated data, the loss of high-frequency features of the cloud end scheme due to data frequency reduction, and the inability of the existing vehicle-cloud collaborative scheme to achieve deep collaboration by mechanically superimposing the functions of the two. A power battery anomaly detection framework based on "high-frequency feature compression-vehicle-cloud dynamic collaboration" is proposed. By establishing a model optimization mechanism combining cloud unsupervised pre-training and vehicle high-frequency fine-tuning, a closed-loop system of vehicle real-time early warning and cloud continuous evolution is constructed, effective fusion of high and low frequency data and dynamic allocation of computing resources are realized, and finally accurate identification and real-time early warning of early battery anomalies are achieved, while ensuring the long-term applicability of the system in cross-model and cross-working condition scenarios.

[0018] In one aspect, the present embodiment provides a power battery early warning method, as shown in Figure 1 The method comprises the following steps: S101, acquiring low-frequency operation data of power batteries uploaded by multiple electric vehicles.

[0019] It can be understood that the collected operating data of the power battery includes the following characteristics: total voltage of the power battery pack, individual cell voltage of the power battery pack, average cell voltage of the power battery pack, highest cell voltage of the power battery pack, lowest cell voltage of the power battery pack, total current of the power battery pack, temperature of each battery probe, average temperature of the power battery pack, highest probe temperature of the power battery pack, lowest probe temperature of the power battery pack, and the like. The low-frequency operating data collected by the cloud generally is data with a frequency of seconds, i.e., 1-10 seconds / time, for example, 3 seconds, 5 seconds, or 10 seconds per frame of data. Obtaining the low-frequency operating data uploaded by multiple electric vehicles is the basis for constructing a general representation of the battery state, which functions in the following aspects: first, the low-frequency data sampled at a second level can retain the key state characteristics (such as voltage / temperature distribution and change trend) of the battery operation under the premise of meeting the bandwidth limitation of vehicle-cloud communication, providing benchmark information for the model to judge the battery health state; second, by aggregating the low-frequency data of a large number of vehicles, the normal operating mode under different vehicle models, battery types, and use scenarios can be covered, enabling the model to learn the common feature expression across dimensions; finally, the low-frequency data, as a stable data source that can be obtained by the cloud, has the characteristic of continuous accumulation, which provides data scale guarantee for unsupervised pre-training of the model, avoiding the overfitting risk caused by relying on a small amount of labeled data. This low-frequency big data strategy achieves an optimal balance between communication cost and feature effectiveness, laying a reliable foundation for subsequent high-frequency feature fusion.

[0020] In S102, a neural network model based on an encoder and a decoder structure is constructed, the low-frequency operating data is taken as input, the neural network model is trained in an unsupervised learning manner, model optimization is performed by minimizing the mean square error between the low-frequency operating data input and the decoder reconstruction output, a pre-trained model is obtained, and the encoder of the pre-trained model has the ability to extract general features of the power battery under normal working conditions.

[0021] It can be understood that by constructing a neural network model based on an encoder and a decoder structure (such as a Transformer or a CNN-LSTM autoencoder), the low-frequency operation data uploaded on multiple electric vehicles is taken as input, and the model is trained in an unsupervised learning manner. In a specific implementation, the model compresses the input low-frequency data into a low-dimensional feature representation through the encoder, and then reconstructs the original input data based on the feature through the decoder. In the training process, the optimization objective is to minimize the mean square error (MSE) between the input data and the reconstructed output. This mechanism forces the encoder to capture the essential feature patterns of the normal operation of the battery in the data compression process, so that it has the ability to extract normal working condition general features of the power battery (such as common state representation across vehicle models and battery systems). The role of this step is: 1) Through unsupervised learning using massive unlabeled data, it breaks through the dependence of supervised learning on fault labeled data; 2) The general feature extraction capability established provides a high-quality feature space for subsequent high-frequency data fusion and anomaly detection; 3) The normal working condition reference features learned by the encoder can be used as a reference for identifying abnormal states, solving the false alarm problem caused by the lack of normal state reference in traditional methods. The feature extraction capability formed by this pre-training process is an important foundation for subsequent efficient vehicle cloud collaborative early warning.

[0022] It should be noted that the autoencoder structure used in the present application can not be limited, as long as it meets the conditions of being able to extract features and reconstruct data and being able to process time series data.

[0023] In some embodiments, the method further includes: obtaining high-frequency operation data of the power battery; taking the high-frequency operation data as input, freezing the decoder parameters of the pre-trained model, only optimizing the encoder parameters, fine-tuning the model by minimizing the error between the high-frequency operation data input and the decoder reconstruction output, and obtaining an encoder model adapted to high-frequency feature extraction.

[0024] It can be understood that the pre-trained model is fine-tuned by high-frequency data to enable the encoder to capture the high-frequency transient characteristics of the power battery. The specific process is as follows: a special data acquisition device is deployed on an experimental vehicle or a real vehicle to collect a small amount of operating parameters (such as total voltage, single cell voltage, total current, temperature, etc.) of the power battery at a high frequency of 10-100 Hz. The high-frequency data (millisecond-level sampling frequency) and the low-frequency data in the cloud (second-level sampling frequency) are complementary, and the high-frequency data contain more transient details (such as voltage fluctuation, temperature gradient change, etc.). It should be noted that the high-frequency data are real-time data collected by the vehicle-side BMS (Battery Management System, battery management system). The low-frequency data are data uploaded by the vehicle-side through the communication module to the cloud for long-term storage. Due to the limitation of communication bandwidth, the data is reduced to second-level. The high-frequency feature is the feature data compressed and reduced by the encoder model. For example, the high-frequency operating data is input into the encoder model in the form of a time window, and after compression and reduction, a low-dimensional, non-time sequence data is formed, which greatly reduces the data amount.

[0025] The encoder and decoder model architecture trained in step S102, such as the CNN-LSTM (Convolutional Neural Network-Long Short Term Memory Network) autoencoder, maintains its learned low-frequency general feature extraction capability. The decoder parameters are frozen (i.e., only as a fixed target for high-frequency data reconstruction), and only the encoder parameters are optimized, forcing the encoder to learn the compression representation ability of high-frequency data in addition to preserving low-frequency general features. By minimizing the mean square error (MSE) of the high-frequency input data and the decoder reconstruction output, the encoder is driven to extract the key dynamic features of the high-frequency data (such as transient voltage drop, temperature jump, etc.). The fine-tuned encoder has dual-modal feature extraction capability: it retains the global state representation of low-frequency data (such as battery health baseline), and adds the detail capturing ability of high-frequency data (such as abnormal transient signals). The low-dimensional feature vector output by the encoder implicitly contains high-frequency information, providing more rich input features for the subsequent early warning model.

[0026] The role is: the low-frequency data in the cloud (second level) cannot capture the millisecond-level abnormal signal (such as the initial micro-voltage fluctuation of internal short circuit). Through high-frequency fine-tuning, the encoder learns to extract key transient features from high-frequency data, indirectly making up for the shortcomings of low-frequency data in the cloud. The high-frequency data at the vehicle end is compressed by the encoder to generate low-dimensional features (rather than raw data), which not only reduces the bandwidth pressure of vehicle-cloud communication, but also ensures that the cloud model can use high-frequency information. For example, the vehicle end uploads a 10Hz data feature vector (such as 128 dimensions), rather than raw millisecond-level time series data (such as 1000 data points / second). The features extracted by the fine-tuned encoder contain high-frequency details, allowing the downstream deep neural network classifier to identify early weak abnormalities. For example, the initial stage of the battery internal short circuit may only show millisecond-level slight fluctuations in the voltage of a single cell, and the high-frequency adaptive encoder can capture such features. By freezing the decoder and limiting the fine-tuning range of the encoder, overfitting of high-frequency data is avoided, ensuring that the model can still process low-frequency input, such as low-frequency data in the cloud. This partial fine-tuning strategy balances specificity and universality, with specificity reflected in high-frequency data adaptation capabilities and universality reflected in wide applicability across data frequencies.

[0027] In some embodiments, the method further comprises: obtaining historical failure data of the power battery and the adaptive high-frequency feature extraction encoder model; inputting the historical failure data into the encoder model to generate a failure feature vector; and using the low-frequency running data and the failure feature vector as dual input sources to jointly train a supervised classifier with failure types as labels to obtain a pre-warning model that fuses high and low frequency features.

[0028] As can be understood, by obtaining the historical failure data of the power battery (including labeled internal short circuit, self-discharge and other abnormal types) and the fine-tuned encoder model, the historical failure data is first input into the encoder model to generate a failure feature vector (which implicitly contains state details obtained through high-frequency fine-tuning). Subsequently, a dual-channel input architecture is adopted: the original low-frequency running data and the failure feature vector output by the encoder model are jointly input into a supervised classifier. Through joint training with failure types as supervision signals, the model learns two key pieces of information: 1) global state representation provided by low-frequency data (such as voltage / temperature trends); 2) high-frequency details contained in the encoded features (such as transient fluctuation patterns). The role is: 1) to realize knowledge fusion of high and low frequency data, enabling the pre-warning model to have both macro state judgment and micro abnormality capture capabilities; 2) to integrate high-frequency information into the classification process in the form of a low-dimensional vector through feature compression by the encoder, solving the computational burden of directly processing high-frequency data; 3) to lay the foundation for subsequent vehicle end deployment, as the combination of encoded features and low-frequency data adapts to the vehicle-cloud collaboration scenario (vehicle end supplements high-frequency features, and cloud provides low-frequency benchmarks). The final pre-warning model significantly improves the detection sensitivity of early weak abnormalities (such as millisecond-level voltage disturbances in the initial stage of internal short circuit).

[0029] It should be noted that the present application only limits the design of double input sources, and does not limit the model itself, as long as it can accept double input sources and can perform supervised model training for fault labels.

[0030] In a specific example, refer to Figure 2 , Figure 2 is a flowchart of the model construction and training process provided by the embodiment of the present application, and a power battery early warning method, the model training stage includes the following steps: S201, pre-training the auto-encoder model to obtain a pre-trained model.

[0031] Based on the low-frequency operation data (such as second-level sampling) of the power battery in the sea of clouds, the auto-encoder model (such as the Transformer auto-encoder) is trained through unsupervised learning, so that the encoder has the ability to extract general features of the battery under normal working conditions. The pre-trained model includes an encoder and a decoder.

[0032] S202: Fine-tuning the pre-trained model to obtain an encoder model adapted for high-frequency feature extraction.

[0033] A small amount of high-frequency operation data of the power battery at the vehicle end is used to fine-tune the pre-trained model of step S201 through unsupervised learning, freeze the decoder parameters, and only optimize the encoder, so that it is adapted for high-frequency feature extraction, and the fine-tuned encoder model is obtained.

[0034] S203: Supervised fine-tuning and classifier joint training of the encoder model to obtain a warning model.

[0035] The historical fault data is input into the encoder model of step S202 to generate a fault feature vector, and the original low-frequency operation data is used as a double input source to jointly train a supervised classifier, and a warning model is obtained. The warning model includes an encoder and a classifier.

[0036] In another specific example, refer to Figure 3 , Figure 3 is a schematic diagram of the model architecture provided by the embodiment of the present application. The original data (including high-frequency operation data or low-frequency operation data, such as total voltage of power battery pack, single cell voltage, total current, probe temperature, etc.) is input into the encoder, and the encoder outputs the encoded feature data. The decoder reconstructs the data according to the encoded feature data, which is optimized through unsupervised training, so that the encoder has the ability to extract general features of the power battery under normal working conditions, to construct a pre-trained model, and to input high-frequency operation data, freeze the decoder parameters and only optimize the encoder parameters, to obtain an encoder model adapted for high-frequency feature extraction. The original data and the encoded feature data are used as double input sources to input the classifier, and abnormal detection and early warning are realized.

[0037] In some other embodiments, the method further comprises: receiving the encoded high-frequency feature vector and its associated low-frequency operation data uploaded by the vehicle end; in the case of power battery failure of the vehicle, using the operation data of the failure period as a sample to incrementally train the early warning model; performing lightweight processing on the early warning model optimized by the incremental training, and then issuing the early warning model to the vehicle end to complete the closed-loop update of the early warning model.

[0038] It can be understood that the cloud receives two types of key data periodically uploaded by the vehicle end, i.e., the high-frequency feature vector (millisecond-level data extracted by a neural network into a low-dimensional representation) compressed by the vehicle end encoder and the corresponding second-level low-frequency operation data (such as voltage / temperature average), and a data closed loop of vehicle-cloud cooperation is constructed. When the vehicle has a power battery failure (such as internal short circuit), the cloud marks the operation data before and after the failure period as a sample, and realizes dynamic optimization through the following mechanisms: 1) input the high-frequency feature vector of the failure sample and the original low-frequency data into the early warning model, incrementally train the model with the failure type as a supervision signal, and focus on optimizing the decision boundary of the classifier to make the model learn the mapping relationship between the failure and the high and low frequency features; 2) use knowledge distillation and parameter quantization technology to compress the optimized early warning model into a lightweight version executable by the vehicle end, and update it through the Internet of Vehicles. The role is to: through the continuous accumulation of failure samples, the model gradually covers more abnormal scenarios (such as battery failures at different aging stages); the cloud model is iterated by the real-time data of the vehicle end, and the optimization results of the cloud end benefit the vehicle end, forming a "data-model" closed loop; dynamically adapt to long-term factors such as battery performance degradation and environmental changes to avoid static model failure problems. Finally, the early warning model has the ability to continuously evolve, and maintains high detection sensitivity for a long time.

[0039] It should be noted that the trained early warning model is incrementally trained at this time, and the encoder does not participate in the training process of the cloud end. The encoded feature data is directly returned by the vehicle end and input into the deep learning classifier together with the original low-frequency operation data to train the model parameters. After continuous update training, the updated early warning model is converted into a lightweight version by pruning and quantization techniques, and is issued to the vehicle end for model update, thereby forming a closed-loop update.

[0040] As shown in Figure 4 , the method further comprises: receiving the encoded high-frequency feature vector and its associated low-frequency operation data uploaded by the vehicle end; in the case of power battery failure of the vehicle, using the operation data of the failure period as a sample to incrementally train the early warning model; performing lightweight processing on the early warning model optimized by the incremental training, and then issuing the early warning model to the vehicle end to complete the closed-loop update of the early warning model. Figure 4A power battery early warning model updating architecture of vehicle-cloud collaboration is demonstrated. Vehicle-side models (such as vehicle-side model 1, vehicle-side model 2, …, vehicle-side model N) continuously upload low-frequency operation data (such as total voltage of the power battery pack, each single cell voltage, and the like data with a sampling frequency of 1-10 seconds / time) and high-frequency feature data (generated by encoding high-frequency operation data by an encoder model adapted by the vehicle side to extract high-frequency features) to the cloud system. The cloud system performs incremental training according to fault labels (labeled when the vehicle has a power battery fault) and vehicle-side uploaded data. Then, the early warning model optimized by the incremental training is subjected to lightweight processing, and the updated model is regularly issued to the vehicle-side model to complete the closed-loop updating of the early warning model, realize the bidirectional interactive evolution of data and models between the vehicle and the cloud, ensure that the early warning model continuously adapts to the battery state change and the fault scene, and improve the early warning precision and system applicability.

[0041] In another aspect, the embodiment provides a power battery early warning method, as shown in the method, comprising: Figure 5 S501, deploying the trained and lightweight processed encoder model and early warning model to the vehicle-side chip.

[0042] In a specific example, it further comprises: performing channel pruning and weight quantization processing on the encoder model and the early warning model to adapt to the computing power constraint of the vehicle-side chip.

[0043] It can be understood that this step first performs lightweight processing on the trained encoder model and early warning model, including removing redundant connections in the neural network by using pruning technology, and reducing the model calculation precision requirement by using quantization technology. The processed encoder model and early warning model will adapt to the computing power limit of the vehicle-side chip, and then be deployed to the vehicle-side ECU (Electronic Control Unit), MCU (Microcontroller Unit), or HPC (High Performance Computer) chip. The role is to enable large models originally trained in the cloud to run in a resource-constrained vehicle-mounted environment, to provide a hardware foundation for subsequent real-time anomaly detection, and to ensure efficient execution of the model on the vehicle side.

[0044] S502, obtaining high-frequency operation data of the power battery, and inputting the high-frequency operation data to the encoder model to generate a high-frequency feature vector.

[0045] ​It can be understood that this step collects high-frequency operation data of the power battery in real time (such as parameters including single cell voltage, current, temperature, etc.), and inputs the high-frequency operation data into the deployed encoder model. The encoder model performs feature extraction and compression on the high-frequency data, and outputs a high-frequency feature vector. That is, the lightweight encoder model and the early warning model are deployed on the vehicle-side chip. As the electric vehicle runs, the early warning model of the power battery also runs in real time, continuously collects high-frequency operation data of the power battery, converts the high-frequency operation data into a high-frequency feature vector through the encoder model deployed on the vehicle side, and inputs the high-frequency feature vector and the low-frequency operation data into the early warning model for abnormality analysis. Once a risk is detected, the vehicle-side warning and early warning mechanism is triggered to ensure driving safety. The function is to realize data dimension reduction through the encoder model, on the one hand to retain key feature information in the high-frequency data, and on the other hand to greatly reduce the amount of data to be processed, to provide the possibility for subsequent real-time analysis, and also to lay the foundation for communication load optimization when uploading data to the cloud.

[0046] S503, inputting the high-frequency feature vector and the acquired low-frequency operation data of the power battery into the early warning model, and outputting an abnormal probability of the power battery through the early warning model.

[0047] In a specific example, further comprising: uploading the high-frequency feature vector and the associated low-frequency operation data to the cloud; and receiving an optimized early warning model issued by the cloud to replace the early warning model deployed locally on the vehicle side.

[0048] It can be understood that this step merges the high-frequency feature vector generated in S502 and the simultaneously acquired low-frequency operation data, and inputs the high-frequency feature vector and the low-frequency operation data into the early warning model for analysis. The early warning model comprehensively evaluates the high-frequency and low-frequency features, and outputs an abnormal probability value of the current power battery system. When the abnormal probability value exceeds a set threshold, the system immediately triggers a vehicle-side warning mechanism (such as a dashboard warning, a voice prompt, etc.). The function is to realize real-time monitoring and immediate response of the power battery state, and to ensure driving safety. At the same time, the vehicle-side system packages the high-frequency feature vector and the low-frequency operation data used for analysis and uploads the high-frequency feature vector and the low-frequency operation data to the cloud, to provide data support for subsequent model optimization, and to form a closed-loop system of vehicle-cloud cooperation.

[0049] In some embodiments, as shown in FIG. 6, Figure 6 A power battery early warning method, comprising: S601, constructing a pre-trained model on the cloud.

[0050] A large amount of low-frequency operation data of the power battery uploaded by the vehicle end is collected in the cloud server (such as the total voltage of the power battery pack and the voltage of each single battery with a sampling frequency of 1-10 seconds / time), and the cloud pre-training of the battery operation data reconstruction autoencoder model is performed based on the low-frequency operation data. The purpose is to train a large amount of low-frequency operation data, so that the encoder of the autoencoder model can preliminarily have the ability to extract general features of the power battery under normal working conditions, and obtain a pre-trained model.

[0051] S602, fine-tuning the pre-trained model to obtain an encoder model adapted to high-frequency data.

[0052] A small amount of high-frequency operation data of the power battery (such as operation data with a sampling frequency of 10-100 Hz) is obtained through an experimental vehicle or other means, and the pre-trained model in the cloud is fine-tuned to have the ability to process high-frequency data. The purpose is to make the pre-trained model based on the low-frequency operation data pre-training adapt to the high-frequency data feature extraction requirement, and obtain an encoder model adapted to high-frequency feature extraction. It should be noted that the pre-trained model includes an encoder and a decoder, and the decoder parameters are frozen during fine-tuning, only the encoder parameters are optimized, and a fine-tuned encoder model is obtained.

[0053] S603, based on the feature data output by the encoder model, jointly training a pre-warning model.

[0054] Based on historical fault data, combined with low-frequency operation data and its encoder encoded feature data, a pre-warning model for anomaly detection is trained with faults as labels. The purpose is to build a pre-warning model that can fuse high and low frequency features (i.e. low-frequency operation data and high-frequency features extracted by the encoder), and classify and predict based on fault types, which is used for subsequent detection and pre-warning of battery anomalies.

[0055] S604, deploying the encoder model and the pre-warning model on the vehicle end and uploading high-frequency feature data on the vehicle end.

[0056] The encoder model and the early warning model are pruned and quantized and deployed to the vehicle end, so that the vehicle end can detect and warn of abnormalities, and the high-frequency running data of the vehicle end is continuously uploaded to the cloud after being encoded by the encoder model. The purpose is to enable the vehicle end to have the ability to detect battery abnormalities in real time (by deploying the model to analyze the vehicle end data in real time), and to upload data to the cloud to provide a data basis for subsequent model optimization (vehicle end and cloud form a closed loop, vehicle end realizes real-time detection and response, and cloud receives data to support optimization). Specifically, the model is deployed after pruning redundant connections, quantizing to reduce the calculation precision requirement, and adapting to the hardware computing power of the vehicle end; when the vehicle end is running, high-frequency running data is collected in real time and converted into high-frequency feature vectors by the encoder model, and the original low-frequency running data is input into the early warning model for analysis, and the risk is detected to trigger an alarm to ensure driving safety, and only low-frequency running data and compressed high-frequency feature vectors are transmitted to reduce communication load.

[0057] S605, the cloud continuously updates the early warning model.

[0058] After the cloud collects all the high-frequency feature data uploaded by the vehicle end, the early warning model is trained in cycles based on the high-frequency features and in combination with the low-frequency running data, to realize continuous updating of the early warning model. The purpose is to use the data (i.e. low-frequency running data and high-frequency feature data) uploaded by the vehicle end continuously, which contains fault scenarios, to perform incremental training and optimization of the early warning model, then to process the optimized model to be lightweight and distribute it to the vehicle end, to complete the closed-loop updating of the early warning model, so that the model can continuously adapt to new battery fault modes and changes in data features, and improve the accuracy and reliability of early warning.

[0059] In summary, the application provides a power battery early warning method based on vehicle-cloud collaboration. First, a high-low frequency data fusion training mechanism is proposed, which pre-trains an autoencoder based on cloud massive low-frequency running data to establish basic feature representation capability, and then fine-tunes the model with a small amount of vehicle end high-frequency data to realize the organic combination of transient feature extraction and general representation. Second, a vehicle-cloud layered computing architecture is adopted, a lightweight model is deployed on the vehicle end to realize real-time detection and early warning of high-frequency data, and the cloud is responsible for aggregating multi-vehicle feature data to continuously optimize the model, forming a collaborative mode of "vehicle end + cloud". Finally, a closed-loop updating scheme is designed, the vehicle end only uploads low-frequency running data and high-frequency feature vectors (not the original high-frequency running data), which greatly reduces the communication overhead, the cloud performs incremental training based on the high-frequency feature vectors and feeds back to the vehicle end through knowledge distillation, to build a complete optimization closed loop of "data collection-model iteration-vehicle end update", effectively solving the coordination problem of high-frequency data processing, real-time response and model continuous optimization in power battery monitoring.

[0060] Therefore, the application realizes multi-scale feature representation of the battery state through high-low frequency data joint training (low frequency operation data pre-training of the cloud + high frequency feature data fine-tuning), effectively solves the problem of missing high frequency information caused by data down-conversion of the traditional cloud model; adopts a vehicle-cloud layered architecture (real-time processing of the vehicle-end lightweight model + global optimization of the cloud) to break through the vehicle-end computing power limit while ensuring millisecond-level early warning response; a feature-level interaction mechanism (only uploading low frequency operation data and high frequency feature data) greatly reduces communication load, and through the closed-loop optimization system of "data collection-cloud iteration-vehicle-end update", the system continuously adapts to dynamic scenes such as battery aging, and finally realizes breakthrough in three dimensions of early warning accuracy, real-time performance and system sustainability, providing an industrial-level solution with high performance and low cost for vehicle-cloud collaborative technology.

[0061] The embodiment also provides an electric vehicle, comprising: a vehicle-mounted BMS system configured to acquire high frequency operation data and low frequency operation data of a power battery; a vehicle-end chip loaded with a lightweight encoder model and an early warning model; a vehicle networking communication module configured to upload the high frequency feature vector and the associated low frequency operation data to the cloud, and receive an optimized early warning model issued by the cloud; a memory configured to store the encoder model, the early warning model and a computer program; a processor connected with the vehicle-mounted BMS system, the vehicle networking communication module and the memory, respectively, and the computer program is executed by the processor to implement the power battery early warning method according to any one of the above embodiments.

[0062] The embodiment also provides a computer readable storage medium having a computer program stored thereon, and the computer program is loaded by a processor to execute the steps in the method according to any one of the above embodiments.

[0063] In the embodiments of the application, the storage medium can be a disk, an optical disk, a read only memory (ROM) or a random access memory (RAM).

[0064] In the above embodiments, the description of each embodiment has its own focus, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0065] The power battery early warning method, the electric vehicle and the storage medium provided by the embodiments of the application are described in detail above, and the principles and implementation manners of the application are described by applying specific examples. The above embodiment description is only used to help understand the method of the application and its core idea; meanwhile, for those skilled in the art, according to the idea of the application, the specific implementation manner and application range will be changed, and the above description should not be understood as a limitation on the application.

Claims

1. A power battery early warning method, characterized in that, The method comprises the following steps: acquiring low-frequency operation data of power batteries uploaded by multiple electric vehicles; constructing a neural network model based on an encoder and a decoder structure, taking the low-frequency operation data as input, training the neural network model in an unsupervised learning manner, optimizing the model by minimizing the mean square error of the low-frequency operation data input and the decoder reconstruction output, and obtaining a pre-trained model, wherein the encoder of the pre-trained model has the ability to extract general features of power batteries under normal working conditions.

2. The power battery early warning method according to claim 1, characterized in that, Further comprising: acquiring high-frequency operation data of the power battery; taking the high-frequency operation data as input, freezing the decoder parameters of the pre-trained model, optimizing only the encoder parameters, fine-tuning the model by minimizing the error of the high-frequency operation data input and the decoder reconstruction output, and obtaining an encoder model adapted to high-frequency feature extraction.

3. The power battery early warning method according to claim 2, characterized in that, Further comprising: acquiring historical fault data of the power battery and the encoder model adapted to high-frequency feature extraction; inputting the historical fault data into the encoder model to generate a fault feature vector, taking the low-frequency operation data and the fault feature vector as double input sources, jointly training a supervised classifier with fault types as labels, and obtaining a pre-warning model that fuses high and low frequency features.

4. The power battery early warning method according to claim 3, characterized in that, Further comprising: receiving the encoded high-frequency feature vector uploaded by the vehicle end and its associated low-frequency operation data; in the case of power battery failure of the vehicle, taking the operation data of the fault period as a sample to incrementally train the pre-warning model; after the pre-warning model is optimized by incremental training, the pre-warning model is subjected to lightweight processing and then is issued to the vehicle end to complete the closed-loop update of the pre-warning model.

5. The power battery early warning method according to claim 3, characterized in that, The operation data of the power battery includes total voltage of the power battery pack, voltage of each single battery of the power battery pack, average voltage of each single battery of the power battery pack, highest voltage of each single battery of the power battery pack, lowest voltage of each single battery of the power battery pack, total current of the power battery pack, temperature of each battery probe, average temperature of the power battery pack, highest probe temperature of the power battery pack, and lowest probe temperature of the power battery pack. The sampling frequency of the low-frequency operation data is 1-10 seconds / time, and the sampling frequency of the high-frequency operation data is 10-100 Hz.

6. A power battery early warning method, characterized in that, The method comprises the following steps: deploying the encoder model and the pre-warning model trained and subjected to lightweight processing by the power battery pre-warning method according to any one of claims 3-5 to the vehicle end chip; acquiring high-frequency operation data of the power battery and inputting the high-frequency operation data into the encoder model to generate a high-frequency feature vector; inputting the high-frequency feature vector and the acquired low-frequency operation data of the power battery into the pre-warning model, and outputting an abnormal probability of the power battery by the pre-warning model.

7. The power battery early warning method according to claim 6, characterized in that, Further comprising: uploading the high-frequency feature vector and its associated low-frequency operation data to the cloud end; receiving the optimized pre-warning model issued by the cloud end to replace the pre-warning model deployed locally at the vehicle end.

8. The power battery early warning method according to claim 6, characterized in that, Further comprising: performing channel pruning and weight quantization processing on the encoder model and the pre-warning model to adapt to the computing power constraint of the vehicle end chip.

9. An electric vehicle, characterized by The method comprises the following steps: a vehicle-mounted BMS system for acquiring high-frequency operation data and low-frequency operation data of the power battery; a vehicle end chip loaded with a lightweight encoder model and a pre-warning model. The Internet of Vehicles communication module is configured to upload the high-frequency feature vector and the associated low-frequency operation data to the cloud, and receive an optimized early warning model issued by the cloud. The memory is configured to store the encoder model, the early warning model, and a computer program. The processor is connected with the vehicle-mounted BMS system, the Internet of Vehicles communication module, and the memory, respectively.

10. A computer-readable storage medium, characterized in that, The computer program is loaded by the processor to perform the steps of the power battery early warning method according to any one of claims 6-8. The computer program is loaded by the processor to perform the steps of the power battery early warning method according to any one of claims 6-8.

Images