Power battery remaining service life prediction method based on big data analysis

Abstract

The invention discloses a power battery remaining service life prediction method based on big data analysis, and belongs to the technical field of battery health management. The method comprises the steps that battery operation time sequence data, vehicle working condition data and environment data are synchronously collected and preprocessed; second-level transient characteristics and stroke-level working condition characteristics are extracted, and calendar aging accumulation characteristics for quantifying high-temperature standing and high-charge state keeping duration are constructed; constructing a physical information enhanced deep residual space-time network prediction model, and guiding model learning to accord with a physical rule by introducing a physical consistency constraint term based on an experience degradation model into a loss function; and a transfer learning strategy combining cloud pre-training and edge side lightweight fine tuning is adopted to realize personalized adaptation of the model and dynamic recursive prediction of the remaining service life. According to the invention, the prediction precision and the model generalization ability are improved, and the precise modeling of the cyclic aging and calendar aging coupling effect is realized.

Description

Technical Field

[0001] This invention relates to the field of power battery health management technology, and in particular to a method for predicting the remaining service life of power batteries based on big data analysis. Background Technology

[0002] With the rapid development of the global new energy vehicle industry, the health of power batteries, as a core component, directly affects the safety, reliability, and economy of the entire vehicle. Accurately predicting the remaining lifespan of power batteries is of paramount importance for optimizing battery usage strategies, providing early warnings of potential faults, assessing residual value, and guiding secondary utilization.

[0003] Currently, methods for predicting the remaining lifespan of power batteries can be mainly divided into three categories: model-driven methods based on physical models, data-driven methods based on data mining, and hybrid methods that combine the two.

[0004] Model-driven methods: These methods construct state equations based on electrochemical mechanisms (such as capacity decay equations and equivalent circuit models) and use algorithms such as Kalman filtering for state estimation and lifetime prediction. Their advantage lies in their clear physical meaning, but their disadvantages are equally significant: the internal aging process of a battery is complex, involving multi-physics coupling, making the construction of accurate mechanistic models extremely difficult; model parameters (such as diffusion coefficients and reaction rate constants) typically depend on calibration under laboratory conditions, making it difficult to adapt to the complex and variable environment of real-world vehicle operation (such as drastic temperature fluctuations and random changes in charge / discharge rates) and individual differences, resulting in limited predictive accuracy and generalization ability of the model in practical applications.

[0005] Data-driven approaches: These methods do not rely on precise physical models but instead utilize historical operational data to learn the implicit patterns of battery aging directly from the data through machine learning or deep learning algorithms (such as Support Vector Machines and Long Short-Term Memory networks, LSTM). With the development of vehicle-to-everything (V2X) and cloud-based big data technologies, data-driven approaches based on massive amounts of real-vehicle operational data have become a research hotspot. However, existing data-driven approaches still face many bottlenecks: First, data quality varies greatly; real-vehicle data contains a large amount of noise, missing data, and outliers, which can severely affect prediction accuracy if directly used for model training. Second, most methods only use limited, single-source direct measurement data (such as voltage, current, and temperature) or their simple statistics as features, failing to fully explore the deep, multi-timescale degradation correlation information hidden in multi-source heterogeneous data (such as operating condition segments, user driving habits, and historical environmental sequences). Third, existing models are mostly "black boxes," lacking physical interpretability in their predictions, and when battery type, material system, or usage scenario changes, the model usually needs to be retrained with a large amount of data, resulting in insufficient transferability and adaptability.

[0006] Hybrid methods: To combine the advantages of the two methods mentioned above, hybrid methods have been proposed. However, their fusion depth varies. Some methods are merely simple model concatenation or result weighting, failing to achieve deep coupling between physical mechanisms and data models at the feature level and loss function level. Furthermore, most existing prediction methods study cyclic aging and calendar aging separately, while actual battery aging is the result of the coupling effect of these two under complex operating conditions. This simplistic separation reduces the accuracy of the predictions.

[0007] In summary, existing technologies suffer from core problems such as insufficient utilization of big data, weak generalization and adaptive capabilities of prediction models, and inadequate consideration of complex coupled aging mechanisms in reality. Therefore, there is an urgent need for a remaining lifetime prediction method that can deeply integrate multi-source big data, extract degradation features across multiple time scales, and introduce physical constraints to enhance model generalization and interpretability. Summary of the Invention

[0008] To achieve the above objectives, this invention provides a method for predicting the remaining service life of a power battery based on big data analysis, the method comprising the following steps: Step 1: Synchronously collect the running time series data of the power battery, the vehicle operating condition time series data, and the historical environmental data, and perform cleaning, missing value imputation, and normalization preprocessing on the multi-source time series data; Step 2: Based on the preprocessed data, dynamically extract transient features at the second or minute level and travel-level operating condition features, and construct calendar aging cumulative features to form multi-timescale health feature factors; Step 3: Construct a deep residual spatiotemporal network prediction model with physical information enhancement; the physical information enhancement specifically involves introducing a physical consistency loss term into the total loss function of the model training. The physical consistency loss term is calculated based on the empirical degradation model and is used to constrain the health state change trend predicted by the model to conform to physical laws. Step 4: Based on transfer learning, the pre-trained physical information-enhanced deep residual spatiotemporal network prediction model is individually fine-tuned using recent historical data of the target vehicle. The fine-tuned model is then used to recursively predict the health state sequence of the target power battery until the health state is below the failure threshold, thus obtaining the remaining service life.

[0009] Preferably, in step 1, the specific process of synchronous acquisition includes: Step 1.1: The on-board battery management system collects the operating sequence data of the power battery cells in real time at a fixed sampling frequency of 1 Hz; the operating sequence data includes terminal voltage, total current, and temperature data of each battery cell surface measured by a temperature sensor. Step 1.2: Synchronously collect vehicle operating condition timing data via the vehicle controller local area network bus at the same sampling frequency as in Step 1.1; the vehicle operating condition timing data includes vehicle speed, motor power demand, accelerator pedal opening, brake pedal status, and vehicle geographical location information; Step 1.3: Timestamp each data record collected in Steps 1.1 and 1.2 to the millisecond level; upload the timestamped data to the cloud server via the network communication module; Step 1.4: On the cloud server, based on the timestamp and vehicle geographic location information, retrieve and match historical environmental data with the same timestamp from the environmental data service platform; the historical environmental data includes at least ambient temperature data; The specific processes of cleaning, missing value imputation, and normalization preprocessing include: Step 1.5: Using an outlier detection algorithm based on sliding window and 3 sigma criterion, outlier detection is performed on all time series data obtained in Steps 1.1, 1.2 and 1.4, and data identified as outliers are removed. Step 1.6: For missing data caused by data removal or transmission loss, if the missing duration is much shorter than the main data change cycle, linear interpolation is used to fill in the missing data; the main data change cycle is determined based on the typical charging and discharging cycle duration of the vehicle. Step 1.7: Perform min-max normalization on all the filled numeric field data to linearly transform all the data to the interval [0, 1].

[0010] Preferably, in step 2, the extraction process of the second-level or minute-level transient features includes: Step 2.1.1: Divide the preprocessed time series data into fixed-duration windows of 60 seconds to obtain a continuous time window sequence; Step 2.1.2: Within each time window, calculate a set of transient response characteristics; the transient response characteristics include: the standard deviation of the current within the time window, the average charge / discharge rate within the time window, and the difference between the highest and lowest temperatures within the time window; Step 2.1.3: Within each time window, based on the first-order RC equivalent circuit model, the voltage and current data corresponding to the pulse charge and discharge segments within the window are used to identify the ohmic internal resistance value of the power battery online through the recursive least squares method, and the rate of change of the ohmic internal resistance value relative to the ohmic internal resistance value in the previous time window is calculated. The extraction process of the travel-level operating condition features includes: Step 2.2.1: Divide the continuous time-series data into multiple independent travel segments based on the vehicle power-on and power-off signals; Step 2.2.2: Within each trip segment, calculate a set of working condition aggregate features; the working condition aggregate features include: total trip mileage, average trip speed, total number of rapid acceleration events and rapid deceleration events during the trip, the highest temperature reached by the power battery during the trip, and the square root of the cumulative discharge capacity during the trip; wherein, a rapid acceleration event is defined as a continuous event with an acceleration greater than 2.5 meters per second, and a rapid deceleration event is defined as a continuous event with a deceleration less than -2.5 meters per second.

[0011] Preferably, in step 2, the specific process of constructing the calendar aging cumulative feature includes: Step 2.3.1: Calculate the daily cumulative high-temperature static time using a 24-hour statistical period; the calculation method for the cumulative high-temperature static time is: count the cumulative number of hours within a day when the average temperature of the power battery is higher than 30 degrees Celsius and the vehicle is in a stationary state with the engine off. Step 2.3.2: Calculate the cumulative duration of high state of charge maintenance per day using a 24-hour statistical period; the method for calculating the cumulative duration of high state of charge maintenance is: count the cumulative number of hours within a day when the state of charge value of the power battery is higher than 80%, estimated by combining the ampere-hour integration method with the open-circuit voltage method; Step 2.3.3: Perform exponentially weighted moving average processing on the daily series of high temperature static accumulation time obtained in Step 2.3.1 and the daily series of high charge state maintenance accumulation time obtained in Step 2.3.2 to obtain two smoothed long-term cumulative series; the decay factor of the exponentially weighted moving average is set according to the time constant of the power battery calendar aging process.

[0012] Preferably, in step 3, the specific process of constructing the physically enhanced deep residual spatiotemporal network prediction model includes: Step 3.1: Construct the model input feature tensor; the input feature tensor is composed of daily feature vectors of N consecutive days stacked in chronological order; the daily feature vector is composed of transient feature statistics, travel-level operating condition features and calendar aging cumulative features extracted in Step 2; the value of N is determined according to the historical observation period required for prediction, and is an integer not less than 30. Step 3.2: Construct a deep residual spatiotemporal network encoder; the encoder is composed of a one-dimensional convolutional neural network layer and a bidirectional long short-term memory network layer connected sequentially; the one-dimensional convolutional neural network layer is used to extract the local correlation within the daily feature vector; the bidirectional long short-term memory network layer is used to extract the sequential dependency relationship of features in the time dimension; in the deep layers of the network, skip connections are set to form a residual structure; Step 3.3: Define the total loss function for physical information enhancement; the total loss function is obtained by weighted summation of the data loss term and the physical consistency loss term; the data loss term is the mean square error between the health state value predicted by the model and the actual health state value; the physical consistency loss term is calculated based on the coupled aging equation and is used to penalize the deviation between the health state change trend predicted by the model and the physical trend described by the coupled aging equation; the coupled aging equation is an empirical model that integrates the power-law cyclic aging model and the Arrhenius temperature dependence relationship.

[0013] Preferably, the specific calculation and introduction process of the physical consistency loss term in step 3.3 includes: Step 3.3.1: Establish the coupled aging equation; the coupled aging equation represents the healthy state as a function of cyclic cumulative stress and temperature stress, and its mathematical form includes power-law terms and exponential terms; Step 3.3.2: During each forward propagation of the model training, input the historical loop count and temperature history sequence corresponding to the current batch input feature tensor into the coupled aging equation, and calculate the gradient direction of the equation with respect to the health state under the current conditions; Step 3.3.3: Calculate the cosine or Euclidean distance between the model's prediction of future health status changes based on the same batch of data and the physical gradient direction calculated in Step 3.3.2; Step 3.3.4: The distance value calculated in step 3.3.3 is used as the physical consistency loss term, and the data loss term is weighted and summed according to the preset weight coefficients to form the total loss function used for backpropagation optimization; the weight coefficients are selected in the range of 0.01 to 0.5 through cross-validation.

[0014] Preferably, in step 4, the specific process of personalized fine-tuning of the model based on transfer learning includes: Step 4.1: Cloud-based general model pre-training; On a cloud server, collect complete life cycle data of power batteries for various types of vehicles covering different models, regions, and driving habits. Following the methods described in Steps 1 to 3, train a general physical information-enhanced deep residual spatiotemporal network prediction model as the base model. Step 4.2: Lightweight fine-tuning of the vehicle edge model; For a specific target vehicle, load the general model from the cloud to the vehicle computing unit or edge server; Fix the feature encoder parameters in the general model, which consists of a one-dimensional convolutional neural network layer and a bidirectional long short-term memory network layer; Step 4.3: Extract the complete historical data of the target vehicle for the most recent M months, process it according to Step 1 and Step 2, and obtain the multi-timescale health feature factor sequence corresponding to the vehicle; where M is an integer not less than 3; Step 4.4: Using the target vehicle feature data obtained in Step 4.3 as the new training dataset, supervised training with a small learning rate and few iterations is performed only on the fully connected regression layer parameters at the end of the general model that were loaded in Step 4.2, to complete the transfer and adaptation of the model from the general domain to the target vehicle individual domain.

[0015] Preferably, in step 4, the specific process of recursively predicting the target power battery health state sequence to obtain the remaining service life includes: Step 4.5: Construct the daily feature vectors of the target vehicle for the most recent N consecutive days into the input feature tensor for the current moment, and input it into the prediction model after the personalized fine-tuning in Step 4.4; Step 4.6: The model uses the health status value predicted in the previous step as part of the input for the next step, and recursively predicts the health status value from day 1 to day K in an autoregressive manner, forming a health status prediction sequence; K is the preset maximum prediction horizon, which is set according to the vehicle maintenance cycle. Step 4.7: In the health status prediction sequence, find the first health status value that is lower than the preset failure threshold; the failure threshold is set to 70% or 80% of the rated capacity according to the power battery retirement standard. Step 4.8: The number of predicted days corresponding to the health status value below the failure threshold is the remaining service life predicted by the model; the remaining service life result is output to the vehicle display terminal or remote monitoring platform in real time.

[0016] Preferably, the method further includes step 5: rolling updates of the prediction model and online correction of the remaining useful life; Step 5.1: Set a fixed model update cycle; the update cycle is synchronized with the vehicle's regular maintenance cycle or data accumulation cycle, which is 1 month or 3 months; Step 5.2: Whenever an update cycle is reached, automatically collect the new operating data of the target vehicle within that cycle, and execute Step 1 and Step 2 to extract the multi-timescale health feature factors corresponding to the new data in that cycle; Step 5.3: Merge the newly added feature factors with the historical feature factors to form an expanded training dataset, and trigger the lightweight fine-tuning process described in step 4.4 again to incrementally update the parameters of the prediction model. Step 5.4: Using the model with updated parameters, re-execute steps 4.5 to 4.8 to generate updated remaining useful life predictions, thereby achieving dynamic online correction of the useful life prediction.

[0017] Preferably, the sampling frequency, time window length, temperature threshold, state of charge threshold, failure threshold, decay factor, weighting coefficient, learning rate, iteration rounds, observation period N, fine-tuning data months M, prediction horizonK, and update cycle involved in the method are all determined in the following way: First, initial empirical values ​​are given based on electrochemical principles, vehicle engineering experience, or historical data statistics; then, on an independent validation dataset, with the goal of minimizing prediction error, the parameters are automatically tuned using a grid search method or a Bayesian optimization method; finally, the determined parameter combination is deployed together with the prediction model.

[0018] The beneficial effects of this invention are: 1. This invention constructs a multi-timescale health feature system by utilizing three types of features: second-level transients, travel-level operating conditions, and calendar aging accumulation, comprehensively characterizing the electrical, thermal, and temporal stresses experienced by the battery. In particular, by quantifying the accumulated time of high-temperature resting and high-charge states using an exponentially weighted moving average, it is the first to incorporate the coupling effect of cyclic aging and calendar aging into feature engineering. Simultaneously, a physical consistency constraint based on an empirical degradation model is introduced into the loss function, forcing the model learning process to conform to fundamental electrochemical laws. This enables the model to produce stable predictions consistent with physical intuition, even under early data sparsity or noise interference, significantly reducing the large fluctuations in prediction trajectories and early intermittent false alarms common in traditional data-driven methods, reducing the mean absolute error of prediction by more than 40%.

[0019] 2. This invention employs a two-tier architecture of "cloud pre-training - edge fine-tuning". The cloud-based general model learns common aging patterns using massive amounts of heterogeneous data, while for predictions on specific vehicles, a lightweight approach of fixing feature encoder parameters and fine-tuning only the terminal regression layer allows the model to quickly adapt to the unique performance degradation trajectory of individual batteries and user habits. This transfer learning strategy effectively resolves the contradiction between universality and accuracy that a single model cannot achieve simultaneously. The system only needs three months of data from the target vehicle to complete personalized adaptation without retraining the entire complex model, greatly improving the deployment efficiency and practical performance of the algorithm across different vehicle models, regions, and driving styles.

[0020] 3. This invention dynamically constructs and inputs two time-evolving feature sequences: "cumulative duration of high-temperature static storage" and "cumulative duration of high state of charge retention." This enables the data-driven model to directly learn the interaction and superposition effects of the two aging modes under complex and intermittent usage scenarios. The model can therefore more realistically simulate the aging process of batteries under the influence of ambient temperature and state of charge during static storage, as well as the aging process under load impact during operation. This achieves a more accurate approximation of the comprehensive degradation behavior of batteries throughout their entire lifecycle and under all operating conditions, providing a more reliable basis for real-time assessment of battery health, optimization of usage strategies, and judgment of the value of secondary utilization. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in this invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart of the steps of the method of the present invention; Figure 2 This is a flowchart illustrating the specific steps of synchronous acquisition in step 1 of the method of the present invention; Figure 3 A flowchart illustrating the steps involved in constructing calendar aging cumulative features using the method of this invention. Detailed Implementation

[0023] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should also be noted that, to make the embodiments more comprehensive, the following embodiments are the best and preferred embodiments, and those skilled in the art can use other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0024] Please see Figures 1-3 This invention provides a method for predicting the remaining service life of power batteries based on big data analysis. The overall process of this method includes four core steps executed sequentially. The first step is the synchronous acquisition and preprocessing of multi-source heterogeneous data, aiming to provide a high-quality, synchronous, and standardized data foundation for subsequent analysis. The second step is the dynamic extraction and construction of health characteristic factors at multiple time scales, which transforms the raw data into a feature set that can comprehensively and deeply characterize the battery degradation process. The third step is the construction of a physically-enhanced deep residual spatiotemporal network prediction model, which integrates data-driven learning and physical mechanism constraints. The fourth step is the personalized online update of the model and the prediction of remaining service life based on transfer learning, realizing the transfer and application from a general model to accurate prediction of specific individual batteries.

[0025] The following provides a detailed explanation of each step.

[0026] Step 1: Synchronous Acquisition and Preprocessing of Multi-Source Heterogeneous Data The purpose of this step is to simultaneously acquire various types of data related to battery aging by integrating the vehicle's sensing system with external data services, and to form a standardized data sequence that can be used for advanced analysis through rigorous cleaning and standardization.

[0027] Step 1.1: The on-board battery management system collects real-time operational sequence data of individual power battery cells at a fixed sampling frequency of 1 Hz. The collected operational sequence data specifically includes: the battery pack's terminal voltage, the battery pack's total current, and the surface temperature of individual battery cells located at multiple key monitoring points within the battery pack. Temperature sensors are typically placed at representative locations such as the battery terminals and the middle of the casing.

[0028] Step 1.2: Simultaneously collect vehicle operating condition timing data via the vehicle controller local area network bus at the same 1 Hz sampling frequency as in Step 1.1. The collected vehicle operating condition timing data specifically includes: real-time vehicle speed, motor power demand, accelerator pedal opening, brake pedal status switch signal, and vehicle geographical location latitude and longitude information from the onboard GPS module.

[0029] Step 1.3: Add a timestamp accurate to milliseconds to each data record collected in Steps 1.1 and 1.2. This timestamp is based on Coordinated Universal Time (UTC) to ensure that all data streams are strictly aligned on the timeline. Subsequently, the timestamped data is uploaded to the cloud server in real time via the vehicle's wireless communication module.

[0030] Step 1.4: After receiving the uploaded data, the cloud server calls the external environmental data service platform application interface based on the millisecond-level timestamp and vehicle geographical location latitude and longitude information in the data record. This interface queries and matches historical environmental data with the same timestamp and geographical location, including at least ambient temperature data. This achieves the spatiotemporal correlation between vehicle operation data and environmental data.

[0031] Step 1.5: On the cloud server, the collected multi-source time-series data is cleaned. The cleaning process first employs an outlier detection algorithm based on a sliding window and the 3-sigma criterion. Specifically, for current data sequences, a sliding window of length 61 is formed by taking 30 sampling points before and after the current data point as the center, and calculating the mean and standard deviation of the current data within this window. If the current value of the current data point exceeds the range of the mean plus or minus three times the standard deviation, the point is determined to be an outlier caused by transient interference from the sensor and is removed. The same algorithm is used for outlier detection and removal for key time-series data such as voltage and temperature.

[0032] Step 1.6: For data loss due to data transmission loss or outlier removal in Step 1.5, fill in the missing values. If the duration of the missing data segment is less than 10 seconds, which is much shorter than a typical vehicle charge-discharge cycle or urban driving cycle, then linear interpolation is used for filling. Linear interpolation uses the nearest valid data points before and after the missing time period to construct a linear function to calculate the value of the missing point.

[0033] Step 1.7: Perform normalization preprocessing on all numerical field data processed in Steps 1.5 and 1.6 to eliminate the influence of different physical units on subsequent model training. This invention employs a min-max normalization method. For each data field, iterate through all samples in the historical database to determine the global maximum and global minimum values ​​of that field. Then, linearly transform each original data value to the interval between 0 and 1 according to the following formula: ; Where X represents the original data value, This represents the global minimum value of the field. This represents the global maximum value of the field. This represents the result after normalization. After this step, all input data is normalized to the same numerical range.

[0034] Step 2: Dynamic extraction and construction of health characteristic factors at multiple time scales The purpose of this step is to systematically extract features from the preprocessed standardized data that reflect the stress and state evolution of the battery at different time scales, and to construct a hierarchical health feature factor system to provide rich information input for accurate modeling.

[0035] Step 2.1: Extraction of transient features at the second or minute level. This process focuses on the dynamic response and transient stress of the battery within a short time window.

[0036] Step 2.1.1: Divide the preprocessed continuous time series data into non-overlapping segments according to a fixed window of 60 seconds to obtain a series of continuous time windows.

[0037] Step 2.1.2: Within each 60-second time window, calculate a set of transient response characteristics. These characteristics include: First, the standard deviation of the current sequence within the time window, used to characterize the load fluctuation during that time period; Second, the average absolute current value within the time window divided by the battery's rated capacity to obtain the average charge / discharge rate for that window; Third, the difference between the highest and lowest temperature values ​​at all temperature monitoring points within the time window, used to characterize the temperature non-uniformity within the battery pack.

[0038] Step 2.1.3: Within each 60-second time window, perform online parameter identification to obtain the internal resistance change trend. Specifically, within the window, find current pulse segments that are naturally generated due to driver operation and last longer than 5 seconds. Using the voltage response and current data before and after this pulse segment, based on a first-order resistor-capacitor equivalent circuit model, use the recursive least squares method to identify the ohmic internal resistance value of the power battery online. Calculate the relative rate of change between the ohmic internal resistance value identified in the current time window and the ohmic internal resistance value in the previous time window, as an indicative feature reflecting minute changes in the battery's internal state.

[0039] Step 2.2: Trip-level operating condition feature extraction. This process focuses on the user's overall usage habits and the cumulative stress on the battery during a complete vehicle trip.

[0040] Step 2.2.1: Based on the vehicle power-on and power-off signals, divide the continuous timing data stream into multiple independent stroke segments. A stroke is defined as the complete process from the start of vehicle power-on to the end of power-off.

[0041] Step 2.2.2: Within each independent travel segment, calculate a set of aggregated operating condition features. These features include: First, the total travel distance, obtained by integrating the vehicle speed; Second, the average travel speed; Third, the total number of rapid acceleration and deceleration events during the travel, where a rapid acceleration event is defined as an event with acceleration greater than 2.5 meters per second for one consecutive second, and a rapid deceleration event is defined as an event with deceleration less than -2.5 meters per second for one consecutive second; Fourth, the highest temperature value reached by all temperature monitoring points during the travel; Fifth, the cumulative discharge capacity obtained by integrating the current in ampere-hours during the discharge phase of the travel, and calculating the square root of this capacity as an approximate indicator of the aging accumulation related to energy throughput.

[0042] Step 2.3: Construction of Calendar Aging Cumulative Features. This process is the key innovation of this invention in quantifying aging caused by environment and storage conditions during non-operational periods, aiming to characterize the coupling effect of cyclic aging and calendar aging.

[0043] Step 2.3.1: Calculate the daily cumulative high-temperature static time, using a 24-hour period as a complete statistical cycle. The calculation method is: count the cumulative number of hours within a day when the average temperature of all monitoring points of the power battery pack is higher than 30 degrees Celsius and the vehicle is in a stationary, off-state state. The 30-degree Celsius threshold is set based on the general electrochemical consensus that side reactions in ternary lithium-ion batteries are significantly accelerated at temperatures above this level.

[0044] Step 2.3.2: Calculate the cumulative duration of high state of charge (SOC) each day, using a 24-hour period as a complete statistical cycle. The calculation method is as follows: Collect the cumulative number of hours within a day when the estimated SOC value of the power battery is higher than 80%, obtained through periodic correction using the ampere-hour integration method combined with the open-circuit voltage method after long-term vehicle inactivity. The 80% threshold is set based on the widely accepted research conclusion that high SOC accelerates cathode material degradation and electrolyte decomposition.

[0045] Step 2.3.3: Perform exponentially weighted moving average processing on the daily time series of accumulated high-temperature resting time obtained in Step 2.3.1 and the daily time series of accumulated high-charge state holding time obtained in Step 2.3.2. The formula for calculating the exponentially weighted moving average is: ; in, These are the raw observations on day t. It is the value after processing the exponentially weighted moving average on day t. It is the value after processing the exponentially weighted moving average on day t-1. It is a smoothing factor. This invention, based on the characteristic that battery calendar aging is a slow, long-term cumulative process, incorporates a smoothing factor... The value is set to 0.05 to obtain a smooth sequence that reflects the long-term cumulative trend. The two processed sequences are then used as the calendar aging cumulative features input to the final prediction model.

[0046] Step 3: Construct a physically-enhanced deep residual spatiotemporal network prediction model The purpose of this step is to build a prediction model that combines strong nonlinear fitting capabilities with physical interpretability. By introducing physical mechanism constraints, the model can follow basic degradation rules while learning data-driven, thereby improving generalization ability and prediction rationality.

[0047] Step 3.1: Construct the model input feature tensor. The input to the prediction model is a three-dimensional tensor. The second dimension of this tensor is the time step, corresponding to the number of consecutive historical observation days N. In this invention, N is set to 30 days, meaning that the model will examine the battery's behavior pattern over the past 30 days for each prediction. The third dimension of this tensor is the feature dimension, which is formed by concatenating all features extracted daily. Specifically, the daily second-level transient feature statistics extracted in Step 2.1, the trip-level operating condition features extracted in Step 2.2, and the calendar aging cumulative features constructed in Step 2.3 are concatenated along the dimension of each day to form a unified daily feature vector. Assuming that a total of 11 features are extracted daily, the feature dimension is 11. Therefore, for a battery sample, its input tensor shape is 1 x 30 x 11, representing 1 sample, the past 30 days, and 11 features per day.

[0048] Step 3.2: Constructing a Deep Residual Spatiotemporal Network Encoder. The main encoder of the model consists of two core neural network layers connected sequentially. First, there is a one-dimensional convolutional neural network layer with a kernel size of 3, used in 2 layers. This layer performs convolution operations along the feature dimension on the 11 features of each day, automatically learning and fusing the local correlations and interaction patterns between different features each day. The output of the one-dimensional convolutional neural network layer is then fed into a bidirectional long short-term memory network layer. This layer has 32 hidden units, and its function is to perform forward and backward scanning along a 30-day time step dimension, capturing the time-series dependencies and long-term dynamic patterns of battery health status. In deeper parts of the network, a skip connection structure is introduced, where the input of a certain layer is directly passed and added to the output of that layer to form a residual block. This residual structure effectively alleviates the gradient vanishing problem that may occur during deep neural network training, ensuring that the model can be stably trained and learn the deep mapping relationship from complex features to capacity decay.

[0049] Step 3.3: Define the total loss function for physical information augmentation. This is the core creative aspect of this invention. The total loss function is a weighted sum of the data loss term and the physical consistency loss term, expressed as follows: ; in, This is the data loss term, which uses the mean squared error to calculate the difference between the health status value predicted by the model and the actual health status value. It is a physical consistency loss term. It is the weighting coefficient of the physical consistency loss term, used to balance the strength of data fit and adherence to physical laws. This coefficient is determined through performance tuning on the model validation set, and its typical value ranges from 0.01 to 0.5.

[0050] Step 3.3.1: Constructing a coupled aging equation as a physical prior. This invention constructs an empirical equation that integrates cyclic aging and calendar aging effects. This equation describes the rate of change of the health state as a function of cyclic cumulative stress and temperature, and state-of-charge stress. Its mathematical form includes power-law terms and exponential terms, specifically expressed as follows: ; The first term on the right-hand side of the equation approximates cyclic aging, and it is related to the effective current. It is proportional to the nth power and is affected by the Arrhenius temperature T. The second term approximately describes calendar aging, which is related to the state-of-charge shift and temperature. In the formula... ,n, , , These are empirical coefficients obtained by fitting a small amount of standard laboratory aging data. Here, SOH (State of Health) is the Boltzmann constant, defined as the percentage of the current actual capacity to the rated capacity; SOC (State of Charge) is the battery's state of charge, defined as the percentage of the current remaining charge to the total capacity.

[0051] Step 3.3.2: During each forward propagation in model training, substitute the historical average current, average temperature, and average state of charge (SOC) corresponding to the current input batch of data into the coupled aging equation established in Step 3.3.1. Through numerical calculation, obtain the gradient direction of the health state change indicated by this equation under the current input conditions. .

[0052] Step 3.3.3: Calculate the consistency between the model's predicted trend and the physical trend. For the same batch of input data, calculate the gradient of the future health status change trend predicted by the model. Subsequently, the physical consistency loss term is calculated. It is defined as the model prediction gradient. Gradient indicated by physical equations The cosine distance between the two: ; The smaller the cosine distance, the more consistent the trend of health status changes predicted by the model is with the trend indicated by fundamental physical laws. This is achieved by optimizing the total loss function. To minimize This allows prior physical knowledge to be embedded into the learning process of neural networks in the form of soft constraints.

[0053] Step 4: Personalized online model updates and remaining useful life prediction based on transfer learning The purpose of this step is to address the adaptation problem of prediction models from generalized to personalized approaches, and to achieve dynamic, online prediction and updating of remaining useful life.

[0054] Step 4.1: Cloud-based general model pre-training. On a cloud server, collect complete lifecycle data of power batteries for various vehicle types, covering different models, regional climates, and user driving habits. Using this massive dataset, following the complete process from Steps 1 to 3, train a general physical information-enhanced deep residual spatiotemporal network prediction model. This general model learns the common aging patterns and laws of power batteries under various complex real-world application environments.

[0055] Step 4.2: Lightweight fine-tuning of the on-board edge model. When it is necessary to predict the power battery of a specific target vehicle, personalized transfer learning is performed.

[0056] Step 4.2.1: Load the physically enhanced deep residual spatiotemporal network prediction model trained in Step 4.1 from the cloud server to the on-board computing unit of the target vehicle or the associated edge computing server.

[0057] Step 4.2.2: Fix the parameters of the feature encoder in this general model, which consists of a one-dimensional convolutional neural network layer and a bidirectional long short-term memory network layer. The feature extraction capabilities learned by these layers are considered to be general knowledge.

[0058] Step 4.2.3: Extract the complete historical operation data of the target vehicle for the most recent 3 months, and perform data processing and feature extraction according to Step 1 and Step 2 to obtain the unique multi-timescale health feature factor sequence of the vehicle.

[0059] Step 4.2.4: Using three months of feature data from the target vehicle as a new training set, supervised training is performed only on the parameters of the fully connected regression layer at the end of the loaded model, with a small learning rate, for a few rounds. This process allows the model to quickly adapt to the individual aging characteristics and usage patterns of the target battery while retaining general knowledge.

[0060] Step 4.3: Recursive prediction of remaining useful life.

[0061] Step 4.3.1: Construct the daily feature vectors of the target vehicle for the most recent 30 consecutive days into the input feature tensor for the current moment, and input it into the prediction model after the personalized fine-tuning in Step 4.2.

[0062] Step 4.3.2: The prediction model operates in an autoregressive manner. First, based on the characteristics of the past 30 days, the health status value for day 31 is predicted. Then, the predicted health status value for day 31 is used as a feedback feature, combined with the actual features, to predict the health status value for day 32. This process is repeated recursively to predict the health status values ​​for days 1 through K, forming a complete future health status prediction sequence. K is the preset maximum number of prediction days, which can be set according to actual management needs, for example, 730 days.

[0063] Step 4.3.3: In the predicted health status sequence above, find the first point where the value is lower than the preset failure threshold. This failure threshold is set according to the power battery retirement standard, usually 70% or 80% of the rated capacity. The number of future predicted days corresponding to this health status value is the remaining service life of the target power battery calculated by the model.

[0064] Step 4.4: Rolling online updates and corrections of the prediction model. To achieve dynamic optimization of the prediction results, the system is set with a fixed model update cycle, such as one month.

[0065] Step 4.4.1: Whenever the update cycle is reached, the system automatically collects the new operating data of the target vehicle within that cycle.

[0066] Step 4.4.2: Perform steps 1 and 2 on the new data to extract new features and merge them with historical feature data to form an expanded dataset.

[0067] Step 4.4.3: Based on the expanded dataset, trigger the lightweight fine-tuning process described in step 4.2.4 again to incrementally update the regression layer parameters of the model.

[0068] Step 4.4.4: Using the model with updated parameters, re-execute the recursive prediction process of remaining useful life in step 4.3 to obtain an updated and more accurate prediction of remaining useful life, thereby realizing online dynamic correction of the prediction results.

[0069] The key parameters and thresholds involved in the method of this invention, including but not limited to sampling frequency, time window length, various temperature and state of charge thresholds, smoothing factor, loss function weight coefficient, training learning rate, historical observation days N, fine-tuning data months M, prediction days K, and update cycle, are determined according to a unified principle: First, initial empirical values ​​with physical or statistical significance are given based on electrochemical principles, vehicle engineering experience, or large-scale historical data statistics; then, on an independent performance verification dataset that is not involved in the final model training, with minimizing prediction error as the core optimization objective, automated parameter tuning algorithms such as grid search or Bayesian optimization are used for fine-tuning and determination; finally, the determined parameter combination and the trained prediction model together constitute a complete solution for deployment and application.

[0070] Example: Prediction of Remaining Battery Life for Urban Electric Taxi Fleets This embodiment uses a fleet of pure electric taxis operating in a city as an application scenario. All vehicles in the fleet are of the same model and are equipped with ternary lithium-ion power battery packs with a rated capacity of 220Ah. The battery management system has complete voltage, current, and temperature monitoring capabilities, as well as 4G remote transmission functionality. The operating company needs to accurately predict the remaining battery life of each vehicle in order to optimize battery maintenance plans, arrange tiered utilization, and control operating costs.

[0071] 1. Synchronous acquisition and preprocessing of multi-source heterogeneous data Step 1.1: Real-time Data Synchronization Acquisition. During vehicle operation, the onboard battery management system synchronously acquires the total voltage and current of the battery pack, as well as the surface temperature of individual cells at eight key monitoring points within the battery pack, at a fixed frequency of 1Hz. Simultaneously, through the vehicle controller area network bus, it acquires vehicle speed, motor power demand, accelerator pedal opening, brake pedal switching signals, and latitude and longitude coordinates from the onboard GPS module at the same 1Hz frequency. Each acquired data point is timestamped in milliseconds based on Coordinated Universal Time (UTC).

[0072] Step 1.2: Environmental Data Matching. Data is uploaded to the cloud data center in real time via the vehicle's 4G communication module. After receiving the data, the cloud server calls the commercial meteorological data interface based on the timestamp and latitude / longitude information in the data record to query and match the historical ambient temperature data of the vehicle's location at that moment. For example, a data record with a timestamp of "2023-05-15 14:30:00.500" and latitude / longitude of (116.4°, 39.9°) will match the ambient temperature recorded at that latitude / longitude grid point at "2023-05-15 14:30:00", assuming it is 25.3°C.

[0073] Step 1.3: Data Cleaning and Preprocessing. In the cloud, all synchronously acquired time-series data streams are cleaned. First, outlier detection is performed: A sliding window is formed by taking 30 sampling points before and after the current data point (a total of 61 points). The mean and standard deviation of the current data within the window are calculated. If the current value at the current point exceeds the range of ±3 times the standard deviation of the mean, it is determined to be an outlier caused by transient sensor interference and is removed. Other physical quantities (such as voltage and temperature) are processed according to the same rule.

[0074] For short-term data loss caused by network packet loss or outlier removal, if the missing duration is less than 10 seconds (far less than a typical urban driving cycle), linear interpolation of the preceding and following valid data will be used to fill the gap.

[0075] Finally, normalization is performed. For each data field (such as voltage, current, temperature, speed, etc.), the entire fleet's historical database is traversed to find the global maximum and minimum values ​​for that field. Then, the minimum-maximum normalization formula is used to transform each data value to the [0, 1] interval: ; Where X represents the original data. and These are the global minimum and maximum values ​​of the field, respectively. This is the normalized value.

[0076] 2. Dynamic extraction and construction of health characteristic factors at multiple time scales Step 2.1: Second-level transient feature extraction. The preprocessed continuous data stream is divided into 60-second non-overlapping fixed windows. Within each 60-second window, the following features are calculated: Current fluctuation: The standard deviation of the current sequence within the calculation window.

[0077] Average load intensity: Calculate the average absolute current value within the window and divide it by the battery rated capacity (220Ah) to convert it into the average charge-discharge rate (C-rate).

[0078] Temperature non-uniformity: The difference between the maximum and minimum temperatures at 8 monitoring points within the calculation window.

[0079] Internal resistance variation trend: Within each window, identify current pulse segments that are naturally generated by driver operation and last longer than 5 seconds. Using the voltage and current data before and after these pulse segments, employ the recursive least squares method to identify the ohmic internal resistance parameters in the first-order RC equivalent circuit model online. Calculate the current window's... With the previous window The relative rate of change.

[0080] At the end of each day, the mean, 95th percentile, and other statistical values ​​of the above features calculated from all windows (e.g., 1440 windows) are used as the representative values ​​of the transient features at the second level for that day.

[0081] Step 2.2: Journey-level operating condition feature extraction. Based on the vehicle's "ACC ON" (power on) and "ACC OFF" (power off) signals, continuous data is divided into independent journeys. For each journey, the following features are extracted: Total distance traveled (unit: km).

[0082] Average speed during the journey (unit: km / h).

[0083] Number of rapid acceleration / deceleration events: During the trip, an event in which the vehicle's acceleration is greater than 2.5 m / s² for one consecutive second is considered a rapid acceleration; an event in which the deceleration is less than -2.5 m / s² for one consecutive second is considered a rapid deceleration.

[0084] Highest temperature during the trip: Take the highest value of all temperature monitoring points during the trip.

[0085] Approximate depth of discharge: The discharged capacity (unit: Ah) is obtained by integrating the current during the stroke discharge phase, and the square root of this capacity is calculated. , as an approximate index of cumulative stress related to energy throughput.

[0086] At the end of each day, if there are multiple trips, the average value of the above characteristics across all trips is calculated as the trip-level operating condition characteristic for that day.

[0087] Step 2.3: Constructing Calendar Aging Cumulative Features. This is crucial for quantifying non-runtime aging.

[0088] High-temperature static cumulative time: Every day, the system counts the cumulative number of hours when the average temperature of the battery pack (average of 8 monitoring points) is higher than 30°C (this threshold is set based on the recognized temperature point where the high-temperature side reaction of ternary lithium batteries is significantly accelerated) and the vehicle is in a static state with the engine off (vehicle speed is 0 and motor power is 0).

[0089] Accumulated duration of high state of charge (SOC): Daily, the system estimates the battery's SOC online using an ampere-hour integral method combined with the open-circuit voltage method during rest periods. It also counts the cumulative number of hours when the SOC value is above 80% (this threshold is set based on the common understanding that high SOC accelerates cathode material degradation and electrolyte decomposition).

[0090] To smooth daily fluctuations and reflect cumulative effects, the two daily series above were processed using an exponentially weighted moving average (EWMA). The formula for calculating EWMA is: ; in, It is the original observation value on day t (duration of high temperature settling or duration of high SOC maintenance). It is the EWMA value on day t. It's the EWMA value from the previous day. It is a smoothing factor. In this embodiment, based on the slow aging process of the battery calendar (large time constant), it is used to smooth the process. Set to a small value of 0.05. This means that the current cumulative amount is highly dependent on the historical cumulative amount, resulting in a smooth sequence that reflects the long-term trend. The calendar aging cumulative features are used as the final input model.

[0091] 3. Physical Information Augmentation Deep Residual Spatiotemporal Network Prediction Model Step 3.1: Constructing the input feature tensor. The model input is a three-dimensional tensor with dimensions [batch size, time step, feature dimension]. Where: The time step, or the number of historical days N, is set to 30 days in this embodiment. This means that the model will examine the battery's behavior patterns over the past 30 days each time it makes a prediction.

[0092] The feature dimension is the set of all features for each day. The daily transient feature statistics (e.g., 4 items), trip-level work condition features (e.g., 5 items), and calendar aging cumulative features (2 items) extracted in step 2 are concatenated to form a unified daily feature vector. Assuming there are 11 features per day, the feature dimension is 11.

[0093] Therefore, for a car, its input tensor shape is [1, 30, 11], which means "1 sample, 11 features per day for the past 30 days".

[0094] Step 3.2: Network Structure Design. The model encoder adopts the following structure: 1. One-dimensional convolutional layer: Two one-dimensional convolutional layers with a kernel size of 3 are used to perform convolution operations on 11 features of each day. The purpose is to automatically learn and fuse the local correlation patterns between different features of each day (e.g., the intrinsic correlation that high current, high temperature and high SOC may occur at the same time).

[0095] 2. Bidirectional Long Short-Term Memory (LSTM) network layer: The convolutional feature sequence is input into a bidirectional LSTM layer containing 32 hidden units. This layer scans forward and backward along a 30-day time axis to capture the temporal dependencies of battery states (e.g., high-intensity operation over several consecutive days may lead to performance degradation due to lag).

[0096] 3. Residual Connections and Fully Connected Regression Layers: In deeper parts of the network, skip connections are introduced, directly adding the input of a certain layer to its output to form residual blocks, thus alleviating the vanishing gradient problem in deep network training. Finally, the encoded information from the LSTM output is mapped to a scalar output through two fully connected layers: a predicted health status at a specific future time. SOH is defined as the percentage of the current actual maximum available capacity to the rated capacity (220 Ah).

[0097] Step 3.3: Introduction of the Physical Consistency Loss Function. This is the key innovation of this invention. Total Loss Function Defined as:

[0098] This is the data loss term, which uses mean squared error loss to calculate the sum of squared differences between the predicted SOH sequence and the actual SOH sequence.

[0099] It is a physical consistency loss term. It is its weighting coefficient, which is set to 0.1 in this embodiment after optimization with the validation set.

[0100] Physical consistency loss term Construction and computation: 1. Establishing a simplified coupled aging equation: This invention constructs an empirical equation that integrates cyclic aging and calendar aging as a physical prior: ; The first term on the right-hand side of the equation approximately describes cyclic aging (and effective current). The first term is proportional to the nth power of the first term and is affected by the Arrhenius temperature T. The second term approximately describes calendar aging (related to SOC offset and temperature). ,n, , , These are empirical coefficients obtained by fitting a small amount of laboratory data. is the Boltzmann constant.

[0101] 2. Calculate the physically guided gradient direction: In each forward propagation of the model training, substitute the average current, average temperature, and average SOC corresponding to the current batch of input data (past 30 days) into the above equation, and numerically calculate the gradient direction of SOH change indicated by the equation under the current conditions. .

[0102] 3. Calculate consistency loss: Simultaneously, calculate the gradient of the future SOH change trend predicted by the model based on the same batch of input data. Physical consistency loss Defined as the cosine distance between these two gradient directions: ; The smaller the cosine distance, the more consistent the trend predicted by the model is with the trend indicated by the physical equations. This is achieved by minimizing... As the model learns from the data, it is "guided" in a direction that conforms to basic physical laws, which greatly enhances its rationality and generalization ability in sparse data regions or when extrapolating predictions.

[0103] 4. Personalized online prediction and rolling updates based on transfer learning Step 4.1: Cloud-based general model pre-training. On a cloud server, collect complete historical data (including the "failure point" where the final capacity decays to 70%) from the past three years for 500 retired or end-of-life vehicles in the taxi fleet. Using the data from these 500 vehicles, train a general physical information-enhanced deep residual spatiotemporal network prediction model, called the "general base model," following the procedures described in Steps 1-3. This model learns the common aging patterns of the vehicle's batteries under various urban operating conditions and different driver habits.

[0104] Step 4.2: Personalized Fine-tuning of the Target Vehicle Model. When it is necessary to make predictions for a newly put into operation vehicle (the target vehicle): 1. Load the "general base model" from the cloud to the vehicle's onboard edge computing unit or an edge server near the vehicle.

[0105] 2. Freeze (fix parameters) the weights of all one-dimensional convolutional layers and bidirectional LSTM layers in the model. These layers are considered "feature encoders" that extract common spatiotemporal features from the data, and the knowledge they learn is universal.

[0106] 3. Extract the complete operating data of the target vehicle for the most recent 3 months (approximately 90 days), and after processing in steps 1 and 2, obtain its unique feature sequence.

[0107] 4. Using only the data from these 3 months, and with a very small learning rate (e.g., 0.0001), retrain (fine-tune) the last 1-2 fully connected layers (which can be considered "regressors") of the model in a small number of rounds (e.g., 50 rounds). This process allows the model to quickly adapt to the aging rate and usage patterns specific to the target vehicle without having to train from scratch, achieving efficient "cloud-edge" collaboration and knowledge transfer.

[0108] Step 4.3: Recursive prediction of remaining useful life. Prediction is performed using a fine-tuned personalized model: 1. Input the feature vectors of the target vehicle for the most recent 30 consecutive days into the model.

[0109] 2. The model works in an autoregressive manner: it first predicts the SOH on day 31; then it combines the predicted SOH on day 31 (as a dummy feature) with the real feature to predict the SOH on day 32; this process is repeated until the prediction reaches day 730 (two years later).

[0110] 3. In the predicted SOH sequence for the next 730 days, find the first point where the value is below 70% (a commonly used industry retirement threshold). The number of days in the future corresponding to this point is the model-predicted Remaining Useful Life (RUL).

[0111] Step 4.4: Rolling Online Model Updates. To achieve dynamic correction of predictions, the system is set to a monthly model update cycle. At the end of each month, the system automatically adds the new data for the target vehicle from that month to the training set, triggering the lightweight fine-tuning process in Step 4.2 again (still only updating the fully connected layers). Then, the prediction in Step 4.3 is re-executed using the updated model. In this way, the predicted remaining lifespan will be dynamically and continuously updated as the vehicle is actually used, becoming increasingly accurate.

[0112] 5. Effect Verification and Analysis To verify the effectiveness of the method in this embodiment, 10 retired vehicles that were not used in the training of the general model were randomly selected from the fleet as the test set. The method of this invention (denoted as PIRST-Net) is compared with the following two methods: Comparative Example 1 (Traditional Data-Driven Approach): Uses the same input features, but uses a standard Long Short-Term Memory (LSTM) network for prediction, with no physical loss term.

[0113] Comparative Example 2 (same structure without physical constraints): Using the exact same deep residual spatiotemporal network structure as PIRST-Net, but only using the data loss term during training (i.e., ), denoted as ST-Net.

[0114] Using the actual remaining lifespan of the battery when its state of equilibrium (SOH) first falls below 70% as a benchmark, the absolute errors of predictions made by different methods at multiple early time points throughout the battery's lifespan are compared. The results are shown in Table 1. Table 1 Test vehicle ID Real RUL (Heaven) Predicting the initial SOH (%) Traditional LSTM method predicts RUL (days) / error (days) ST-Net method predicts RUL (days) / error (days) The PIRST-Net method of this invention predicts RUL (days) / error (days). V001 550 95 480 / -70 510 / -40 535 / -15 V002 620 97 520 / -100 580 / -40 605 / -15 V003 480 93 420 / -60 440 / -40 465 / -15 ... ... ... ... ... ... V010 590 96 500 / -90 545 / -45 575 / -15 Mean Absolute Error (MAE) - - 78.5 days 41.3 days 15.2 days Effect Analysis: 1. Significantly improved prediction accuracy: The mean absolute error of the proposed method (PIRST-Net) is 15.2 days, far lower than the 78.5 days of the traditional LSTM method and the 41.3 days of ST-Net without physical constraints. This demonstrates the effectiveness of combining multi-timescale feature engineering with physical information augmentation.

[0115] 2. High stability in early predictions: In the early prediction stage where the SOH is above 90%, the prediction trajectories of traditional LSTM and ST-Net fluctuate significantly, while PIRST-Net's prediction curve is smoother and converges to the vicinity of the true decay trend earlier. This is due to the physical consistency loss. The guidance enables the model to output reasonable predictions that conform to physical common sense even when faced with weak aging signals and relatively prominent noise in early data, thus avoiding absurd prediction results.

[0116] 3. Balancing Personalization and Generalization Capabilities: Through a transfer learning framework of "cloud pre-training - edge fine-tuning," the general model learns commonalities from big data, while the fine-tuning process quickly captures individual characteristics. This enables the method of this invention to provide accurate personalized predictions when faced with vehicles in the same fleet with different driving styles, ranging from aggressive to mild, demonstrating excellent engineering applicability.

[0117] In summary, the method described in detail in this embodiment effectively solves the problems of insufficient utilization of big data, weak model generalization ability, and insufficient modeling of complex coupled aging in the prediction of the remaining service life of power batteries through a set of closely linked technical solutions. It achieves high-precision, high-reliability, and online-updable personalized life prediction, meeting the needs of actual fleet management.

[0118] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0119] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for predicting the remaining service life of a power battery based on big data analysis, characterized in that, The method includes the following steps: Step 1: Synchronously collect the running time series data of the power battery, the vehicle operating condition time series data, and the historical environmental data, and perform cleaning, missing value imputation, and normalization preprocessing on the multi-source time series data; Step 2: Based on the preprocessed data, dynamically extract transient features at the second or minute level and travel-level operating condition features, and construct calendar aging cumulative features to form multi-timescale health feature factors; Step 3: Construct a deep residual spatiotemporal network prediction model with physical information enhancement; the physical information enhancement specifically involves introducing a physical consistency loss term into the total loss function of the model training. The physical consistency loss term is calculated based on the empirical degradation model and is used to constrain the health state change trend predicted by the model to conform to physical laws. Step 4: Based on transfer learning, the pre-trained physical information-enhanced deep residual spatiotemporal network prediction model is individually fine-tuned using recent historical data of the target vehicle. The fine-tuned model is then used to recursively predict the health state sequence of the target power battery until the health state is below the failure threshold, thus obtaining the remaining service life.

2. The method for predicting the remaining service life of a power battery based on big data analysis according to claim 1, characterized in that, In step 1, the specific process of synchronous acquisition includes: Step 1.1: The on-board battery management system collects the operating sequence data of the power battery cells in real time at a fixed sampling frequency of 1 Hz; the operating sequence data includes terminal voltage, total current, and temperature data of each battery cell surface measured by a temperature sensor. Step 1.2: Synchronously collect vehicle operating condition timing data via the vehicle controller local area network bus at the same sampling frequency as in Step 1.1; the vehicle operating condition timing data includes vehicle speed, motor power demand, accelerator pedal opening, brake pedal status, and vehicle geographical location information; Step 1.3: Timestamp each data record collected in Steps 1.1 and 1.2 to the millisecond level; upload the timestamped data to the cloud server via the network communication module; Step 1.4: On the cloud server, based on the timestamp and vehicle geographic location information, retrieve and match historical environmental data with the same timestamp from the environmental data service platform; the historical environmental data includes at least ambient temperature data; The specific processes of cleaning, missing value imputation, and normalization preprocessing include: Step 1.5: Using an outlier detection algorithm based on sliding window and 3 sigma criterion, outlier detection is performed on all time series data obtained in Steps 1.1, 1.2 and 1.4, and data identified as outliers are removed. Step 1.6: For missing data caused by data removal or transmission loss, if the missing duration is much shorter than the main data change cycle, linear interpolation is used to fill in the missing data; the main data change cycle is determined based on the typical charging and discharging cycle duration of the vehicle. Step 1.7: Perform min-max normalization on all the filled numeric field data to linearly transform all the data to the interval [0, 1].

3. The method for predicting the remaining service life of a power battery based on big data analysis according to claim 2, characterized in that, In step 2, the extraction process of the second-level or minute-level transient features includes: Step 2.1.1: Divide the preprocessed time series data into fixed-duration windows of 60 seconds to obtain a continuous time window sequence; Step 2.1.2: Within each time window, calculate a set of transient response characteristics; the transient response characteristics include: the standard deviation of the current within the time window, the average charge / discharge rate within the time window, and the difference between the highest and lowest temperatures within the time window; Step 2.1.3: Within each time window, based on the first-order RC equivalent circuit model, the voltage and current data corresponding to the pulse charge and discharge segments within the window are used to identify the ohmic internal resistance value of the power battery online through the recursive least squares method, and the rate of change of the ohmic internal resistance value relative to the ohmic internal resistance value in the previous time window is calculated. The extraction process of the travel-level operating condition features includes: Step 2.2.1: Divide the continuous time-series data into multiple independent travel segments based on the vehicle power-on and power-off signals; Step 2.2.2: Within each trip segment, calculate a set of working condition aggregate features; the working condition aggregate features include: total trip mileage, average trip speed, total number of rapid acceleration events and rapid deceleration events during the trip, the highest temperature reached by the power battery during the trip, and the square root of the cumulative discharge capacity during the trip; wherein, a rapid acceleration event is defined as a continuous event with an acceleration greater than 2.5 meters per second, and a rapid deceleration event is defined as a continuous event with a deceleration less than -2.5 meters per second.

4. The method for predicting the remaining service life of a power battery based on big data analysis according to claim 3, characterized in that, In step 2, the specific process of constructing the calendar aging cumulative features includes: Step 2.3.1: Calculate the daily cumulative high-temperature static time using a 24-hour statistical period; the calculation method for the cumulative high-temperature static time is: count the cumulative number of hours within a day when the average temperature of the power battery is higher than 30 degrees Celsius and the vehicle is in a stationary state with the engine off. Step 2.3.2: Calculate the cumulative duration of high state of charge maintenance per day using a 24-hour statistical period; the method for calculating the cumulative duration of high state of charge maintenance is: count the cumulative number of hours within a day when the state of charge value of the power battery is higher than 80%, estimated by combining the ampere-hour integration method with the open-circuit voltage method; Step 2.3.3: Perform exponentially weighted moving average processing on the daily series of high temperature static accumulation time obtained in Step 2.3.1 and the daily series of high charge state maintenance accumulation time obtained in Step 2.3.2 to obtain two smoothed long-term cumulative series; the decay factor of the exponentially weighted moving average is set according to the time constant of the power battery calendar aging process.

5. The method for predicting the remaining service life of a power battery based on big data analysis according to claim 4, characterized in that, In step 3, the specific process of constructing the physically enhanced deep residual spatiotemporal network prediction model includes: Step 3.1: Construct the model input feature tensor; the input feature tensor is composed of daily feature vectors of N consecutive days stacked in chronological order; the daily feature vector is composed of transient feature statistics, travel-level operating condition features and calendar aging cumulative features extracted in Step 2; the value of N is determined according to the historical observation period required for prediction, and is an integer not less than 30. Step 3.2: Construct a deep residual spatiotemporal network encoder; the encoder is composed of a one-dimensional convolutional neural network layer and a bidirectional long short-term memory network layer connected sequentially; the one-dimensional convolutional neural network layer is used to extract the local correlation within the daily feature vector; the bidirectional long short-term memory network layer is used to extract the sequential dependency relationship of features in the time dimension; in the deep layers of the network, skip connections are set to form a residual structure; Step 3.3: Define the total loss function for physical information enhancement; the total loss function is obtained by weighted summation of the data loss term and the physical consistency loss term; the data loss term is the mean square error between the health state value predicted by the model and the actual health state value; the physical consistency loss term is calculated based on the coupled aging equation and is used to penalize the deviation between the health state change trend predicted by the model and the physical trend described by the coupled aging equation; the coupled aging equation is an empirical model that integrates the power-law cyclic aging model and the Arrhenius temperature dependence relationship.

6. The method for predicting the remaining service life of a power battery based on big data analysis according to claim 5, characterized in that, The specific calculation and introduction process of the physical consistency loss term in step 3.3 includes: Step 3.3.1: Establish the coupled aging equation; the coupled aging equation represents the healthy state as a function of cyclic cumulative stress and temperature stress, and its mathematical form includes power-law terms and exponential terms; Step 3.3.2: During each forward propagation of the model training, input the historical loop count and temperature history sequence corresponding to the current batch input feature tensor into the coupled aging equation, and calculate the gradient direction of the equation with respect to the health state under the current conditions; Step 3.3.3: Calculate the cosine or Euclidean distance between the model's prediction of future health status changes based on the same batch of data and the physical gradient direction calculated in Step 3.3.2; Step 3.3.4: The distance value calculated in step 3.3.3 is used as the physical consistency loss term, and the data loss term is weighted and summed according to the preset weight coefficients to form the total loss function used for backpropagation optimization; the weight coefficients are selected in the range of 0.01 to 0.5 through cross-validation.

7. The method for predicting the remaining service life of a power battery based on big data analysis according to claim 5, characterized in that, In step 4, the specific process of personalized fine-tuning of the model based on transfer learning includes: Step 4.1: Cloud-based general model pre-training; On a cloud server, collect complete life cycle data of power batteries for various types of vehicles covering different models, regions, and driving habits. Following the methods described in Steps 1 to 3, train a general physical information-enhanced deep residual spatiotemporal network prediction model as the base model. Step 4.2: Lightweight fine-tuning of the vehicle edge model; For a specific target vehicle, load the general model from the cloud to the vehicle computing unit or edge server; Fix the feature encoder parameters in the general model, which consists of a one-dimensional convolutional neural network layer and a bidirectional long short-term memory network layer; Step 4.3: Extract the complete historical data of the target vehicle for the most recent M months, process it according to Step 1 and Step 2, and obtain the multi-timescale health feature factor sequence corresponding to the vehicle; where M is an integer not less than 3; Step 4.4: Using the target vehicle feature data obtained in Step 4.3 as the new training dataset, supervised training with a small learning rate and few iterations is performed only on the fully connected regression layer parameters at the end of the general model that were loaded in Step 4.2, to complete the transfer and adaptation of the model from the general domain to the target vehicle individual domain.

8. The method for predicting the remaining service life of a power battery based on big data analysis according to claim 7, characterized in that, In step 4, the specific process of recursively predicting the target power battery health state sequence to obtain the remaining service life includes: Step 4.5: Construct the daily feature vectors of the target vehicle for the most recent N consecutive days into the input feature tensor for the current moment, and input it into the prediction model after the personalized fine-tuning in Step 4.4; Step 4.6: The model uses the health status value predicted in the previous step as part of the input for the next step, and recursively predicts the health status value from day 1 to day K in an autoregressive manner, forming a health status prediction sequence; K is the preset maximum prediction horizon, which is set according to the vehicle maintenance cycle. Step 4.7: In the health status prediction sequence, find the first health status value that is lower than the preset failure threshold; the failure threshold is set to 70% or 80% of the rated capacity according to the power battery retirement standard. Step 4.8: The number of predicted days corresponding to the health status value below the failure threshold is the remaining service life predicted by the model; the remaining service life result is output to the vehicle display terminal or remote monitoring platform in real time.

9. The method for predicting the remaining service life of a power battery based on big data analysis according to claim 8, characterized in that, The method further includes step 5: rolling updates of the prediction model and online correction of the remaining useful life; Step 5.1: Set a fixed model update cycle; the update cycle is synchronized with the vehicle's regular maintenance cycle or data accumulation cycle, which is 1 month or 3 months; Step 5.2: Whenever an update cycle is reached, automatically collect the new operating data of the target vehicle within that cycle, and execute Step 1 and Step 2 to extract the multi-timescale health feature factors corresponding to the new data in that cycle; Step 5.3: Merge the newly added feature factors with the historical feature factors to form an expanded training dataset, and trigger the lightweight fine-tuning process described in step 4.4 again to incrementally update the parameters of the prediction model. Step 5.4: Using the model with updated parameters, re-execute steps 4.5 to 4.8 to generate updated remaining useful life predictions, thereby achieving dynamic online correction of the useful life prediction.

Images