Adaptive multi-working-condition battery SOC-SOH state joint estimation system and adaptive multi-working-condition battery SOC-SOH state joint estimation method
By combining an integrated battery aging estimation module with a deep neural network, the problem of joint estimation of battery SOC and SOH in electric vehicles is solved, achieving high accuracy and interpretability under small sample data, and improving the accuracy and reliability of driving range prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-04
- Publication Date
- 2026-04-03
AI Technical Summary
Existing technologies struggle to achieve high-precision, interpretable joint estimation of battery SOC and SOH in electric vehicles with limited onboard data, especially under complex dynamic conditions where estimation accuracy and model generalization ability are insufficient.
An integrated battery aging estimation module is adopted, which combines a battery capacity degradation model and a LightGBM residual correction model. Through multi-dimensional feature extraction and sliding window analysis technology, driving conditions are identified in real time, and deep neural networks are used for SOC estimation to construct an adaptive multi-condition battery SOC-SOH joint estimation system.
It achieves high-precision estimation of battery state of health (SOH) under small sample data conditions, significantly improving the accuracy and reliability of driving range prediction, and solving the accuracy and adaptability problems of traditional methods under complex operating conditions.
Smart Images
Figure CN121784596A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of electric vehicle battery management technology, specifically relating to an adaptive multi-condition battery SOC-SOH joint estimation system and method. Background Technology
[0002] Batteries are the energy source for electric vehicles, and accurate estimation of their internal states (such as State of Charge (SOC) and State of Health (SOH)) is a core function of battery management systems, directly impacting the accuracy of range prediction, vehicle energy management strategies, and battery safety and lifespan. However, the complex electrochemical reactions within batteries prevent direct measurement of their state parameters using sensors; estimation must be indirect, relying on external observations such as voltage, current, and temperature. Furthermore, various battery state parameters are interconnected and influence each other. Among these, the degradation of State of Health (SOH) significantly affects the accuracy of SOC estimation, representing a key challenge for practical battery management systems. Increased battery usage time and charge / discharge cycles lead to capacity decay and increased internal resistance, directly reducing usable energy and charging / discharging efficiency. For example, a 10% capacity decay in a vehicle's battery could theoretically reduce the driving range by 15%-20%, and aging can also cause decreased battery pack consistency, further exacerbating range fluctuations. In practical automotive applications, battery state estimation becomes even more complex, primarily manifesting in the following three aspects.
[0003] (1) There is a deep coupling estimation problem between SOC and SOH. SOH characterizes the performance degradation of the battery due to aging, and its core indicator - the current maximum usable capacity Q of the battery - is the basis for the accurate calculation of SOC by the ampere-hour integration method. If the SOH is not estimated accurately, it will directly lead to irreversible cumulative error in SOC, causing the range display to be "falsely high" or suddenly "jump", which will seriously aggravate the user's range anxiety.
[0004] (2) The vehicle environment faces a serious bottleneck of "small sample data". Battery capacity is a slowly varying parameter, and it is difficult to obtain a large amount of complete charge and discharge cycle data for data-driven model learning in actual vehicle operation. This "small sample" dilemma makes it difficult for many data-driven methods that rely on big data training to generalize, while traditional physical models are difficult to accurately describe the complex and ever-changing actual aging trajectory.
[0005] (3) Dynamic driving conditions and driving style have a profound impact on battery status. Different road conditions (such as congestion and highway) and the driver's operating behavior (such as aggressive and mild) jointly determine the instantaneous power demand and energy consumption mode of the vehicle, which is directly reflected in the charging and discharging current, voltage change rate and temperature rise of the battery, and has a significant impact on the instantaneous value of SOC and the long-term decay of SOH.
[0006] Existing technologies typically take the software algorithm module in the vehicle's battery management system (BMS) as their core, while their hardware foundation is the electronic control unit (ECU), which includes a microprocessor (MCU), current / voltage / temperature sampling circuits, and storage units. At the methodological level, the mainstream technical approaches are mainly model-driven and data-driven methods.
[0007] The core structure of the model-driven approach is to construct an equivalent circuit model of the battery (such as a first-order or second-order RC model). This model consists of a voltage source, an ohmic internal resistance, and several RC parallel networks connected in series to simulate the battery's dynamic characteristics such as open-circuit voltage and polarization effects. Its main workflow is as follows: First, the initial values of the model parameters are identified through offline experiments; during online operation, the system acquires the battery's terminal voltage and load current at a fixed sampling period (e.g., 100ms); then, using algorithms such as Kalman filters (e.g., EKF), the model-predicted terminal voltage is compared with the actually measured terminal voltage. By minimizing the prediction error, the state variables in the model (usually including SOC and the voltage across the RC networks) are recursively updated and corrected, and the SOH is indirectly inferred (e.g., by tracking changes in capacity or internal resistance).
[0008] The data-driven approach relies on a complex machine learning model (such as LSTM or CNN) pre-trained on a cloud or high-performance computing platform, which is then deployed within the Battery Management System (BMS). The workflow is as follows: During the training phase, massive amounts of historical battery operating data (voltage, current, and temperature sequences) under various operating conditions and their corresponding true state labels are collected. Supervised learning is then used to establish a non-linear mapping from input data to the target state (SOC, SOH). During the deployment phase, the BMS inputs real-time vehicle operating data into this trained model, directly outputting state estimates.
[0009] However, the aforementioned methods often struggle to balance estimation accuracy, model generalization ability, and physical interpretability when addressing the combined challenges of small sample sizes and complex dynamic operating conditions in vehicles. Therefore, there is an urgent need in this field for an innovative method that can operate stably under small sample conditions, integrate driving condition information, and clearly characterize battery aging mechanisms, thereby achieving high-precision and high-reliability joint estimation of SOC and SOH. Summary of the Invention
[0010] To address the problems existing in the prior art, this invention provides an adaptive multi-condition battery SOC-SOH joint estimation system and method. Under the condition of small sample data in the vehicle, it achieves high-precision and interpretable online estimation of battery state of health (SOH), and performs real-time correction of battery state of charge (SOC) based on this, thereby significantly improving the accuracy and reliability of range prediction throughout the entire life cycle.
[0011] To achieve the above objectives, the present invention provides the following solution: An adaptive multi-condition battery SOC-SOH joint estimation system, the system comprising: an integrated battery aging estimation module, a driving information dynamic characterization module, and a battery SOC estimation module; The integrated battery aging estimation module is used to estimate the battery health status. The driving information dynamic representation module is used to identify and quantify the vehicle's driving conditions in real time through multi-dimensional feature extraction and sliding window analysis technology. The battery SOC estimation module is used to estimate the battery's state of charge based on the battery health status estimation results and the vehicle's driving conditions. It also uses the SOC estimation step to calculate the SOC change value, dynamically updates the charge change in combination with the battery's current maximum available capacity, and feeds it back to the integrated battery aging estimation module to update the battery aging status.
[0012] Preferably, the integrated battery aging estimation module includes: a battery capacity degradation model and a LightGBM residual correction model; Among them, the battery capacity degradation model: ; In the formula, for Physically predicted battery capacity at the given time; This refers to the battery's rated capacity at the time of manufacture. This is the time decay coefficient; This refers to the cumulative battery usage time. To account for the degradation coefficient due to charging and discharging habits; For each charge / discharge cycle of the battery; LightGBM residual correction model: ; In the formula, This represents the actual maximum usable capacity of the battery at present. For residuals; Final battery aging condition estimate: .
[0013] Preferably, the dynamic representation module for driving information includes: a data cleaning unit, a sliding window feature extraction unit, and a one-hot encoding output unit; The data cleaning unit is used to standardize and clean the collected driving behavior data. The sliding window feature extraction unit is used to dynamically analyze the cleaned continuous driving behavior data using sliding window technology, and capture the time-varying characteristics of driving conditions through statistical features within a local window. The one-hot encoding output unit is used to establish a hierarchical driving condition classification system based on the feature vector extracted by the sliding window, and to discretize continuous driving information into standardized driving condition categories.
[0014] Preferably, the process of dynamically analyzing continuous driving behavior data after cleaning using sliding window technology, and capturing the time-varying characteristics of driving conditions through statistical features within a local window, includes: ; ; ; In the formula, For average speed, To determine the activity level of the pedals, The average current is given, and WindowSize is the size of the sampling window. For the velocity information at the sampling time, This is the sampling time sequence within the window. and These are lists of drive / brake pedal states for the sampled time sequences. and These represent the total number of drive / brake pedal activation states in the sampled time sequence. This refers to the current information at the sampling time.
[0015] Preferably, the battery SOC estimation module adopts a deep neural network (DNN) model, which includes: an input feature preprocessing layer, a deep residual feature extraction layer, a feature compression and nonlinear mapping layer, and an SOC regression output layer. The input feature preprocessing layer is used to preprocess and fuse real-time sensor time-series features, battery health global features, and driving condition category features to obtain a fused feature vector. The deep residual feature extraction layer is used to extract the dynamic deep spatiotemporal features of the battery through residual connections based on the fused feature vector. The feature compression and nonlinear mapping layer is used to reduce the dimensionality of deep spatiotemporal features and complete the nonlinear mapping transition to the final regression target. The SOC regression output layer is used to map the compressed feature representation to the final SOC estimate.
[0016] This invention also provides an adaptive multi-condition battery SOC-SOH joint estimation method, which is implemented through the aforementioned system and includes: By using multi-dimensional feature extraction and sliding window analysis technology, the driving conditions of vehicles can be identified and quantified in real time. An integrated battery aging estimation module based on the battery capacity degradation model and the LightGBM residual correction model was established, using the current maximum available capacity of the battery as the standard for evaluating the battery aging state. Based on driving conditions, the current maximum usable battery capacity output in real time by the integrated battery aging estimation module is directly input into the battery SOC estimation module as input features. The dynamic capacity that changes over time provides a dynamic representation of the battery aging state for SOC estimation. The battery SOC estimation module outputs the initial and final SOC values and uses the estimated SOC difference to calculate the change in charge, which is then input into the integrated battery aging estimation module as input features to update the battery aging state.
[0017] Preferably, methods for real-time identification and quantification of vehicle driving conditions through multi-dimensional feature extraction and sliding window analysis techniques include: The collected driving behavior data is standardized and converted to a new format, and then cleaned. The sliding window technique is used to dynamically analyze continuous driving behavior data after cleaning, and the time-varying characteristics of driving conditions are captured by statistical features within a local window. Based on the feature vectors extracted by the sliding window, a hierarchical working condition classification system is established to discretize continuous driving information into standardized working condition categories.
[0018] Preferably, the method of dynamically analyzing continuous driving behavior data after cleaning using sliding window technology, and capturing the time-varying characteristics of driving conditions through statistical features within a local window, includes: ; ; ; In the formula, For average speed, To determine the activity level of the pedals, The average current is given, and WindowSize is the size of the sampling window. For the velocity information at the sampling time, This is the sampling time sequence within the window. and These are lists of drive / brake pedal states for the sampled time sequences. and These represent the total number of drive / brake pedal activation states in the sampled time sequence. This refers to the current information at the sampling time.
[0019] Preferred battery capacity degradation model: ; In the formula, for Physically predicted battery capacity at the given time; This refers to the battery's rated capacity at the time of manufacture. This is the time decay coefficient; This refers to the cumulative battery usage time. To account for the degradation coefficient due to charging and discharging habits; For each charge / discharge cycle of the battery; LightGBM residual correction model: ; In the formula, This represents the actual maximum usable capacity of the battery at present. For residuals; Final battery aging condition estimate: .
[0020] Preferably, the battery SOC estimation module adopts a deep neural network (DNN) model, which includes: an input feature preprocessing layer, a deep residual feature extraction layer, a feature compression and nonlinear mapping layer, and an SOC regression output layer. The input feature preprocessing layer is used to preprocess and fuse real-time sensor time-series features, battery health global features, and driving condition category features to obtain a fused feature vector. The deep residual feature extraction layer is used to extract the dynamic deep spatiotemporal features of the battery through residual connections based on the fused feature vector. The feature compression and nonlinear mapping layer is used to reduce the dimensionality of deep spatiotemporal features and complete the nonlinear mapping transition to the final regression target. The SOC regression output layer is used to map the compressed feature representation to the final SOC estimate.
[0021] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention fundamentally solves the problem of SOC prediction accuracy caused by the fixed battery capacity parameters and poor adaptability to operating conditions in traditional methods by introducing a coupled prediction technique that considers driving information and SOC-SOH. Through residual learning of a physical model and LightGBM, an integrated battery aging estimation model applicable to small sample data is constructed, balancing prediction accuracy and interpretability. The constructed DNN model for SOC prediction is coupled with the integrated battery aging estimation model to achieve closed-loop collaborative updating of SOC and SOH.
[0022] In real-world road condition verification, the implementation of this invention has yielded significant results in three aspects: improved prediction accuracy, enhanced long-term stability, and broadened engineering applicability. Attached Figure Description
[0023] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0024] Figure 1 This is a schematic diagram of an adaptive multi-condition battery SOC-SOH joint estimation system according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating the joint estimation of battery SOC-SOH state considering driving information in an embodiment of the present invention. Figure 3 This is a schematic diagram of the calibration results of the battery aging degradation model in an embodiment of the present invention; Figure 4 This is a schematic diagram of the integrated battery aging estimation results according to an embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the training and validation results of the SOC estimation model in an embodiment of the present invention. Detailed Implementation
[0025] 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 some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0026] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0027] Example 1 In actual operation of electric vehicles, accurate estimation of battery state of health (SOH) faces the challenge of scarce effective data, leading to significant lag and inaccuracy in updating the battery's maximum usable capacity. This problem further perpetuates the estimation process of battery state of charge (SOC), causing accumulated errors in the ampere-hour integration method, ultimately resulting in decreased and fluctuating accuracy in vehicle range prediction, exacerbating users' range anxiety. To address these issues, such as... Figure 1As shown, this invention provides an adaptive multi-condition battery SOC-SOH joint estimation system. The system is deployed in the vehicle's onboard BMS and receives data from the CAN bus in real time. It collects data during vehicle operation (which can include any segment of parking charging and normal driving), and sequentially performs data processing, condition identification, capacity prediction, and SOC joint estimation according to the workflow of this invention. Ultimately, the system can achieve accurate and continuous SOC estimation, providing drivers with a reliable reference for driving range. The system includes: an integrated battery aging estimation module, a dynamic driving information characterization module, and a battery SOC estimation module. An integrated battery aging estimation module is used to estimate the battery's health status. The driving information dynamic representation module is used to identify and quantify the vehicle's driving conditions in real time through multi-dimensional feature extraction and sliding window analysis technology. The battery SOC estimation module is used to estimate the battery's state of charge based on the battery health status estimation results and the vehicle's driving conditions. It also uses the SOC estimation step to calculate the SOC change value, dynamically updates the charge change in combination with the battery's current maximum available capacity, and feeds it back to the integrated battery aging estimation module to update the battery aging status.
[0028] Specifically: 1. Integrated battery aging estimation module This module integrates physical mechanisms and data-driven methods to construct a two-level estimation architecture, namely the battery capacity degradation model and the LightGBM residual correction model, to achieve high-precision and interpretable estimation of battery health status under small sample conditions.
[0029] 1.1 Battery Capacity Degradation Model Based on the battery aging mechanism, a battery capacity degradation model considering time degradation and charging / discharging habits is established: In formula (1), for Physically predicted battery capacity (Ah) at the current time. The rated battery capacity (Ah) at the time of manufacture. This is the time decay coefficient; This refers to the cumulative battery usage time (in hours). To account for the degradation coefficient due to charging and discharging habits; This represents the change in battery charge / discharge capacity (Ah) for each charge / discharge cycle.
[0030] 1.2 Battery Capacity Degradation Model Calibration By using sensors to collect real-time operating data of electric vehicles, and through the effective charge and discharge segments that meet specific conditions recorded by the onboard BMS system, the maximum usable battery capacity is indirectly calculated. The specific calculation formula is as follows: In formula (2) The change in electrical charge (Ah) is calculated using the ampere-hour integration method. The net change in SOC for this segment is (%). The SOC estimation method can be the conventional SOC-OCV estimation, but the specific estimation method will not be explained in detail here.
[0031] Data segment selection criteria: Prioritize complete charge-discharge cycles to avoid prolonged periods of inactivity or fragmented charge-discharge switching; ensure that the average charge-discharge rate of the selected data segments is within a reasonable range, avoiding drastic transient power processes such as rapid acceleration and deceleration to reduce errors caused by battery polarization effects; ensure that the average temperature of the battery pack is within its normal operating range, and use the SOC results after the car has been inactive for a long time as much as possible as the initial and final SOC values.
[0032] Each selected segment generates a capacity-time-power change data point. ),in This represents the moment in time when the segment occurs. Over time, the vehicle will accumulate a series of sparse but reliable data points of this kind, which constitute the small sample battery capacity dataset required by the integrated battery aging estimation module.
[0033] The battery capacity degradation model was personalized and calibrated using a small sample battery capacity dataset to obtain the unknown coefficients in formula (1). , A calibrated battery physical degradation model can make a preliminary prediction of the macroscopic degradation trend of battery capacity based on the battery's service life and the amount of charge and discharge per cycle.
[0034] 1.3 LightGBM Residual Correction Model Because battery capacity degradation models cannot accurately describe complex nonlinear aging factors, their estimates inevitably deviate from the true values. To overcome this limitation, this invention introduces a lightweight LightGBM machine learning model specifically designed to learn and correct systematic biases in the physical model's predictions. The specific correction mechanism and training process are as follows: The learning objective of the LightGBM model is to obtain the prediction residuals of the battery capacity degradation model, i.e., to construct a mapping function from a series of relevant features to the residual values. Defined as the true value of the battery's current maximum usable capacity at a specific time t. Compared with the physical model predictions The difference between them: Construct an input feature set containing the following types: maximum operating current, voltage extremes, absolute value of segment energy, battery capacity degradation model prediction, and time. Train a residual prediction model using the LightGBM algorithm, with the above multidimensional features as input and the residual correction value of the battery capacity degradation model prediction as output.
[0035] The final battery aging state estimate is determined by a combination of physical model baseline prediction and machine learning residual correction.
[0036] This module maintains the rationality of the physical model's mechanism while capturing complex nonlinear aging modes through a data-driven approach, achieving high-precision battery health state estimation under small sample conditions.
[0037] 2. Dynamic representation module of driving information This module utilizes multi-dimensional feature extraction and sliding window analysis techniques to identify and quantify vehicle driving conditions in real time, providing refined operating condition context information for battery state estimation. The module employs a three-level processing architecture: a data cleaning unit, a sliding window feature extraction unit, and a one-hot encoded output unit.
[0038] 2.1 Data Cleaning Unit The collected data sources should include the following: speed sequence (km / h), current sequence (A), and accelerator (drive) pedal state sequence. (Boolean value 0 or 1), brake pedal state sequence (Boolean value 0 or 1), timestamp sequence (Unix timestamp), sampling frequency (Hz).
[0039] In the data preprocessing stage, the first step is to standardize the format of the collected raw data. Since the raw data is usually stored as strings or in special encoded formats, it needs to be parsed into a numerical list structure to ensure consistency in subsequent processing. Specific conversion methods include string splitting, type conversion, and structure verification.
[0040] After the format conversion is completed, the data cleaning process is executed.
[0041] Outlier identification is performed on velocity and current sequences using the 3σ principle. The mean (μ) and standard deviation (σ) of each sequence are calculated, and data points deviating from the mean by more than 3σ are removed to eliminate outliers caused by sensor noise and transmission errors. Data gaps resulting from outlier removal are smoothed using cubic spline interpolation. This method maintains the continuity and smoothness of the data curves and avoids introducing abrupt changes. Numerical validity checks are performed on the pedal state sequences to ensure they contain only predefined state values (0 or 1), and invalid values are marked and uniformly removed. Monotonicity and continuity checks are performed on the timestamp sequences. Abnormal time intervals are identified by calculating the interval between adjacent timestamps, and interpolation or segment truncation is performed based on actual conditions to ensure strictly increasing timestamps. The time alignment and length consistency of each sequence are checked to ensure the synchronization of multi-source data.
[0042] 2.2 Sliding Window Feature Extraction Unit The sliding window technique is used to dynamically analyze continuous driving behavior data, and the time-varying characteristics of driving conditions are captured by statistical features within a local window.
[0043] The window parameter configuration provided in this solution is as follows: Window size: 15 sampling points; Sliding step size: 1 sampling point; Window positioning: The center position of the window corresponds to the output position of driving feature information at any given time.
[0044] Analyze the vehicle driving information at the center position of the window using all vehicle operation data within the window, and calculate the window-level average speed using the speed sequence of the input window. (km / h), used to characterize the operating conditions of a vehicle: In formula (5), WindowSize is the size of the sampled window, and the WindowSize defined in this invention is 15; The speed information (km / h) at the sampling time. This is the sampling time sequence within the window, which is 1 to 15 in this case.
[0045] The pedal activity level is calculated using the drive pedal status list and brake pedal status list from the input window to quantify the driver's driving style. In formula (6) and Each represents a list of drive / brake pedal states (Boolean value 0 or 1) for the sampled time sequence. and These represent the total number of drive / brake pedal activation states in the sampled time sequence (the number of states with a corresponding boolean value of 1 in the list). Calculate the window-level average current using the current sequence of the input window. (A) is used to characterize the charging and discharging state of a vehicle: In formula (7) This represents the current information (A) at the sampling time.
[0046] The average speed of the above three core features Pedal activity level and average current Together, they form the basis for the dynamic representation of driving information, which can effectively reflect the vehicle's operating characteristics under different driving conditions.
[0047] 2.3 One-hot encoding output unit Based on feature vectors extracted using a sliding window, a hierarchical driving condition classification system is established, discretizing continuous driving behavior data into standardized driving condition categories. The following are the hierarchical classification rules for driving conditions: Level 1: Special Working Condition Identification (1) Parking and charging conditions: Level 2: Driving Status Judgment (1) Idle speed when parked: (2) Driving at low speed: (3) Medium speed driving: (4) High-speed driving: Level 3: Driving Style Recognition (1) Gentle driving: (2) Aggressive driving: Through the above classification, eight standardized operating condition combinations are formed: parking and charging, parking and idling, low speed and gentle, low speed and aggressive, medium speed and gentle, medium speed and aggressive, high speed and gentle, and high speed and aggressive.
[0048] Each operating condition corresponds to an 8-dimensional one-hot encoded vector: Table 1. Operating condition characterization of the eight-dimensional one-hot encoding When the sampling frequency is 0.1Hz (10 seconds / point) and the window size is 15 sampling points: The window coverage time span is 15 × 10 = 150 seconds; the output update frequency is one working condition code every 10 seconds.
[0049] This encoded output provides standardized operating condition input for the subsequent battery state estimation module, ensuring the accuracy and adaptability of battery aging estimation under different operating conditions. The operating condition category at the initial position of the window is set to medium speed and mild [0,0,0,0,1,0,0,0] by default. Since this part of the data accounts for a small portion of the data generated by the actual operation of the vehicle, it will not affect the overall joint estimation effect.
[0050] 3. Battery SOC estimation module The deep neural network (DNN) model used in this invention comprises four core modules: an input feature preprocessing layer, a deep residual feature extraction layer, a feature compression and nonlinear mapping layer, and a SOC regression output layer. This architecture, by introducing residual connections, significantly improves training stability while maintaining the model's expressive power, achieving high-precision and robust estimation of the battery's state of charge under complex operating conditions.
[0051] 3.1 Input Feature Preprocessing Layer The input feature preprocessing layer is responsible for receiving and fusing multi-dimensional features from different sources with different physical meanings and statistical properties. This layer employs differentiated preprocessing strategies to address the differences in feature data structures, ultimately forming a unified, normalized feature vector that provides high-quality input for subsequent deep networks.
[0052] The raw input features include three main categories: real-time sensor time-series features, global battery health features, and driving condition category features.
[0053] Real-time sensor timing characteristics include total voltage (volt), minimum / maximum single-cell voltage (min_volt / max_volt), current, and ampere-hour integral (Ah).
[0054] Global characteristics of battery health, i.e., predicted battery capacity This characteristic remains constant within a segment, characterizing the overall health of the battery.
[0055] The driving condition category feature is a driving pattern represented by one-hot encoding, which is a sparse eight-dimensional vector.
[0056] Real-time sensor time-series characteristics Z-score standardization is performed to eliminate dimensional differences and accelerate model convergence: in, The mean vector of continuous features in the training set. This is the standard deviation vector of continuous features in the training set. (Standardized features) It is directly fed into subsequent deep network layers, providing a stable and normalized input basis for model training.
[0057] Battery Predicted Capacity This characterizes the aging state of a specific battery segment. To prevent its numerical range from dominating model training, it is also standardized. In formula 9, For battery capacity The average value on the training dataset; For battery capacity Standard deviation on the training dataset; This is the standardized battery capacity value.
[0058] While the original one-hot encoded vectors are discrete values, directly inputting them into the model would introduce unnecessary sparsity and high dimensionality. Therefore, this invention performs dimensionality reduction and densification processing, converting them into a single working condition category identifier (i.e., taking argmax): In formula 10 An eight-dimensional one-hot encoded vector representing driving condition information; The maximum value index operation returns the index of the maximum value in the input vector. The operating condition category identifier is an integer calculated from an eight-dimensional one-hot encoded vector, used to replace the original eight-dimensional one-hot encoded vector input model.
[0059] This identifier is then used as an independent integer feature in subsequent processing. This method greatly simplifies the model input structure while preserving key operating condition information. All features that have undergone the above specific preprocessing are concatenated along the feature dimension to form a unified, dense input feature vector: in, This represents a vector concatenation operation. The final fused feature vector is then fed into a subsequent deep residual feature extraction layer.
[0060] 3.2 Deep Residual Feature Extraction Layer The deep residual feature extraction layer is the core of this module. It aims to address the gradient vanishing and network degradation problems caused by excessive depth in deep neural networks by introducing a residual connection mechanism, thereby constructing a deeper and more powerful feature extraction network. This layer consists of a feature projection layer and two cascaded residual blocks.
[0061] The preprocessed fused feature vectors are fed into a fully connected layer to increase their dimension to 128, thereby enhancing the initial expressive power of the model. This process can be represented as: in, and The weights and biases of the projection layer are then set. The output is then... Batch normalization and dropout with a ratio of 0.2 are used to improve training stability and generalization ability.
[0062] The first residual block consists of two 96-dimensional fully connected layers, each followed by batch normalization. The specific steps are as follows: enter The process sequentially passes through a 96-dimensional fully connected layer, batch normalization, ReLU activation function, and Dropout layer, before entering a second 96-dimensional fully connected layer and batch normalization. This series of nonlinear transformations aims to learn the desired residual function. .because Dimensions (128) and main path output The dimension (96) does not match, and this invention uses a dedicated linear projection matrix. right Dimensionality reduction is performed to make it consistent with the output dimension of the main path. This process is represented as follows: The output of the main path is added element-wise to the output of the shortcut path, and then the ReLU activation function is applied to obtain the final output of the first residual block. : This operation enables the effective fusion of the original input information and the learned residual information.
[0063] Output of the first residual block After passing through a 64-dimensional fully connected layer, the input is fed into the second residual block. This block adopts a similar but simplified structure to the first residual block: the input sequentially passes through two 64-dimensional fully connected layers to learn the residual function. Since both the input and output dimensions of this block are 64, the shortcut connection does not require linear projection and can directly use the identity mapping, i.e., directly pass the input. Similarly, the outputs of the two paths are summed and activated to obtain the output of the second residual block: 3.3 Feature Compression and Nonlinear Mapping Layer The purpose of this layer is to reduce and refine the features while preserving core information, and to complete the nonlinear mapping transition to the final regression target. This layer receives the output from the second residual block. The data flows through three fully connected layers in sequence. The data transfer process is as follows: enter First, it is transformed to a lower dimension, and then non-linear activation is performed.
[0064] in, It is a [64×32] weight matrix. This is the bias vector. This operation compresses the feature dimension from 64 to 32. Then, the output of the first layer... Further compressed and refined: in, The output of the second compression layer is... After further compression and refinement, the dimension of the feature matrix was reduced from 32 to 16. The bias vector for the second compression layer (dimension [1×16]) provides independent bias adjustments for each output dimension; It is a [32×16] weight matrix. The feature dimension is further compressed to 16. The final mapping layer does not perform dimensionality compression, but instead performs a final nonlinear combination and mapping on the refined 16-dimensional features to prepare for the output layer.
[0065] in, The bias vector for the third layer (with dimensions [1×16]) provides independent bias adjustments for each output dimension; It is a [16×16] weight matrix. Output It is a 16-dimensional, highly refined feature vector.
[0066] 3.4 SOC Regression Output Layer The output layer is a single-neuron fully connected layer responsible for mapping the compressed feature representation to the final SOC estimate: in, It is a trainable weight vector of dimension [16×1], realizing a linear combination from 16-dimensional features to 1-dimensional SOC values; Provides a baseline offset for SOC estimation as a trainable bias scalar; This is the estimated state of charge (SOC) value output by the model. This layer uses a linear activation function to directly regress the SOC percentage.
[0067] 4. Battery SOC-SOH Dynamic Prediction Coupling Mechanism The integrated battery aging estimation module outputs the current maximum usable battery capacity in real time. The battery SOC estimation module directly inputs the data as features, and the dynamic capacity Q, which changes over time, provides an accurate dynamic representation of the battery's aging state for SOC estimation. The battery SOC estimation module outputs precise initial and final SOC values, and uses the estimated SOC difference to calculate the change in charge, which is then input into the integrated battery aging estimation module as features. This replaces the method of simply calculating charge using ampere-hour integration, avoiding the cumulative error problem caused by the ampere-hour integration method.
[0068] This invention provides an adaptive multi-condition battery state joint estimation system; the specific technical approach is detailed below. Figure 2 This includes the following steps: (1) Dynamic representation of driving information Driving features are dynamically extracted using a sliding window technique (15 sampling points per window, 1 sampling point per sliding step), including window-level average speed, pedal activity (calculated from accelerator and brake pedal states), and average current. Based on preset classification rules, the current driving state is divided into eight typical operating conditions: low-speed-mild, low-speed-aggressive, medium-speed-mild, medium-speed-aggressive, high-speed-mild, high-speed-aggressive, parking and charging, and parking and idling, generating corresponding 8-dimensional one-hot encoded vectors. This encoded output is updated every 10 seconds, providing standardized operating condition context information for battery state estimation.
[0069] (2) Estimation of the aging state of integrated batteries An integrated battery aging estimation module based on a battery capacity degradation model and a LightGBM model was established, using the battery's current maximum available capacity Q as the standard for evaluating the battery's aging state. To ensure the accuracy of the battery capacity results, the capacity change was updated using segments with a SOC change greater than 20%, and the validity of the prediction results was verified to ensure that the capacity value was within a reasonable physical range.
[0070] (3) Joint estimation of SOC-SOH The integrated battery aging estimation module outputs the current maximum usable battery capacity Q(t) in real time, which is directly input into the SOC estimation module as an input feature. The dynamic capacity Q, which changes over time, provides an accurate dynamic representation of the battery aging state for SOC estimation. The SOC estimation module outputs fine-grained initial and final SOC values and uses the estimated SOC difference to calculate the change in charge, which is then input into the integrated battery aging estimation module as an input feature.
[0071] The data used for verification in this invention is based on a dataset of real-world operation of pure electric vehicles. This dataset was collected from data of 20 pure electric vehicles operating on actual roads, and includes complete multi-dimensional information such as vehicle status, battery system, and driving behavior. It is characterized by strong authenticity, complete data, and rich scenarios, providing reliable data support for the verification of this invention.
[0072] Figure 3 The calibration results of the constructed battery capacity degradation model that takes into account time degradation and charge / discharge habits are presented.
[0073] Figure 4 The prediction results of an integrated battery aging estimation model based on the battery capacity degradation model and the LGBM model are presented.
[0074] Figure 5 The training results of the constructed deep learning network model are presented. On the validation set, the model exhibits excellent fitting ability, with a mean absolute error (MAE) of 1.4792% and a root mean square error (RMSE) of 2.1472% for SOC predictions, indicating that the model has good basic prediction accuracy.
[0075] The above verification results show that the method proposed in this invention has more reliable engineering applicability and generalization ability compared with traditional methods under real road conditions.
[0076] Based on the above analysis, this invention aims to precisely address the core pain point of existing technologies, namely, how to construct a battery state estimation system that combines physical reliability and data adaptability under conditions of small sample vehicle data: An integrated battery aging estimation model is proposed. To address the issues of low-quality and limited battery capacity data, a battery capacity degradation model considering time decay and charging / discharging habits is established and refined using existing limited data. This is achieved by combining a physically consistent battery capacity degradation model with a lightweight LightGBM machine learning model. The former ensures that the prediction results conform to physical common sense, while the latter learns to capture difficult-to-capture nonlinear factors to correct the physical model. This architecture combines the advantages of both physical mechanisms and data-driven approaches, overcoming the "mismatch" problem of pure model methods and solving the "small sample size" and "black box" problems of pure data-driven methods. This enables high-precision and interpretable estimation of SOH (State of Health) with limited sample data.
[0077] A dual-state coupled estimation framework of SOC-SOH is constructed. On the one hand, the current maximum usable capacity Q of the battery is dynamically updated using the integrated battery aging estimation model mentioned above, and fed back to the SOC estimation stage to characterize the battery aging state. On the other hand, the SOC estimation stage calculates the SOC change value, combines it with the current maximum usable capacity Q to dynamically update the charge change E, and feeds it back to the integrated battery aging estimation model to update the battery aging state, thus solving the cumulative error problem caused by the ampere-hour integration method in calculating charge change. This framework corrects the cumulative SOC error caused by capacity decay and the traditional ampere-hour integration method, ensuring the long-term accuracy of SOC estimation throughout the entire battery life cycle.
[0078] We design a "feature engineering" system for real-time driving environments, employing a sliding window to capture dynamic driving information sequences and using one-hot encoding to represent the driving information. This feature engineering enables the SOC estimation process to perceive and adapt to different energy consumption patterns, thereby achieving high-precision and robust joint prediction of SOC and SOH in various real-world driving scenarios.
[0079] Example 2 This invention also provides an adaptive multi-condition battery SOC-SOH joint estimation method, which is implemented using the system described in Embodiment 1. The method includes: By using multi-dimensional feature extraction and sliding window analysis technology, the driving conditions of vehicles can be identified and quantified in real time. An integrated battery aging estimation module based on the battery capacity degradation model and the LightGBM residual correction model was established, using the current maximum available capacity of the battery as the standard for evaluating the battery aging state. Based on driving conditions, the current maximum usable battery capacity output in real time by the integrated battery aging estimation module is directly input into the battery SOC estimation module as input features. The dynamic capacity that changes over time provides a dynamic representation of the battery aging state for SOC estimation. The battery SOC estimation module outputs the initial and final SOC values and uses the estimated SOC difference to calculate the change in charge, which is then input into the integrated battery aging estimation module as input features to update the battery aging state.
[0080] In this embodiment, the method for real-time identification and quantification of vehicle driving conditions through multi-dimensional feature extraction and sliding window analysis technology includes: The collected driving behavior data is standardized and converted to a new format, and then cleaned. The sliding window technique is used to dynamically analyze continuous driving behavior data after cleaning, and the time-varying characteristics of driving conditions are captured by statistical features within a local window. Based on the feature vectors extracted by the sliding window, a hierarchical working condition classification system is established to discretize continuous driving information into standardized working condition categories.
[0081] In this embodiment, the method of dynamically analyzing continuous driving behavior data after cleaning using sliding window technology and capturing the time-varying characteristics of driving conditions through statistical features within a local window includes: ; ; ; In the formula, For average speed, To determine the activity level of the pedals, The average current is given, and WindowSize is the size of the sampling window. For the velocity information at the sampling time, This is the sampling time sequence within the window. and These are lists of drive / brake pedal states for the sampled time sequences. and These represent the total number of drive / brake pedal activation states in the sampled time sequence. This refers to the current information at the sampling time.
[0082] In this embodiment, the battery capacity degradation model is as follows: ; In the formula, for Physically predicted battery capacity at the given time; This refers to the battery's rated capacity at the time of manufacture. This is the time decay coefficient; This refers to the cumulative battery usage time. To account for the degradation coefficient due to charging and discharging habits; For each charge / discharge cycle of the battery; LightGBM residual correction model: ; In the formula, This represents the actual maximum usable capacity of the battery at present. For residuals; Final battery aging condition estimate: .
[0083] In this embodiment, the battery SOC estimation module adopts a deep neural network (DNN) model, which includes: an input feature preprocessing layer, a deep residual feature extraction layer, a feature compression and nonlinear mapping layer, and an SOC regression output layer. The input feature preprocessing layer is used to preprocess and fuse real-time sensor time-series features, global battery health features, and driving condition category features to obtain the fused feature vector. A deep residual feature extraction layer is used to extract the deep spatiotemporal features of the battery dynamics through residual connections based on the fused feature vectors. Feature compression and nonlinear mapping layers are used to reduce the dimensionality of deep spatiotemporal features and complete the nonlinear mapping transition to the final regression target; The SOC regression output layer is used to map the compressed feature representation to the final SOC estimate.
[0084] In summary, the advantages achieved by this invention are as follows: 1. An integrated battery aging estimation paradigm. It does not rely on pure black-box models or pure physical models, but instead uses a physical model for basic battery aging prediction and lightweight machine learning (LGBM) to learn its residuals. This solves the problem of "small sample" battery capacity data in the vehicle environment, while ensuring the model's interpretability and ability to capture complex nonlinear factors. 2. Mutual correction between SOC and SOH. The estimation result of SOH (represented here by the battery's current maximum usable capacity Q) is dynamically and in real-time fed back to the SOC estimation module. The SOC estimation module, using rich features as input, outputs SOC, calculates the SOC change value, and uses Q to calculate the charge change, avoiding the cumulative error caused by directly using ampere-hour integration, and providing clean basic data for SOH estimation. 3. Embedded feature representation of real-time dynamic driving conditions. Through sliding windows and regularized feature engineering, continuous driving behavior is transformed into standardized numerical features in real time and encoded as input that the model can understand, rather than performing complex pattern classification that relies on historical data.
[0085] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.
Claims
1. An adaptive multi-condition battery SOC-SOH joint estimation system, characterized in that, The system includes: an integrated battery aging estimation module, a driving information dynamic characterization module, and a battery SOC estimation module; The integrated battery aging estimation module is used to estimate the battery health status. The driving information dynamic representation module is used to identify and quantify the vehicle's driving conditions in real time through multi-dimensional feature extraction and sliding window analysis technology. The battery SOC estimation module is used to estimate the battery's state of charge based on the battery health status estimation results and the vehicle's driving conditions. It also uses the SOC estimation step to calculate the SOC change value, dynamically updates the charge change in combination with the battery's current maximum available capacity, and feeds it back to the integrated battery aging estimation module to update the battery aging status.
2. The system according to claim 1, characterized in that, The integrated battery aging estimation module includes: a battery capacity degradation model and a LightGBM residual correction model; Among them, the battery capacity degradation model: ； In the formula, for Physically predicted battery capacity at any given time; This refers to the battery's rated capacity at the time of manufacture. This is the time decay coefficient; This refers to the cumulative battery usage time. To account for the degradation coefficient due to charging and discharging habits; For each charge / discharge cycle of the battery; LightGBM residual correction model: ； In the formula, This represents the actual maximum usable capacity of the battery at present. For residuals; Final battery aging condition estimate: 。 3. The system according to claim 1, characterized in that, The dynamic representation module for driving information includes: a data cleaning unit, a sliding window feature extraction unit, and a one-hot encoding output unit; The data cleaning unit is used to standardize and clean the collected driving behavior data. The sliding window feature extraction unit is used to dynamically analyze the cleaned continuous driving behavior data using sliding window technology, and capture the time-varying characteristics of driving conditions through statistical features within a local window. The one-hot encoding output unit is used to establish a hierarchical driving condition classification system based on the feature vector extracted by the sliding window, and to discretize continuous driving information into standardized driving condition categories.
4. The system according to claim 3, characterized in that, The process of dynamically analyzing continuous driving behavior data after cleaning using sliding window technology, and capturing the time-varying characteristics of driving conditions through statistical features within a local window, includes: ； ； ； In the formula, For average speed, To determine the activity level of the pedals, The average current is given, and WindowSize is the size of the sampling window. For the velocity information at the sampling time, This is the sampling time sequence within the window. and These are lists of drive / brake pedal states for the sampled time sequences. and These represent the total number of drive / brake pedal activation states in the sampled time sequence. This represents the current information at the sampling time.
5. The system according to claim 1, characterized in that, The battery SOC estimation module adopts a deep neural network (DNN) model, which includes: an input feature preprocessing layer, a deep residual feature extraction layer, a feature compression and nonlinear mapping layer, and an SOC regression output layer. The input feature preprocessing layer is used to preprocess and fuse real-time sensor time-series features, battery health global features, and driving condition category features to obtain a fused feature vector. The deep residual feature extraction layer is used to extract the dynamic deep spatiotemporal features of the battery through residual connections based on the fused feature vector. The feature compression and nonlinear mapping layer is used to reduce the dimensionality of deep spatiotemporal features and complete the nonlinear mapping transition to the final regression target. The SOC regression output layer is used to map the compressed feature representation to the final SOC estimate.
6. An adaptive multi-condition battery SOC-SOH joint estimation method, said method being implemented using the system described in any one of claims 1-5, characterized in that, The method includes: By using multi-dimensional feature extraction and sliding window analysis technology, the driving conditions of vehicles can be identified and quantified in real time. An integrated battery aging estimation module based on the battery capacity degradation model and the LightGBM residual correction model was established, using the current maximum available capacity of the battery as the standard for evaluating the battery aging state. Based on driving conditions, the current maximum usable battery capacity output in real time by the integrated battery aging estimation module is directly input into the battery SOC estimation module as input features. The dynamic capacity that changes over time provides a dynamic representation of the battery aging state for SOC estimation. The battery SOC estimation module outputs the initial and final SOC values and uses the estimated SOC difference to calculate the change in charge, which is then input into the integrated battery aging estimation module as input features to update the battery aging state.
7. The method according to claim 6, characterized in that, Methods for real-time identification and quantification of vehicle driving conditions using multi-dimensional feature extraction and sliding window analysis techniques include: The collected driving behavior data is standardized and converted to a new format, and then cleaned. The sliding window technique is used to dynamically analyze continuous driving behavior data after cleaning, and the time-varying characteristics of driving conditions are captured by statistical features within a local window. Based on the feature vectors extracted by the sliding window, a hierarchical working condition classification system is established to discretize continuous driving information into standardized working condition categories.
8. The method according to claim 7, characterized in that, Methods for dynamically analyzing continuous driving behavior data after cleaning using the sliding window technique, and capturing the time-varying characteristics of driving conditions through statistical features within a local window, include: ； ； ； In the formula, For average speed, To determine the activity level of the pedals, The average current is given, and WindowSize is the size of the sampling window. For the velocity information at the sampling time, This is the sampling time sequence within the window. and These are lists of drive / brake pedal states for the sampled time sequences. and These represent the total number of drive / brake pedal activation states in the sampled time sequence.
9. The method according to claim 6, characterized in that, Battery capacity degradation model: ； In the formula, for Physically predicted battery capacity at any given time; This refers to the battery's rated capacity at the time of manufacture. This is the time decay coefficient; This refers to the cumulative battery usage time. To account for the degradation coefficient due to charging and discharging habits; For each change in the amount of charge / discharge of the battery, This refers to the current information at the sampling time. LightGBM residual correction model: ； In the formula, This represents the actual maximum usable capacity of the battery at present. For residuals; Final battery aging condition estimate: 。 10. The method according to claim 6, characterized in that, The battery SOC estimation module adopts a deep neural network (DNN) model, which includes: an input feature preprocessing layer, a deep residual feature extraction layer, a feature compression and nonlinear mapping layer, and an SOC regression output layer. The input feature preprocessing layer is used to preprocess and fuse real-time sensor time-series features, battery health global features, and driving condition category features to obtain a fused feature vector. The deep residual feature extraction layer is used to extract the dynamic deep spatiotemporal features of the battery through residual connections based on the fused feature vector. The feature compression and nonlinear mapping layer is used to reduce the dimensionality of deep spatiotemporal features and complete the nonlinear mapping transition to the final regression target. The SOC regression output layer is used to map the compressed feature representation to the final SOC estimate.