Battery data hybrid compression and reconstruction system and method based on end-cloud collaboration

Abstract

The invention discloses a battery data hybrid compression and reconstruction system and method based on end-cloud collaboration, and the method comprises the steps: a vehicle-mounted end firstly collects battery state data, carries out the differential processing, automatically judges whether the current state is in a steady state or a complex working condition through the calculation of a nonlinear fluctuation index of data in a sliding time window, and generates a corresponding identifier; if the working condition is a steady-state working condition, compressing the data by adopting algebraic iteration linear regression to obtain a slope and an intercept coefficient, and packaging the slope and the intercept coefficient into a steady-state data packet to be uploaded; and if the working condition is a complex working condition, extracting a hidden space feature vector of the data by using a lightweight quantization encoder, and packaging the hidden space feature vector into an abnormal data packet to be uploaded. And the cloud server respectively calls a linear generator or a high-precision decoder according to the data packet identifier, reconstructs the received parameters or feature vectors into original differential time sequence data segments, and finally seamlessly splices all the data segments according to a time sequence to restore complete and continuous battery state time sequence data for storage and advanced analysis.

Description

Technical Field

[0001] This invention relates to the field of new energy vehicle battery management and vehicle-to-everything (V2X) communication, specifically to a battery data hybrid compression and reconstruction system and method based on end-to-cloud collaboration. Background Technology

[0002] With the explosive growth of the new energy vehicle industry, the safety and reliability of power batteries, as core components, have become a major concern. To monitor battery status in real time, national standards (such as GB / T 32960) and major automakers require vehicles to collect and upload monitoring data (including cell voltage, current, and temperature) from the battery management system (BMS) at high frequencies. However, with the increasing scale of battery packs (a single battery pack often contains hundreds of cells) and the increase in sampling frequency (from 0.1Hz to 10Hz or even higher), the amount of time-series data generated is growing exponentially.

[0003] Faced with the contradiction between the pressure of transmitting massive amounts of time-series data and the limited bandwidth of in-vehicle communication (4G / 5G traffic costs), existing technical solutions mainly fall into two categories: 1. Traditional compression methods based on fixed rules: This is currently the most commonly used solution for in-vehicle terminals (T-BOX). Typical algorithms include deadband compression, rotating door algorithm (SDT), or simple linear piecewise fitting. These methods usually set a fixed threshold; when the data change is less than the threshold, the data point is discarded or replaced with a straight line, and only key points with changes exceeding the threshold are retained.

[0004] 2. Cloud-based big data processing model: The vehicle only performs simple data packaging and uploading; complex feature extraction and fault diagnosis logic are all handled by high-performance cloud servers. The onboard BMS (Battery Management System) primarily handles data collection and basic protection functions, lacking advanced data analysis capabilities. Summary of the Invention

[0005] The purpose of this invention is to provide a battery data hybrid compression and reconstruction system and method based on end-to-cloud collaboration, which can resolve the contradiction between limited computing power on the vehicle side and limited bandwidth on the cloud side.

[0006] To achieve this objective, the present invention provides a battery data hybrid compression and reconstruction system based on edge-cloud collaboration, comprising: The vehicle-mounted data processing module is used to acquire the status time-series data of the power battery pack in real time according to a preset sampling frequency using the battery management system on the vehicle. It performs a first-order difference operation on the status time-series data to obtain differential time-series data, performs a sliding window framing operation on the differential time-series data, and calculates the nonlinear fluctuation index of the differential time-series data in the current time window in real time. When the nonlinear fluctuation index is less than or equal to a preset safety threshold, the power battery pack is in a stable operating condition in the current time window, and a corresponding stable data identifier is generated. Otherwise, the power battery pack is in a complex operating condition in the current time window, and an abnormal data identifier is generated. The linear regression module is used to perform steady-state data linear regression compression based on algebraic iteration on the differential time series data within the current time window when the power battery pack is in a stable operating condition within the current time window. This results in a univariate linear regression model of the differential time series data within the current time window. The optimization objective is to minimize the mean square error. A loss function is defined for all coefficients in the univariate linear regression model. All coefficients in the univariate linear regression model are iteratively updated using the expanded partial derivative formula. All coefficients in the iteratively updated univariate linear regression model and their corresponding stable data identifiers are packaged to obtain a stable data package, which is then uploaded to the cloud server. The feature extraction module is used to input the differential time series data in the current time window into the autoencoder model and perform forward propagation operation when the power battery pack is in a complex working condition in the current time window. It obtains the latent space feature vector corresponding to the differential time series data in the current time window, packages the latent space feature vector and the corresponding abnormal data identifier to obtain an abnormal data packet, and uploads the abnormal data packet to the cloud server. The cloud-based data processing module performs linear data segment reconstruction on stable data packets, converting them into differential time-series data within the corresponding time window. It also performs deep decoding reconstruction on abnormal data packets, restoring them to differential time-series data within the corresponding time window. Finally, it concatenates the differential time-series data within the time window of the stable data packets with the differential time-series data within the time window of the abnormal data packets in chronological order to obtain complete timing data of the power battery pack status.

[0007] The beneficial effects of this invention are: 1. This invention resolves the inherent trade-off between high compression ratio and high fault detection rate, significantly improving battery safety early warning capabilities. Traditional dead-zone compression or rotating door (SDT) algorithms, in order to save bandwidth, often treat minute voltage drops in the early stages of battery thermal runaway as noise filtering, resulting in the loss of fault features. This invention introduces an adaptive current-sharing mechanism based on MSE. Under steady-state conditions, it utilizes linear regression to achieve extreme compression; once nonlinear fluctuations (suspected faults) are detected, it switches to deep learning feature extraction mode within milliseconds. In terms of technical effectiveness, this mechanism ensures that while saving over 90% of steady-state data bandwidth, it can completely capture and reconstruct waveform details of fault precursors (such as micro-short-circuit features), enabling the cloud to detect potential battery safety hazards in advance and prevent missed detections.

[0008] 2. It has broken through the computing power bottleneck of automotive-grade edge computing chips, enabling the low-cost deployment of complex algorithms. Existing deep learning compression models have a large number of parameters, making them difficult to run in real time on resource-constrained automotive MCUs (low clock speed, small memory). Traditional least squares methods involve complex matrix inversion operations, resulting in high CPU usage. This invention offers dual optimizations for edge computing environments: firstly, it transforms the linear regression algorithm into a matrix-free algebraic expansion iterative formula, significantly reducing computational complexity; secondly, it performs model pruning and 8-bit quantization on the LSTM (Long Short-Term Memory) autoencoder. In terms of technical performance, this enables the entire hybrid compression algorithm to run smoothly on ordinary ARM Cortex-M series chips without the need for additional expensive AI acceleration hardware (NPU), reducing the hardware cost and system power consumption of the BMS.

[0009] 3. A novel paradigm for latent space feature vector transmission is proposed, enhancing the value of cloud-based data mining. Traditional methods only transmit sample points of the original data, requiring tedious feature engineering in the cloud before it can be used for lifetime prediction. This invention transmits a latent vector extracted by the encoder under abnormal operating conditions. This vector is essentially a high-level abstraction of the state of battery health (SOH). In terms of technical effectiveness, this vector is not only used to reconstruct waveforms, but can also be directly used as input features for cloud-based battery life prediction (RUL) models, reducing redundant computational overhead in the cloud and realizing an efficient data value chain through edge-cloud collaboration.

[0010] 4. By utilizing differential preprocessing and an asymmetric encoding / decoding architecture, bandwidth adaptation under all operating conditions is achieved. First-order differential preprocessing eliminates the DC component of the data, combined with the superior reconstruction capabilities of the cloud-based high-precision FP32 decoder. In terms of technical effectiveness, even in weak network environments with poor signal (low bandwidth), a high-fidelity battery operating curve can be reconstructed in the cloud by transmitting only a minimal amount of fitting parameters or feature vectors, ensuring the continuity and integrity of data monitoring. Attached Figure Description

[0011] Figure 1 This is a diagram of the edge-cloud collaborative system architecture of the present invention; Figure 2 This is a flowchart of the main method of the present invention; Figure 3 This is a flowchart of the deep learning encoding and decoding process based on edge-cloud collaboration of the present invention; Figure 4 This is a schematic diagram of the structure of the present invention. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to represent selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0013] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: Example 1 like Figure 4 As shown, a battery data hybrid compression and reconstruction system based on edge-cloud collaboration includes: The vehicle-mounted data processing module is used to acquire the status time-series data of the power battery pack in real time according to a preset sampling frequency using the battery management system on the vehicle. It performs a first-order difference operation on the status time-series data to obtain differential time-series data, performs a sliding window framing operation on the differential time-series data, and calculates the nonlinear fluctuation index of the differential time-series data in the current time window in real time. When the nonlinear fluctuation index is less than or equal to a preset safety threshold, the power battery pack is in a stable operating condition in the current time window, and a corresponding stable data identifier is generated. Otherwise, the power battery pack is in a complex operating condition in the current time window, and an abnormal data identifier is generated. The linear regression module is used to perform steady-state data linear regression compression based on algebraic iteration on the differential time series data within the current time window when the power battery pack is in a stable operating condition within the current time window. This results in a univariate linear regression model of the differential time series data within the current time window. The optimization objective is to minimize the mean square error. A loss function is defined for all coefficients in the univariate linear regression model. All coefficients in the univariate linear regression model are iteratively updated using the expanded partial derivative formula. All coefficients in the iteratively updated univariate linear regression model and their corresponding stable data identifiers are packaged to obtain a stable data package, which is then uploaded to the cloud server. The feature extraction module is used to input the differential time series data in the current time window into the autoencoder model and perform forward propagation operation when the power battery pack is in a complex working condition in the current time window. It obtains the latent space feature vector corresponding to the differential time series data in the current time window, packages the latent space feature vector and the corresponding abnormal data identifier to obtain an abnormal data packet, and uploads the abnormal data packet to the cloud server. The cloud-based data processing module performs linear data segment reconstruction on stable data packets, converting them into differential time-series data within the corresponding time window. It also performs deep decoding reconstruction on abnormal data packets, restoring them to differential time-series data within the corresponding time window. Finally, it concatenates the differential time-series data within the time window of the stable data packets with the differential time-series data within the time window of the abnormal data packets in chronological order to obtain complete timing data of the power battery pack status.

[0014] The following shortcomings exist in the existing technology: 1. Lossy compression of traditional power battery pack data leads to the loss of early-stage fault characteristics: Severe faults such as thermal runaway, micro-short circuits, or lithium plating often manifest in the early stages as minute voltage drops or nonlinear fluctuations over a very short period (i.e., fault characteristic data). Traditional SDT or fixed-threshold compression algorithms are insensitive to these characteristics. They often misjudge these minute but crucial nonlinear fluctuations as "sensor noise" and smooth them out. As a result, while the data received by the cloud saves bandwidth, it loses the most critical fault fingerprint information, rendering the safety warning system ineffective and making it impossible to trace the true cause of the accident.

[0015] 2. Mismatch between in-vehicle edge computing power and complex AI models: Although compression algorithms based on deep learning (such as LSTM and Transformer) can extract nonlinear features well, these models have a huge number of parameters and high computational complexity. However, the existing in-vehicle BMS main control chips (mostly ARM Cortex-M or R series) have limited resources and cannot support real-time inference of large-scale neural networks, making it difficult to deploy advanced AI algorithms on the vehicle side, and vehicles are still in the stage of simple uploading.

[0016] 3. Lack of adaptive transmission mechanism based on operating condition awareness: Existing solutions adopt a one-size-fits-all approach. Regardless of whether the vehicle is in a stable high-speed cruising state (low data value), or in a state of violent acceleration, deceleration, or fault threshold (high data value), the same compression logic is used. This results in wasted bandwidth under stable operating conditions, while under fault conditions, excessive compression leads to information distortion, lacking flexibility.

[0017] This invention utilizes edge computing technology to introduce a hybrid architecture of lightweight deep learning models and linear regression algorithms at the resource-constrained BMS end, enabling real-time perception of battery conditions: high compression ratio transmission is used in steady state, and automatic switching to high-fidelity feature extraction mode is performed in abnormal state, thereby maximizing the value of data under limited bandwidth and computing power.

[0018] This invention operates within an asymmetric computing power collaborative architecture comprised of an in-vehicle edge computing terminal, a wireless communication link, and a cloud-based big data platform. This architecture aims to resolve the contradiction between limited computing power on the vehicle side and limited bandwidth on the cloud side. The architecture includes the following modules: 1. The in-vehicle edge computing terminal acts as the producer and primary processor of data. This terminal relies on the main control microcontroller (MCU, such as the Infineon Aurix TC3xx series or ARM Cortex-M7 core chip) of the vehicle's battery management system (BMS). This module includes a multi-dimensional data acquisition module, a differential preprocessing module, and a dual-mode hybrid compression unit. Specifically, the multi-dimensional data acquisition module is responsible for interfacing with the AFE (analog front-end) and current sensor to acquire time-series physical quantities such as voltage, current, and temperature in real time; the differential preprocessing module is used to perform first-order differential operations on the raw time-series data to eliminate the DC component of the data and reduce information entropy; the operating condition identification and current splitting module is equipped with lightweight decision logic based on mean square error (MSE) to evaluate the nonlinearity of the data in real time and control the data flow direction; the dual-mode hybrid compression unit includes two parallel units: a linear regression unit is used to process steady-state data and perform gradient descent fitting with algebraic expansion; the lightweight AI encoding unit is equipped with an LSTM encoder with 8-bit quantization pruning to extract latent space feature vectors under complex operating conditions.

[0019] 2. The wireless communication link relies on the vehicle-mounted remote terminal (T-BOX) and utilizes 4G / 5G cellular networks or V2X communication technology to establish a highly reliable data transmission channel between the vehicle and the cloud. This link is responsible for transmitting compressed data packets containing different packet header identifiers.

[0020] 3. The cloud-based big data platform, acting as both the final consumer and advanced processor of data, is deployed on a high-performance server cluster. It possesses unlimited computing resources. This module includes a data receiving gateway module, an AI decoding cluster module, and a data reconstruction and fusion module. Specifically, the data receiving gateway module is responsible for the access and protocol parsing of high-concurrency data packets; the AI ​​decoding cluster module deploys a full-precision (FP32) deep neural network decoder, responsible for reconstructing the received latent vectors into high-fidelity waveforms; the data reconstruction and fusion module is responsible for time-series concatenation of linearly reconstructed data and deep-decoded data based on timestamps, and writing the complete lifecycle data into a time-series database (TSDB).

[0021] In some embodiments, such as Figure 1 As shown, the edge-cloud collaborative system architecture of this invention is as follows: the voltage (obtained through a voltage sampling line), current (obtained through a Hall sensor), and temperature signals (obtained through a temperature probe) of the power battery pack are connected to the DMA controller (memory access controller), SRAM (semiconductor memory), and Flash (semiconductor memory) of the BMS core at the vehicle edge via physical connections such as power lines, data lines, and CAN bus. The vehicle edge uses a microcontroller (MCU) as the computing center, which integrates a DSP (digital signal processor), an FPU (floating-point unit), and a dedicated CMSRS-NN processing unit (a hardware accelerator customized for neural network inference). The data acquisition module completes data acquisition and preprocessing through multiple sensors and interfaces such as I2C (communication bus protocol) and ADC (digital-to-analog converter), and performs operations including sliding window analysis and mean square error algorithm (MSE) on the preprocessed data. The system performs real-time operating condition discrimination, including nonlinear exponential calculation, and then drives the hybrid compression module for split processing: in steady state, a linear regression algorithm is executed for fitting, while in abnormal conditions, an AI encoder (lightweight encoder) extracts features; the processed data is uploaded to the cloud platform via a 4G / 5G wireless link, along with necessary model weight update (OTA) requests; after receiving and parsing the data, the cloud platform uses an AI decoder to reconstruct the uploaded feature vectors, and performs full storage and time-series splicing with the linear reconstruction segment in the data fusion module. The cloud platform trains the model based on historical data and continuously optimizes the vehicle-side model through the model weight update (OTA) mechanism. Finally, it is used to provide complete and traceable battery status information and early warning services for the central control screen warning, mobile APP and Web monitoring backend of the application layer, thereby realizing digital management of the entire battery life cycle.

[0022] In some technical solutions, methods for performing first-order difference operations on the state-time data of the power battery pack include: ; in, For the state timing data of the power battery pack The output value of the differential time series data after differential processing. for The raw physical quantity values ​​of the power battery pack status data collected at all times. The start time of the current data transmission frame The original physical quantity values ​​of the power battery pack state data collected at the previous sampling time t are: At that time, directly retain the original value. This serves as the baseline anchor point data for the current frame, used for initial positioning during cloud reconstruction.

[0023] Performing first-order difference operations on the state-time data of the power battery pack can transform the original absolute physical quantity sequence into a relative change quantity sequence. This effectively eliminates the inherent DC component and long-term drift in the data, significantly reducing the dynamic range and information entropy of the data, laying a good foundation for subsequent data compression. At the transmission level, the change quantity data usually has smaller numerical amplitudes and more stable statistical characteristics, which not only greatly improves the efficiency and accuracy of compression algorithms such as linear fitting and AI feature extraction, but also directly reduces the number of bits required for encoding, thereby greatly saving wireless communication bandwidth and traffic costs. At the same time, the design of retaining the original physical value as a reference anchor point at the beginning of the data transmission frame (t=t0) provides a key and accurate initial positioning point for cloud reconstruction, ensuring that the original absolute physical value time-time data can be restored losslessly and with high fidelity through the accumulation operation of the difference sequence. Finally, the overall technology achieves a closed loop of efficient compression, reliable transmission and accurate reconstruction of battery state data under limited bandwidth.

[0024] In some embodiments, the state timing data acquisition process of the power battery pack includes: the data acquisition module within the vehicle battery management system (BMS) acquires multi-dimensional state data of the battery pack in real time at a preset sampling frequency (e.g., 10Hz), wherein the multi-dimensional state data includes, but is not limited to: the voltage sequence of each individual battery cell acquired through the voltage sampling channel. The total current sequence of the battery pack circuit obtained through a current sensor and temperature sequences of key sampling points obtained through temperature probes. To eliminate the dimensional differences between different sensors, the processor timestamps the physical quantity data of the different dimensions mentioned above to construct a unified multidimensional time-series input vector.

[0025] To reduce the dynamic range of data transmission and extract variation features, the processor performs a first-order difference operation on the acquired raw time-series data. This operation transforms the original absolute numerical sequence into a relatively varying numerical sequence. In some technical solutions, methods for performing sliding window framing on the differential time-series data and calculating the nonlinear fluctuation exponent of the differential time-series data within the current time window in real time include: Construct a sliding time window of length N in chronological order, denoted as Where k is the window number, N is the preset window length, and a preset overlap rate is designed between adjacent time windows. For the differential time series data within the current time window, a virtual linear baseline connecting the first and last data points of the sliding time window is constructed: the theoretical projection value corresponding to the virtual linear baseline is calculated based on the difference between the first and last differential time series data within the current time window. The theoretical projection values ​​satisfy: ; Based on the theoretical projection values ​​corresponding to the virtual linear baseline, the mean square error (MSE) algorithm is used to calculate the deviation of the differential time series data within the current time window from the virtual linear baseline: ; in, The first time within the current sliding time window The theoretical projection values ​​of each differential time series data point onto a virtual linear baseline. This is the first differential time series data value within the current sliding time window. This is the second differential time series data value within the current sliding time window. Here, i = 1, 2, ..., N, represents the position index within the time window, corresponding to the 1st to Nth data point within the current sliding time window. This is the last differential time series data value within the current sliding time window. This is the nonlinear fluctuation index of the differential time series data within the current time window. The first time within the current sliding time window Differential time series data.

[0026] In some embodiments, by performing a sliding window framing operation on the differential time-series data stream, the nonlinear fluctuation index of the data within the current time window is calculated in real time, and the corresponding compression processing unit's adaptive condition determination is dynamically activated based on the comparison result of the index with a preset safety threshold. The specific execution logic is as follows: The processor allocates a first-in-first-out (FIFO) buffer queue in memory to perform frame processing on the differential timing data stream. A sliding time window of length N is constructed, denoted as . Where k is the window number and N is the preset window length (for example, N=100, which corresponds to a data duration of 10 seconds). The overlap rate between adjacent time windows is set to a fixed percentage (for example, 20%).

[0027] Read the security threshold preset in the register , calculate and Compare the results and perform the following path switching based on the comparison outcome: Branch A: If If the current condition is determined to be steady state (such as constant speed driving or stationary), the processor activates the linear regression compression unit to perform steady state data linear regression compression based on algebraic iteration on the differential time series data stream of the current time window. Branch B: If If the current condition is determined to be complex (such as rapid acceleration, precursor to thermal runaway, or micro short circuit), the processor activates a lightweight AI encoding unit to perform lightweight deep feature extraction on the differential time series data of the current time window for abnormal operating conditions.

[0028] In some embodiments, the value of the security threshold λ is mainly determined from two aspects: First, the statistical characteristics of differential time-series data of the battery system under known steady-state conditions (such as constant speed cruise and stationary operation) are calculated by analyzing a large amount of historical steady-state data. The distribution of the mean is taken as the upper bound by adding a certain number of standard deviations to its mean. Secondly, by considering the nonlinear fluctuation amplitude caused by early signs of battery failure (such as voltage dips due to micro-short circuits), the threshold can be ensured to sensitively capture such anomalies. Specifically, for example, when the sampling frequency is 10Hz and the window length N=100, the differential voltage sequence under typical steady-state conditions... L is usually in 10 -6 ~10 -4 V 2 The magnitude, while the fluctuations of the precursors to failure It can reach 10 -3 V 2 above.

[0029] Therefore, the value of λ can be, but is not limited to, 0.001 V. 2 (i.e., 1×10) -3 V 2 This value effectively delineates between steady-state noise and early signs of failure. In practical engineering deployments, this threshold can be adjusted through calibration tests based on specific battery pack specifications and sensor accuracy.

[0030] By constructing a virtual linear baseline connecting the beginning and end of a time window and calculating the mean square error between the actual data points and this baseline, the nonlinear fluctuation degree of a time series data segment can be quantified in real time and quantitatively. This provides a reliable and efficient decision-making basis for subsequent adaptive data compression strategies: by comparing the calculated nonlinear fluctuation index with a preset threshold, the system can accurately and automatically identify whether the battery is in a steady-state condition with smooth data or an abnormal condition with drastic data fluctuations. This intelligent judgment based on the characteristics of the data itself allows the system to abandon a one-size-fits-all approach and instead achieve a hybrid compression mode that uses high compression ratio linear fitting in steady-state conditions and switches to high-fidelity feature extraction in abnormal conditions. This fundamentally solves the inherent contradiction in traditional compression methods where high compression ratio and high fault feature fidelity are mutually exclusive. It greatly saves wireless transmission bandwidth and ensures that subtle precursory features of serious faults such as battery thermal runaway can be completely captured and uploaded, significantly improving the reliability of battery safety monitoring. Meanwhile, the method has a simple calculation process, mainly involving difference, linear interpolation and sum of squares operations. It has low computing power requirements for vehicle edge computing devices and meets the stringent requirements of automotive-grade chips for real-time performance and low complexity in resource-constrained environments.

[0031] To accommodate the limited floating-point computing power of the vehicle-mounted BMS microcontroller (MCU) and avoid the high-consumption matrix inversion operation in the traditional least squares method, this invention employs an algebraic expansion gradient descent algorithm to linearly fit the differential data within the time window. In some technical solutions, when the power battery pack is in a stable operating condition within the current time window, methods for performing steady-state data linear regression compression based on algebraic iteration on the differential time-series data within the current time window to obtain a univariate linear regression model of the differential time-series data within the current time window include: For the current time window differential sequence within Establish a univariate linear regression model: ; Represents the relative time index within the time window. ; a is the slope coefficient to be solved, which represents the changing trend of the differential time series data within the current time window; b is the intercept coefficient to be solved, which represents the baseline offset of the differential time series data within the current time window. These are the predicted values ​​from a univariate linear regression model. Using the minimization of mean squared error as the optimization objective, a loss function is defined for all coefficients in the univariate linear regression model. The methods include: .

[0032] By employing a linear regression compression method based on algebraic iteration, the system can significantly reduce the amount of data to be transmitted with an extremely high compression ratio when the power battery is in a stable operating condition. Only two parameters, slope 'a' and intercept 'b', need to be uploaded to completely reconstruct the time-series data of the entire time window in the cloud, thereby saving more than 90% of wireless communication bandwidth and reducing traffic costs. At the same time, this method transforms the complex matrix inversion operation in the traditional least squares method into a scalar iterative operation suitable for embedded processors, which greatly reduces the computational complexity and real-time computing load of the vehicle microcontroller, avoids excessive CPU resource consumption, and enables the advanced data compression algorithm to run stably and efficiently on resource-constrained automotive-grade chips, providing ample computing power for the software function upgrade of the battery management system. By uploading fitting parameters instead of original data points, it not only ensures that the steady-state data waveform can be linearly restored without loss in the cloud, maintaining the mathematical consistency of the data, but also provides a clean and continuous data foundation for subsequent battery state analysis in the cloud, effectively supporting the full life cycle management of the battery.

[0033] In some technical solutions, methods include iteratively updating all coefficients in a univariate linear regression model using the expanded partial derivative formula, and then packaging all coefficients in the iteratively updated univariate linear regression model and their corresponding stable data identifiers to obtain a stable data package. Update intercept coefficient : ; Update slope coefficient : ; Where a is the slope coefficient value of the current iteration step, and b is the intercept coefficient value of the current iteration step. The updated slope coefficient value. These are the updated intercept coefficient values. The learning rate (including but not limited to 0.01) is preset according to the calculation cycle of the BMS system. When the number of iterations reaches the preset convergence condition, the iteration stops, and the updated slope coefficient value obtained at the time of stopping the iteration is used. ) and intercept coefficient value ( ) and the stable data identifier corresponding to the current time window ( ), current time window start time ( ), current time window length ( Stable data packets are obtained by combining them in a preset order.

[0034] In some embodiments, the preset convergence condition includes, but is not limited to, when the number of iterations reaches a preset value (e.g., k=50 times) or the loss function. When convergence occurs, iteration stops. Stable data packet structures include, but are not limited to, those mentioned above. The start time of the current time window is used to provide the starting anchor point of the differential time series data corresponding to the data packet on the time axis. The cloud server determines the number of data points or time span contained in the differential time series data corresponding to the data packet based on the start time and length of the current time window.

[0035] By transforming the matrix inversion operation in the traditional least squares method into a matrix-free scalar accumulation iteration, the computational complexity and CPU usage are significantly reduced. This enables the algorithm to run in real time on resource-constrained automotive BMS microcontrollers (such as the ARM Cortex-M series) without requiring additional expensive hardware. At the same time, this iterative optimization ensures high-precision fitting parameters under preset convergence conditions. This allows for the complete reconstruction of data segments in the cloud with only a small amount of slope and intercept coefficients (rather than a large amount of original data). As a result, a data compression rate of over 90% is achieved under stable operating conditions, greatly saving wireless communication bandwidth and traffic costs, and ensuring the temporal continuity of the reconstructed data. This provides a reliable data foundation for cloud-based battery status analysis, fault tracing, and full lifecycle management.

[0036] In some technical solutions, when the power battery pack is under complex operating conditions within the current time window, the differential time series data within the current time window is input into the autoencoder model and a forward propagation operation is performed to obtain the latent space feature vector corresponding to the differential time series data within the current time window. The method of packaging the latent space feature vector and the corresponding abnormal data identifier to obtain an abnormal data packet includes: taking the differential sequence within the current sliding time window of length N... After normalization, the data is input into the Encoder model and the forward propagation operation is performed: ; in, This is an encoder model that has undergone int8 quantization. The output latent space feature vector The dimension; To create a latent space feature vector that highly condenses the nonlinear characteristics of fault waveforms, For the encoder model, the difference sequence within the current sliding time window The extracted and compressed feature vector of the first data point in the innermost region. For the encoder model, the difference sequence within the current sliding time window The extracted and compressed feature vector of the second data point. For the encoder model, the difference sequence within the current sliding time window The extracted and compressed feature vector of the m-th data point will be used to construct the latent space feature vector. ), the abnormal data identifier corresponding to the current time window ( ) and the start time of the current time window ( The abnormal data packets are obtained by combining them in a preset order.

[0037] In some embodiments, the in-vehicle edge computing unit is pre-configured with a pruned and quantized LSTM autoencoder model. The model uses a Long Short-Term Memory (LSTM) network as its backbone, comprising an input layer, two hidden layers, and a bottleneck layer with reduced dimensions. To adapt to the fixed-point arithmetic characteristics of cores such as ARM Cortex-M, the pre-configured model undergoes post-quantization after training. This operation maps the original 32-bit floating-point (FP32) weight parameters to 8-bit integers, and accelerates the process using embedded neural network instruction sets such as CMSIS-NN. The structure of abnormal data packets includes, but is not limited to, […]. The start time of the current time window is used to provide the starting anchor point of the differential time-series data corresponding to this data packet on the time axis.

[0038] By deploying a lightweight autoencoder with pruned and 8-bit integer quantization at the vehicle edge, efficient and real-time processing of battery abnormal operating condition data is achieved while significantly overcoming the computing power and memory limitations of automotive-grade microcontrollers. The technical advantage lies in using the encoder to nonlinearly map high-dimensional, redundant raw differential time-series data into extremely low-dimensional latent space feature vectors. This enables near-lossless transmission with extremely high compression ratios under abnormal operating conditions, significantly saving wireless communication bandwidth and traffic costs. Simultaneously, it ensures that subtle nonlinear waveform features, such as micro-short circuits and lithium plating, are completely extracted and preserved, without being filtered as noise. The uploaded latent space feature vectors are not only the carrier of data compression but also a highly abstract representation of battery health status. They can be directly used as input to a high-precision decoder in the cloud to reconstruct high-fidelity waveforms and can also seamlessly connect to cloud-based battery life prediction and fault diagnosis models, avoiding the computational overhead of repeated feature engineering in the cloud. This constructs a highly efficient end-to-cloud collaborative value chain from edge-side feature extraction to cloud-based intelligent analysis.

[0039] In some technical solutions, the method of performing linear data segment reconstruction operations on stable data packets to convert them into differential time-series data within the corresponding time window includes: a linear generator on a cloud server, based on the updated slope coefficient value, intercept coefficient value, current time window start time, and current time window length carried in the stable data packet, uses linear equations to generate a cloud-based differential time-series data sequence within the current time window. The formula is as follows: ; in, This represents the prediction value of the univariate linear model for the first difference time series data within the current sliding time window. This represents the prediction value of the second difference time series data within the current sliding time window by the univariate linear model. For a univariate linear model, the first _ ... Predicted values ​​of differential time series data The length of the current time window.

[0040] In some embodiments, the cloud server includes, but is not limited to, a high-performance computing server equipped with an NVIDIA A100 GPU (an NVIDIA high-performance computing chip), and an edge cloud server equipped with an NVIDIA Tesla T4 (a high-performance chip) or an A10 GPU (a high-performance chip).

[0041] After reconstruction, the differential time series data in the cloud differential time series data sequence are accumulated and restored to state time series data: the initial state data of the power battery pack corresponding to the start time of the current time window is added to the prediction value of the first differential time series data in the current sliding time window by the univariate linear model in the cloud differential time series data sequence. The first state time series data is obtained by adding the first state time series data to the prediction value of the second difference time series data within the current sliding time window by the univariate linear model. The corresponding second state time series data is obtained by adding the second state time series data to the prediction value of the third difference time series data within the current sliding time window by the univariate linear model. The corresponding third state time series data, and so on, until the state time series data corresponding to all differential time series data in the cloud differential time series data sequence are obtained.

[0042] By using a cloud-based linear generator to directly reconstruct complete differential time-series data from the simplified fitting parameters (slope and intercept) uploaded from the vehicle, and then accumulating and restoring the data using predetermined initial physical values, high-fidelity recovery of the battery's steady-state operating curve is achieved with almost zero original data transmission cost. The technical advantage lies in fully utilizing the strong linearity of steady-state data, transforming the cloud reconstruction process into deterministic calculations that do not require complex iterations, significantly reducing the computational load and processing latency of the cloud server. Simultaneously, through an accumulation operation that is the inverse of differential preprocessing, the absolute continuity and physical authenticity of the restored voltage, current, temperature, and other state-time data in both numerical and temporal dimensions are ensured. This provides a high-quality data foundation that can be directly utilized for subsequent battery lifecycle analysis, state assessment, and big data mining, thereby achieving the end-cloud collaborative design goal of extreme compression at the vehicle end and efficient restoration in the cloud.

[0043] In some embodiments, the cloud server receives compressed data packets uploaded by the vehicle terminal, identifies the compression mode based on the packet header identifier, and calls a linear generator or a depth decoder to restore the data. Finally, it splices and merges the data segments based on the timestamp.

[0044] The specific execution logic for the data reconstruction and fusion is as follows: Cloud-based parsing and traffic splitting: The cloud access gateway receives the data stream uploaded by the vehicle terminal, extracts the 8-bit header identifier, and uses this identifier to perform traffic splitting operations. Branch A: If The data is determined to be steady-state linear, and the system calls the linear generator to extract the parameters. That is, proceed to logic 2 and perform refactoring; Branch B: If The data was determined to be characteristic data of complex working conditions. The system then invoked a deep decoder to extract latent space feature vectors. Enter logic 3 and perform refactoring.

[0045] In some technical solutions, methods for performing deep decoding and reconstruction on abnormal data packets to restore them to differential time-series data within the corresponding time window include: using a decoder built into a cloud server to analyze the received latent space feature vector based on the start time of the current time window in the abnormal data packet. Mapping back to a high-dimensional time series space, differential time series data within the corresponding time window, including nonlinear details, are reconstructed. : ; in, This is a full-precision deep neural network decoder running on a cloud server.

[0046] By deploying a high-performance server cluster and a full-precision (FP32) deep neural network decoder in the cloud, an asymmetric decoding architecture matching the vehicle-mounted lightweight encoder was constructed. This enabled high-fidelity, lossless reconstruction of battery data under abnormal operating conditions. By fully utilizing the near-limitless computing resources in the cloud, the highly compressed latent space feature vectors uploaded from the vehicle were accurately mapped and restored to the original differential time-series waveforms containing complete nonlinear details. This ensured that, under the premise of extreme compression transmission at the vehicle end, the cloud could completely recover subtle features that indicate early battery faults, such as voltage micro-drops and nonlinear fluctuations. This completely solved the safety hazard of losing fault fingerprints caused by traditional lossy compression. At the same time, this high-precision decoding and reconstruction process provided a high-quality data source for the cloud time-series database, enabling subsequent battery life cycle management, health status assessment, and fault root cause tracing analysis to be based on complete and accurate data, thus opening up a closed loop from edge-side feature abstraction to cloud-based intelligent decision-making.

[0047] In some embodiments, a high-precision decoder (Decoder) jointly trained with the in-vehicle encoder is deployed on the cloud server. This decoder runs on a cloud-based high-performance computing unit (GPU) and employs 32-bit floating-point (FP32) precision. The decoder then processes the received low-dimensional feature vectors... Mapping back to a high-dimensional time space, the original waveform containing nonlinear details is reconstructed. The cloud-based data fusion module seamlessly stitches together the steady-state data segment output by the linear generator and the fault feature data segment output by the deep decoder in chronological order. The resulting complete time-series data stream is stored in the cloud-based time-series database (TSDB) for subsequent battery lifecycle management and fault tracing analysis.

[0048] like Figure 2As shown, the main method flow of this invention is as follows: after the vehicle is ignited, the process begins. The vehicle terminal first performs data acquisition to obtain the original state timing data of the battery pack. The original state timing data is preprocessed differentially to convert it into differential timing data. The differential timing data is framed by a sliding window to form a timing window to be processed. The system identifies the operating condition by calculating the nonlinear fluctuation index of the current window (i.e., determining operating condition a > λ). If the index is not greater than the threshold λ (NO, i.e., steady state), linear regression compression is performed on the data to obtain the coefficients of the univariate linear regression model, and the data is packaged (Header=0x00) to form a steady-state data packet for uploading. If the index is greater than λ (YES, abnormal or transient), lightweight AI encoding is initiated to extract the core features of the data, generate a feature vector Z, and the data is packaged (Header=0x01) to form an abnormal data packet for uploading. After all data packets are received and parsed by the cloud, they are divided into modes according to the packet header identifier: data packets identified as 0x00 are reconstructed by the linear generator into differential time-series data segments, and data packets identified as 0x01 are reconstructed by the deep learning decoder into differential time-series data segments containing nonlinear details. Finally, the cloud seamlessly splices all data segments in chronological order through the time-series data fusion module to restore complete, high-fidelity battery state time-series data, which is ultimately stored in the time-series database for full life-cycle management and analysis. This concludes the process.

[0049] like Figure 3 As shown, the deep learning encoding and decoding process based on edge-cloud collaboration is as follows: When the system inputs complex operating conditions, the data located at the vehicle edge computing terminal is first normalized. Then, a lightweight encoder (Int8 quantization or CMSIS-NN) module is loaded to compress and encode the normalized time series data and perform forward inference operations, finally outputting a latent space feature vector (Z). This low-dimensional feature vector is wirelessly transmitted to the cloud server via 4G / 5G. After receiving VectorZ (latent space feature vector), the cloud calls a deep decoder (FP32 full precision or inverse LSTM) and performs inverse mapping reconstruction operation to output a reconstructed time series waveform. This high-fidelity waveform data is then sent to the time series fusion module to be spliced ​​with the linearly reconstructed steady-state data segment to restore the complete battery state time series sequence.

[0050] Example 2 A method for hybrid compression and reconstruction of battery data based on edge-cloud collaboration according to the system includes: The vehicle-mounted battery management system acquires the state time-series data of the power battery pack in real time according to a preset sampling frequency. First-order difference operation is performed on the state time-series data to obtain differential time-series data. Sliding window framing operation is performed on the differential time-series data, and the nonlinear fluctuation index of the differential time-series data within the current time window is calculated in real time. When the nonlinear fluctuation index is less than or equal to a preset safety threshold, the power battery pack is in a stable operating condition within the current time window, and a corresponding stable data identifier is generated. Otherwise, the power battery pack is in a complex operating condition within the current time window, and an abnormal data identifier is generated. When the power battery pack is in a stable operating condition within the current time window, a linear regression compression of steady-state data based on algebraic iteration is performed on the differential time series data within the current time window to obtain a univariate linear regression model of the differential time series data within the current time window. The mean square error is minimized as the optimization objective. A loss function is defined for all coefficients in the univariate linear regression model. All coefficients in the univariate linear regression model are iteratively updated using the expanded partial derivative formula. All coefficients in the iteratively updated univariate linear regression model and their corresponding stable data identifiers are packaged to obtain a stable data package, which is then uploaded to the cloud server. When the power battery pack is under complex operating conditions within the current time window, the differential time series data within the current time window is input into the automatic encoder model and forward propagation is performed to obtain the latent space feature vector corresponding to the differential time series data within the current time window. The latent space feature vector and the corresponding abnormal data identifier are packaged to obtain an abnormal data packet, and the abnormal data packet is uploaded to the cloud server. For stable data packets, a linear data segment reconstruction operation is performed to convert the stable data packets into differential time-series data within the corresponding time window. For abnormal data packets, a deep decoding reconstruction operation is performed to restore the abnormal data packets into differential time-series data within the corresponding time window. Then, the differential time-series data within the time window of the stable data packets and the differential time-series data within the time window of the abnormal data packets are concatenated in chronological order to obtain the complete timing data of the power battery pack status.

[0051] Example 3 The present invention provides a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, performs the steps of the method described in Embodiment 2.

[0052] This invention can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented in whole or in part as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., a solid-state drive (SSD)).

[0053] Those skilled in the art will readily understand that the above are merely preferred embodiments of the present invention and do not limit the present invention. Any modifications, combinations, substitutions, improvements, etc., made under the spirit and principles of the present invention are included within the protection scope of the present invention.

[0054] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

Claims

1. A battery data hybrid compression and reconstruction system based on edge-cloud collaboration, characterized in that, It includes: The vehicle-mounted data processing module is used to acquire the status time-series data of the power battery pack in real time according to a preset sampling frequency using the battery management system on the vehicle. It performs a first-order difference operation on the status time-series data to obtain differential time-series data, performs a sliding window framing operation on the differential time-series data, and calculates the nonlinear fluctuation index of the differential time-series data in the current time window in real time. When the nonlinear fluctuation index is less than or equal to a preset safety threshold, the power battery pack is in a stable operating condition in the current time window, and a corresponding stable data identifier is generated. Otherwise, the power battery pack is in a complex operating condition in the current time window, and an abnormal data identifier is generated. The linear regression module is used to perform steady-state data linear regression compression based on algebraic iteration on the differential time series data within the current time window when the power battery pack is in a stable operating condition within the current time window. This results in a univariate linear regression model of the differential time series data within the current time window. The optimization objective is to minimize the mean square error. A loss function is defined for all coefficients in the univariate linear regression model. All coefficients in the univariate linear regression model are iteratively updated using the expanded partial derivative formula. All coefficients in the iteratively updated univariate linear regression model and their corresponding stable data identifiers are packaged to obtain a stable data package, which is then uploaded to the cloud server. The feature extraction module is used to input the differential time series data in the current time window into the autoencoder model and perform forward propagation operation when the power battery pack is in a complex working condition in the current time window. It obtains the latent space feature vector corresponding to the differential time series data in the current time window, packages the latent space feature vector and the corresponding abnormal data identifier to obtain an abnormal data packet, and uploads the abnormal data packet to the cloud server. The cloud-based data processing module performs linear data segment reconstruction on stable data packets, converting them into differential time-series data within the corresponding time window. It also performs deep decoding reconstruction on abnormal data packets, restoring them to differential time-series data within the corresponding time window. Finally, it concatenates the differential time-series data within the time window of the stable data packets with the differential time-series data within the time window of the abnormal data packets in chronological order to obtain complete timing data of the power battery pack status.

2. The battery data hybrid compression and reconstruction system based on edge-cloud collaboration according to claim 1, characterized in that: Methods for performing first-order difference operations on the state-time data of a power battery pack include: ； in, For the state timing data of the power battery pack The output value of the differential time series data after differential processing. for The raw physical quantity values ​​of the power battery pack status data collected at all times. The start time of the current data transmission frame The original physical quantity values ​​of the power battery pack state data collected at the previous sampling time t are: At that time, directly retain the original value. This serves as the baseline anchor point data for the current frame, used for initial positioning during cloud reconstruction.

3. The battery data hybrid compression and reconstruction system based on edge-cloud collaboration according to claim 2, characterized in that: The method for performing sliding window framing on the differential time series data and calculating the nonlinear fluctuation index of the differential time series data within the current time window in real time includes: Construct a sliding time window of length N in chronological order, denoted as Where k is the window number and N is the preset window length, and a preset overlap rate is designed between adjacent time windows. For the differential time series data within the current time window, a virtual linear baseline connecting the first and last data points of the sliding time window is constructed: the theoretical projection value corresponding to the virtual linear baseline is calculated based on the difference between the first and last differential time series data within the current time window. The theoretical projection values ​​satisfy: ； Based on the theoretical projection values ​​corresponding to the virtual linear baseline, the mean square error algorithm is used to calculate the deviation of the differential time series data within the current time window from the virtual linear baseline: ； in, The first time within the current sliding time window The theoretical projection values ​​of each differential time series data point onto a virtual linear baseline. This is the first differential time series data value within the current sliding time window. This is the second differential time series data value within the current sliding time window. Here, i = 1, 2, ..., N, represents the position index within the time window, corresponding to the 1st to Nth data point within the current sliding time window. This is the last differential time series data value within the current sliding time window. This is the nonlinear fluctuation index of the differential time series data within the current time window. The first time within the current sliding time window Differential time series data.

4. The battery data hybrid compression and reconstruction system based on edge-cloud collaboration according to claim 3, characterized in that: When the power battery pack is in a stable operating condition within the current time window, the method for performing steady-state data linear regression compression based on algebraic iteration on the differential time series data within the current time window to obtain a univariate linear regression model of the differential time series data within the current time window includes: For the current time window differential sequence within Establish a univariate linear regression model: ； Represents the relative time index within the time window. ; a is the slope coefficient to be solved, which represents the changing trend of the differential time series data within the current time window; b is the intercept coefficient to be solved, which represents the baseline offset of the differential time series data within the current time window. These are the predicted values ​​from a univariate linear regression model. Using the minimization of mean squared error as the optimization objective, a loss function is defined for all coefficients in the univariate linear regression model. The methods include: .

5. A battery data hybrid compression and reconstruction system based on edge-cloud collaboration as described in claim 3 or 4, characterized in that: The method of iteratively updating all coefficients in a univariate linear regression model using the expanded partial derivative formula, and then packaging all coefficients in the iteratively updated univariate linear regression model and their corresponding stable data identifiers to obtain a stable data package includes: Update intercept coefficient : ； Update slope coefficient : ； Where a is the slope coefficient value of the current iteration step, and b is the intercept coefficient value of the current iteration step. The updated slope coefficient value. These are the updated intercept coefficient values. The learning rate is preset based on the calculation cycle of the BMS system. When the number of iterations reaches the preset convergence condition, the iteration stops. The updated slope coefficient and intercept coefficient obtained at the time of stopping the iteration are combined with the stable data identifier corresponding to the current time window, the start time of the current time window, and the length of the current time window in a preset order to obtain a stable data packet.

6. A battery data hybrid compression and reconstruction system based on edge-cloud collaboration according to claim 1 or 3, characterized in that: When the power battery pack is under complex operating conditions within the current time window, the differential time series data within the current time window is input into the autoencoder model and a forward propagation operation is performed to obtain the latent space feature vector corresponding to the differential time series data within the current time window. The method of packaging the latent space feature vector and the corresponding abnormal data identifier to obtain an abnormal data package includes: taking the differential sequence within the current sliding time window of length N... After normalization, the data is input into the Encoder model and the forward propagation operation is performed: ； in, This is an encoder model that has undergone int8 quantization. The output latent space feature vector The dimension; To create a latent space feature vector that highly condenses the nonlinear characteristics of fault waveforms, For the encoder model, the difference sequence within the current sliding time window The extracted and compressed feature vector of the first data point in the innermost region. For the encoder model, the difference sequence within the current sliding time window The extracted and compressed feature vector of the second data point. For the encoder model, the difference sequence within the current sliding time window The extracted and compressed feature vector of the m-th data point is combined with the latent space feature vector, the abnormal data identifier corresponding to the current time window, and the start time of the current time window in a preset order to obtain the abnormal data packet.

7. A battery data hybrid compression and reconstruction system based on edge-cloud collaboration according to claim 5, characterized in that: The method for performing linear data segment reconstruction on stable data packets to convert them into differential time-series data within the corresponding time window includes: a linear generator on a cloud server generating a cloud-based differential time-series data sequence within the current time window using linear equations based on the updated slope coefficient value, intercept coefficient value, current time window start time, and current time window length carried in the stable data packets. The formula is as follows: ； in, This represents the prediction value of the univariate linear model for the first difference time series data within the current sliding time window. This represents the prediction value of the second difference time series data within the current sliding time window by the univariate linear model. For a univariate linear model, the first _ ... Predicted values ​​of differential time series data The length of the current time window; After reconstruction, the differential time series data in the cloud differential time series data sequence are accumulated and restored to state time series data: the initial state data of the power battery pack corresponding to the start time of the current time window is added to the prediction value of the first differential time series data in the current sliding time window by the univariate linear model in the cloud differential time series data sequence. The first state time series data is obtained by adding the first state time series data to the prediction value of the second difference time series data within the current sliding time window by the univariate linear model. The corresponding second state time series data is obtained by adding the second state time series data to the prediction value of the third difference time series data within the current sliding time window by the univariate linear model. The corresponding third state time series data, and so on, until the state time series data corresponding to all differential time series data in the cloud differential time series data sequence are obtained.

8. A battery data hybrid compression and reconstruction system based on edge-cloud collaboration according to claim 6, characterized in that: The method for performing deep decoding and reconstruction on abnormal data packets to restore them to differential time-series data within the corresponding time window includes: using the decoder built into the cloud server to extract the received latent space feature vector based on the start time of the current time window in the abnormal data packet. Mapping back to a high-dimensional time series space, differential time series data within the corresponding time window, including nonlinear details, are reconstructed. : ； in, This is a full-precision deep neural network decoder running on a cloud server.

9. A method for hybrid compression and reconstruction of battery data based on end-to-cloud collaboration according to the system of claim 1, characterized in that, include: The vehicle-mounted battery management system acquires the state time-series data of the power battery pack in real time according to a preset sampling frequency. First-order difference operation is performed on the state time-series data to obtain differential time-series data. Sliding window framing operation is performed on the differential time-series data, and the nonlinear fluctuation index of the differential time-series data within the current time window is calculated in real time. When the nonlinear fluctuation index is less than or equal to a preset safety threshold, the power battery pack is in a stable operating condition within the current time window, and a corresponding stable data identifier is generated. Otherwise, the power battery pack is in a complex operating condition within the current time window, and an abnormal data identifier is generated. When the power battery pack is in a stable operating condition within the current time window, a linear regression compression of steady-state data based on algebraic iteration is performed on the differential time series data within the current time window to obtain a univariate linear regression model of the differential time series data within the current time window. The mean square error is minimized as the optimization objective. A loss function is defined for all coefficients in the univariate linear regression model. All coefficients in the univariate linear regression model are iteratively updated using the expanded partial derivative formula. All coefficients in the iteratively updated univariate linear regression model and their corresponding stable data identifiers are packaged to obtain a stable data package, which is then uploaded to the cloud server. When the power battery pack is under complex operating conditions within the current time window, the differential time series data within the current time window is input into the automatic encoder model and forward propagation is performed to obtain the latent space feature vector corresponding to the differential time series data within the current time window. The latent space feature vector and the corresponding abnormal data identifier are packaged to obtain an abnormal data packet, and the abnormal data packet is uploaded to the cloud server. For stable data packets, a linear data segment reconstruction operation is performed to convert the stable data packets into differential time-series data within the corresponding time window. For abnormal data packets, a deep decoding reconstruction operation is performed to restore the abnormal data packets into differential time-series data within the corresponding time window. Then, the differential time-series data within the time window of the stable data packets and the differential time-series data within the time window of the abnormal data packets are concatenated in chronological order to obtain the complete timing data of the power battery pack status.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method of claim 9.

Images