Method and system for predicting health degree of vehicle-mounted battery of electric vehicle

Abstract

The invention discloses a method and system for predicting the health degree of a vehicle-mounted battery of an electric vehicle, and relates to the technical field of electric vehicle battery management, and the method comprises the steps: collecting multi-source data, and carrying out the preprocessing of the collected multi-source data; extracting features from the preprocessed data and screening the features; constructing a hybrid model fusing deep learning and reinforcement learning according to the screened features; historical data of cross-brand and cross-model electric vehicle batteries under various working conditions are collected to train the hybrid model to obtain an optimal health degree prediction model; and collecting real-time data and inputting the real-time data into the optimal health degree prediction model to obtain an SOH estimation and prediction result. According to the method, multi-source data and feature engineering are combined, driving habits and environmental factors are considered, and the accuracy of battery health degree estimation is improved. A hybrid model of deep learning and reinforcement learning is adopted, the robustness and generalization ability of the model are enhanced, and the method is suitable for different batteries and complex working conditions. The model supports on-line updating and tracks battery performance in real time.

Description

Technical Field

[0001] The present invention relates to the technical field of electric vehicle battery management, and more specifically, to a method and system for predicting the health state of on-vehicle batteries of electric vehicles. Background Art

[0002] With the popularization of electric vehicles, the battery, as a core component, the accurate estimation and prediction of its health state (SOH) are crucial. The SOH is directly related to the driving range, safety, and overall performance of electric vehicles. In the prior art, for traditional model-based methods, it is difficult to model the complex physical and chemical processes of the battery, and it is difficult to adapt to the characteristic changes of the battery under different working conditions; while for data-driven methods, when facing noise interference, data loss, and individual differences of different batteries, the prediction accuracy and generalization ability are poor. Moreover, most of the existing methods do not fully consider the dynamic factors in the battery usage process, such as the influence of user driving habits and environmental temperature changes on the battery health state.

[0003] Therefore, how to propose a method and system for predicting the health state of on-vehicle batteries of electric vehicles, overcome the deficiencies of the prior art, fully consider the dynamic factors in the battery usage process, and improve the safety and overall performance of electric vehicles is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0004] In view of this, the present invention provides a method and system for predicting the health state of on-vehicle batteries of electric vehicles, which overcome the deficiencies of the prior art, improve the accuracy, robustness, and generalization ability of SOH estimation and prediction, fully consider the dynamic factors in the battery usage process, provide reliable data support for the electric vehicle battery management system, and improve the safety and overall performance of electric vehicles. To achieve the above objectives, the present invention adopts the following technical solutions:

[0005] A method for predicting the health state of on-vehicle batteries of electric vehicles includes:

[0006] Collect multi-source data and preprocess the collected multi-source data;

[0007] Extract features from the preprocessed data and perform screening;

[0008] Construct a hybrid model that combines deep learning and reinforcement learning according to the screened features;

[0009] Collect historical data of electric vehicle batteries of different brands and models under various working conditions to train the hybrid model to obtain an optimal health state prediction model;

[0010] Collect real-time data and input it into the optimal health state prediction model to obtain SOH estimation and prediction results.

[0011] Optionally, the multi-source data includes: battery voltage, current, temperature, number of charge and discharge cycles, cumulative driving mileage, acceleration, speed change, number of hard brakes, duration of different driving modes, as well as environmental temperature and humidity.

[0012] Optionally, the preprocessing of the collected multi-source data includes: using an adaptive filtering algorithm to remove noise interference, using an interpolation algorithm to fill in missing data values, and for abnormal data, identifying and correcting it through an anomaly detection model based on statistical analysis and machine learning.

[0013] Optionally, the extraction of features from the preprocessed data includes: extracting features of voltage and current change rates, correlation features between hard acceleration frequency and battery current impact, battery power consumption features under different driving modes, and the impact of environmental temperature on battery internal resistance change from the preprocessed data.

[0014] Optionally, the screening includes: using a method that combines principal component analysis and mutual information analysis to perform dimensionality reduction and screening on the extracted features, removing redundant features, and retaining key features that have a significant impact on SOH.

[0015] Optionally, the construction of a hybrid model that combines deep learning and reinforcement learning based on the screened features includes: the deep learning part uses an improved convolutional long short-term memory network, automatically learning local features in battery data using a convolutional neural network, and then capturing the time series features of the data through a long short-term memory network; the reinforcement learning part introduces a deep Q network, using the current state and operations of the battery as inputs, and through continuous interaction with the environment, learning the optimal SOH estimation and prediction strategy, and dynamically adjusting the model parameters.

[0016] Optionally, it further includes: introducing a Dropout layer and Batch Normalization technology in the improved convolutional long short-term memory network, and at the same time using a regularization method to optimize the loss function of the deep Q network to optimize the overfitting problem in model training.

[0017] Optionally, the training of the hybrid model using historical data of electric vehicle batteries of different brands and models under various working conditions includes: collecting historical data of electric vehicle batteries of different brands and models under various working conditions, including normal use data and fault data, to construct a training data set;

[0018] Using the training data set to perform supervised training on the hybrid model, using the actually measured SOH value as a label, minimizing the error between the model prediction value and the true value, and continuously updating the model parameters through the backpropagation algorithm to enable the model to learn the relationship between battery health and various input features.

[0019] Optionally, it further includes: when new battery data is received, using an incremental learning algorithm to gradually integrate the new data into model training and dynamically adjust the model parameters.

[0020] Optionally, an in-vehicle battery health prediction system for electric vehicles includes:

[0021] An acquisition and processing module: used to acquire multi-source data and preprocess the acquired multi-source data;

[0022] A screening module: used to extract and screen features from the preprocessed data;

[0023] A model construction module: used to construct a hybrid model that integrates deep learning and reinforcement learning based on the screened features;

[0024] A training module: used to collect historical data of electric vehicle batteries across different brands and models under various working conditions to train the hybrid model to obtain an optimal health prediction model;

[0025] A prediction module: used to acquire real-time data and input it into the optimal health prediction model to obtain SOH estimation and prediction results.

[0026] As can be seen from the above technical solutions, compared with the prior art, the present invention discloses an in-vehicle battery health prediction method and system for electric vehicles, having the following beneficial effects:

[0027] The present invention proposes an in-vehicle battery health prediction method for electric vehicles, including: acquiring multi-source data and preprocessing the acquired multi-source data; extracting and screening features from the preprocessed data; constructing a hybrid model that integrates deep learning and reinforcement learning based on the screened features; collecting historical data of electric vehicle batteries across different brands and models under various working conditions to train the hybrid model to obtain an optimal health prediction model; acquiring real-time data and inputting it into the optimal health prediction model to obtain SOH estimation and prediction results. Through multi-source data acquisition and innovative feature engineering, the present invention fully considers the influence of dynamic factors such as user driving habits and environmental factors on battery health. Compared with traditional methods, it can more comprehensively and accurately reflect the actual health status of the battery, improving the accuracy of SOH estimation and prediction. The hybrid model that integrates deep learning and reinforcement learning combines the advantages of both. It can not only automatically learn complex features and time series relationships in battery data but also learn the optimal prediction strategy through interaction with the environment, significantly enhancing the robustness and generalization ability of the model and being able to adapt to different battery individual differences and various complex working conditions. The model has the ability to update online, can track the performance changes of the battery in real time, timely adjust the prediction results, provide more reliable and dynamic support for electric vehicle battery management, help extend the battery service life, reduce the user's usage cost, and promote the development of the electric vehicle industry. Brief Description of the Drawings

[0028] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained according to the provided drawings without creative efforts.

[0029] Figure 1 It is a schematic flow chart of a method for predicting the health state of an in-vehicle battery of an electric vehicle provided by the present invention.

[0030] Figure 2 It is a structural framework diagram of a system for predicting the health state of an in-vehicle battery of an electric vehicle provided by the present invention. Detailed Embodiments

[0031] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0032] An embodiment of the present invention discloses a method for predicting the health state of an in-vehicle battery of an electric vehicle, as Figure 1 shown, including:

[0033] Collect multi-source data and preprocess the collected multi-source data;

[0034] Extract features from the preprocessed data and perform screening;

[0035] Construct a hybrid model that combines deep learning and reinforcement learning according to the screened features;

[0036] Collect historical data of electric vehicle batteries of different brands and models under various working conditions to train the hybrid model to obtain an optimal health state prediction model;

[0037] Collect real-time data and input it into the optimal health state prediction model to obtain SOH estimation and prediction results.

[0038] Furthermore, the multi-source data includes: traditional parameters such as the voltage, current, temperature, charge and discharge times, and cumulative driving mileage of the battery. At the same time, data reflecting the user's driving habits such as the acceleration, speed change, number of emergency brakes, and duration of different driving modes during vehicle driving are collected, as well as environmental data such as environmental temperature and humidity.

[0039] Further, the preprocessing of the collected multi-source data includes: using an adaptive filtering algorithm to remove noise interference, using an interpolation algorithm to fill in missing data values, and for abnormal data, establishing an anomaly detection model based on statistical analysis and machine learning for identification and correction to ensure the accuracy and integrity of the input data.

[0040] In the specific implementation manner, the specific steps of preprocessing the collected multi-source data include:

[0041] S1: Adaptive filtering algorithm, by iteratively adjusting the filter coefficients to minimize the mean square error between the output signal and the desired signal:

[0042] (1) Initialization

[0043] Initialize the input signal as x(n), the observed signal with noise as d(n) = s(n) + v(n), where s(n) is the true signal and v(n) is the noise;

[0044] Initialize the filter weight coefficient vector as w(0) = [w0(0), w1(0), …, w M (0)]T, where M is the filter order;

[0045] (2) Iteratively update the weight coefficients

[0046] For the nth iteration, the input vector is x(n) = [x(n), x(n - 1), …, x(n - M)] T ;

[0047] Calculate the filter output: y(n) = w T (n - 1)x(n);

[0048] Calculate the error: e(n) = d(n) - y(n);

[0049] Update the weight coefficients: w(n) = w(n - 1) + 2μe(n)x(n), where μ is the step size factor.

[0050] (3) Termination condition

[0051] When the error e(n) is less than the set threshold or the maximum number of iterations is reached, stop the iteration and output the denoised signal y(n).

[0052] S2: The goal of the interpolation algorithm is to fill in the missing data points by piecewise cubic polynomial fitting to ensure the smoothness and continuity of the curve.

[0053] (1) Data preparation

[0054] Let the known data points be (x i , y i)(i = 0, 1, …, N), where x0 < x1 <... < x N , missing values need to be filled at the missing value

[0055] (2) Construct a spline function

[0056] For each interval [x i , x i+1 , define a cubic polynomial:

[0057] S i (x) = a i + b i (x - x i ) + c i (x - x i ) 2 + d i (x - x i ) 3 ;

[0058] Satisfy the following conditions:

[0059] Continuity: S i (x i+1 ) = S i+1 (x i+1 );

[0060] First derivative continuity: S′ i (x i+1 ) = S′ i+1 (x i+1 );

[0061] Second derivative continuity: S″ i (x i+1 ) = S″ i+1 (x i+1 );

[0062] Boundary conditions (taking natural boundary as an example): S″0(x0) = 0, S″ N (x N ) = 0.

[0063] (3) Solve for the coefficients

[0064] Let the second derivative M i = S″ i (x i ), solve for M i using the tridiagonal matrix method, and then calculate the coefficients:

[0065]

[0066] (4) Calculate the missing values

[0067] Determine The interval [x k , x k+1 , substitute it into the corresponding S k (x) for calculation

[0068] S3: Anomaly Detection and Correction

[0069] 1) Statistical analysis method (Z-score detection)

[0070] (1) Calculate the statistic Calculate the data mean μ and standard deviation σ;

[0071]

[0072] 2) Detect outliers, set a threshold k (e.g., k = 3), if ∣x i - μ∣ > kσ, then it is determined as an outlier.

[0073] 3) Correct the outlier, replace it with the mean or median:

[0074] (2) Machine learning method

[0075] Construct an isolation tree, randomly select a feature and a random split point of this feature, recursively partition the data until the sample is isolated or reaches the maximum tree depth. Calculate the "path length" h(x) of each sample, which represents the number of splits from the root node to the leaf node.

[0076] Calculate the anomaly score:

[0077]

[0078] Among them, E(h(x)) is the average path length, c(n) is the normalization constant, and γ≈0.5772 is the Euler constant.

[0079] Correct the outlier, for the samples with scores higher than the threshold, replace them with the normal values predicted by the K-nearest neighbor (KNN) algorithm:

[0080] S4: Data Accuracy and Integrity Verification

[0081] Accuracy verification, calculate the root mean square error (RMSE) between the denoised signal and the original true signal:

[0082]

[0083] Integrity verification, check the missing value filling rate: Filling rate = (Number of filled missing values / Total number of missing values) × 100%

[0084] Through the above formula steps, denoising, completion, and anomaly correction of multi-source data are achieved to ensure the accuracy and integrity of data preprocessing.

[0085] Further, the extraction of features from the preprocessed data includes: extracting features of voltage and current change rates from the preprocessed data, innovatively extracting features related to the user's driving habits, such as the correlation feature between the rapid acceleration frequency and the battery current impact, the battery power consumption features under different driving modes, etc., and the correlation features between environmental factors and battery performance, such as the influence feature of environmental temperature on the change of battery internal resistance.

[0086] Further, the screening includes: adopting a method combining principal component analysis and mutual information analysis to perform dimensionality reduction and screening on the extracted features, removing redundant features, and retaining the key features that have a significant impact on SOH, reducing the subsequent model calculation complexity, and improving the model training efficiency and accuracy.

[0087] Further, the construction of a hybrid model integrating deep learning and reinforcement learning based on the screened features includes: the deep learning part uses an improved convolutional long short-term memory network (Conv-LSTM). Utilizing the powerful feature extraction ability of the convolutional neural network (CNN), it automatically learns the local features in the battery data, and then captures the time series features of the data through the long short-term memory network (LSTM) to effectively process the dynamic characteristics of battery health changes; the reinforcement learning part introduces a deep Q network (DQN). Taking the current state and operations of the battery (such as charging and discharging strategies) as inputs, through continuous interaction with the environment (simulating the operation of the battery under different working conditions), it learns the optimal SOH estimation and prediction strategies, dynamically adjusts the model parameters, and improves the adaptability of the model to different usage scenarios.

[0088] Further, it also includes: introducing a Dropout layer and Batch Normalization technology into the improved convolutional long short-term memory network, and at the same time using a regularization method to optimize the loss function of the deep Q network to optimize the overfitting problem in model training.

[0089] In the specific implementation manner, the construction of a hybrid model integrating deep learning and reinforcement learning based on the screened features specifically includes:

[0090] S1: Deep learning part: Improved Conv-LSTM network

[0091] (1) Convolutional layer (CNN feature extraction)

[0092] Input: Battery data sequence

[0093] Among them, T is the time step, H and W are the spatial dimensions, and C is the number of channels, such as voltage, current, temperature, etc.

[0094] Convolution operation:

[0095] Among them, K i is the convolution kernel (learning parameter) at the i-th time step, with a size of k×k×C×F (F is the number of output channels); * is the convolution operation; b u is the bias term; f is the activation function (such as ReLU).

[0096] Output: Local feature map After pooling, the spatial dimensions are compressed to H′ and W′.

[0097] (2) LSTM layer (temporal feature capture)

[0098] Input: Convolution feature sequence {U1, U2, …, U T}

[0099] LSTM cell state update:

[0100] i t = σ(W xi U t + W hi h t-1 + W ci ⊙ c t-1 + b i ),

[0101] f t = σ(W xf U t + W hf h t-1 + W cf ⊙ c t-1 + b f ),

[0102] 0 t = σ(W xo U t + W ho h t-1 + W co ⊙ c t + b o ),

[0103]

[0104] h t = o t ⊙ tanh(c t ).

[0105] Among them, i t, f t , o t are the input gate, forget gate, and output gate; c t , h t are the cell state and hidden state; σ is the Sigmoid activation function, ⊙ is element-wise multiplication; W is the weight matrix, and b is the bias term.

[0106] Output: Temporal feature h T ∈R N , where N is the dimension of the LSTM hidden layer.

[0107] S2: Reinforcement learning part: Deep Q-network (DQN)

[0108] (1) State and action space definition

[0109] State s t : includes the LSTM output h t , the current state of health (SOH) of the battery t , operating condition parameters (such as temperature, charge and discharge current), i.e., s t = [h t , SOH t , operating condition].

[0110] Action a t : charging / discharging strategy (such as constant current charging, pulsed discharging, discrete action space A = {a1, a2,......, a K}).

[0111] (2) Q-value function and network structure

[0112] Q-network input: State s t , and outputs the Q-values Q(s t , a; θ) for each action, where θ is the network parameter.

[0113] Loss function:

[0114] where r is the immediate reward (such as the reciprocal of the SOH prediction error); γ is the discount factor; θ′ is the target network parameter (periodically synchronized to the main network θ).

[0115] (3) Policy update (ε-greedy algorithm)

[0116]

[0117] where ε decays during training to balance exploration and exploitation.

[0118] S3: Overfitting optimization strategy

[0119] (1) Dropout layer in Conv-LSTM

[0120] Randomly discard neurons after the convolutional layer or LSTM layer, with the formula:

[0121] U′ t = Dropout(U t , p d );

[0122] h′ t = Dropout(h t , p d );

[0123] where p d is the dropout probability (value range: 0.2 - 0.5).

[0124] (2) Batch Normalization (BN)

[0125] Normalize the activation values after the convolutional or fully connected layer:

[0126]

[0127] where γ and β are learnable scaling and translation parameters; ∈ is a smoothing term (to prevent the denominator from being zero).

[0128] (3) Regularization of DQN

[0129] Add an L2 regularization term to the loss function:

[0130] L(θ) = original loss + λ ∑ θ θ 2 ;

[0131] where λ is the regularization coefficient (value: 10 -4 ).

[0132] S4: Model training process

[0133] Data preprocessing: Normalize the battery historical data (voltage, current, temperature, SOH, etc.) to [-1, 1] and construct the input sequence X.

[0134] Conv - LSTM forward propagation: Extract spatio - temporal features and output the battery state representation h T .

[0135] DQN policy generation: According to the state s t = [h t , SOH t ,......], select the action a t , execute the simulated working condition and obtain the reward r and the next state s′.

[0136] Experience replay: Transfer (s t , a t , r t , s t+1 ′) is stored in the experience pool, and a batch of data is randomly sampled to update the network.

[0137] Parameter update: Backpropagation is used to optimize the parameters of Conv-LSTM to minimize the SOH prediction loss (such as mean square error); the parameters of DQN are updated using gradient descent to minimize the regularization loss function.

[0138] Target network synchronization: Regularly copy the parameters of the main network to the target network to stabilize the training process.

[0139] Furthermore, the training of the hybrid model using historical data of electric vehicle batteries across different brands and models under various working conditions includes: collecting a large amount of historical data of electric vehicle batteries of different brands and models under various working conditions, including normal use data and fault data, to construct a training dataset;

[0140] The training dataset is used to perform supervised training on the hybrid model, with the actually measured SOH value as the label, to minimize the error between the model prediction value and the true value, such as the root mean square error (RMSE). The model parameters are continuously updated through the backpropagation algorithm, enabling the model to learn the complex relationship between battery health and various input features.

[0141] Furthermore, it also includes: To adapt to the aging and performance changes of the battery during use, the model has the ability to be updated online. When the on-vehicle battery management system receives new battery data, the incremental learning algorithm is used to gradually integrate the new data into the model training and dynamically adjust the model parameters, enabling the model to always maintain the ability to accurately estimate and predict the current health state of the battery.

[0142] In a specific embodiment, an on-vehicle battery health prediction system for electric vehicles, as Figure 2 shown, includes:

[0143] Acquisition and processing module: Used to acquire multi-source data and preprocess the acquired multi-source data;

[0144] Screening module: Used to extract features from the preprocessed data and perform screening;

[0145] Model construction module: Used to construct a hybrid model that combines deep learning and reinforcement learning based on the screened features;

[0146] Training module: Used to collect historical data of electric vehicle batteries across different brands and models under various working conditions to train the hybrid model to obtain an optimal health prediction model;

[0147] Prediction module: used to collect real-time data and input it into the optimal state of health prediction model to obtain the SOH estimation and prediction results.

[0148] Furthermore, it also includes a result output and application module:

[0149] Present the SOH estimation and prediction results output by the model to the user in an intuitive way, such as displaying the current battery health percentage on the vehicle dashboard, and at the same time providing an estimated time for the remaining service life of the battery. For the situation where the battery health is close to the warning threshold, the system issues an alarm in a timely manner to remind the user to perform battery maintenance or replacement.

[0150] Apply the SOH estimation and prediction results to the energy management system of the electric vehicle, dynamically adjust the vehicle's power output strategy according to the battery health status, optimize the charging plan, avoid further damage to the battery caused by overcharging or over-discharging, thereby extending the battery service life and improving the overall performance and safety of the electric vehicle.

[0151] In a specific embodiment, a method for predicting the state of health of an in-vehicle battery of an electric vehicle specifically includes the following steps:

[0152] Step 1: Data collection and preprocessing

[0153] (1) Install high-precision voltage and current sensors on the electric vehicle to collect the voltage and current data of the battery in real time, and set the sampling frequency to 1 Hz. At the same time, use the in-vehicle temperature sensor to obtain the battery temperature data with an accuracy of ±0.5 °C. Obtain information such as the number of charge and discharge cycles and the cumulative driving mileage through the vehicle's CAN bus.

[0154] (2) Install devices such as an acceleration sensor, a speed sensor, and a gyroscope to collect the acceleration and speed change data during the vehicle's driving process for analyzing the user's driving habits. Judge the hard acceleration and hard braking behaviors by analyzing the change rate of acceleration, and count the number of hard acceleration and hard braking times within a certain period. Obtain the usage duration of different driving modes (such as economy mode, sport mode) through the vehicle's on-board computer.

[0155] (3) Install environmental temperature and humidity sensors outside the vehicle to collect environmental data with a sampling frequency of once every 5 minutes.

[0156] (4) For the collected raw data, the adaptive Kalman filtering algorithm is first used to remove the noise interference in the voltage and current data. For the missing data values, if the continuous missing time does not exceed 10 seconds, the linear interpolation algorithm is used for filling; if the missing time exceeds 10 seconds, the statistical model based on historical data is used for estimation and filling. For the abnormal data, an anomaly detection model based on the Isolation Forest algorithm is constructed to identify the data deviating from the normal data distribution range as outliers, and they are corrected according to the change trend of adjacent normal data.

[0157] Step 2: Feature Engineering

[0158] (1) Extract multiple features from the preprocessed data:

[0159] Battery voltage change rate: Calculate the ratio of the difference in battery voltage at adjacent sampling times to the sampling time interval, which reflects the dynamic change of the battery voltage. Current impact feature: Define the current impact index according to the instantaneous change amplitude and duration of the current during rapid acceleration to measure the impact degree of the user's driving habit on the battery current. Battery power consumption characteristics under different driving modes: Calculate the average power output by the battery per unit time under different driving modes such as the economy mode and the sports mode. Correlation feature between ambient temperature and battery internal resistance: Fit the functional relationship between ambient temperature and battery internal resistance through experimental data, and extract the change characteristics of the battery internal resistance at different ambient temperatures.

[0160] (2) Use the principal component analysis (PCA) method to perform dimensionality reduction on all the extracted features, reducing the feature dimension to about 60% of the original to remove redundant information. Then, use the mutual information analysis method to calculate the mutual information value between each feature and the SOH, and select the key features with mutual information values greater than a certain threshold (such as 0.1) as the input for the subsequent model.

[0161] Step 3: Construction and Training of the Estimation and Prediction Model

[0162] Construct a hybrid model integrating deep learning and reinforcement learning. In the Conv-LSTM network part, set 3 convolutional layers with convolutional kernel sizes of 3×3, 5×5, and 7×7 respectively. After each convolutional layer, connect a ReLU activation function and a max-pooling layer to extract the local features of the battery data. Then connect 2 LSTM layers with the number of hidden units being 128 and 64 respectively to capture the time series features of the data. Introduce a Dropout layer after the fully connected layer of the Conv-LSTM network, and set the Dropout probability to 0.2 to prevent overfitting.

[0163] In the reinforcement learning part, a deep Q-network (DQN) is constructed. The input of the DQN is the current state of the battery (including voltage, current, temperature, SOH estimation value, etc.) and the current operation (such as charging and discharging strategies), and the output is the Q value corresponding to different operations. The DQN is set to have 2 hidden layers with the number of neurons being 256 and 128 respectively, and the ReLU activation function is adopted. The mean squared error (MSE) is used as the loss function of the DQN, and the Adam optimizer is used to update the parameters of the DQN, with the learning rate set to 0.001.

[0164] Step 4: Collect battery data of 500 electric vehicles of different brands and models, covering various years of use and mileage, and construct a training data set. The data set is divided into a training set, a validation set, and a test set according to the ratio of 70%, 20%, and 10%. The training set is used to perform supervised training on the hybrid model. During the training process, the output of the Conv-LSTM network is used as the state input of the DQN. By continuously adjusting the model parameters, the root mean square error (RMSE) between the predicted SOH value of the model and the true SOH value is minimized. During the training process, every 10 training epochs, the validation set is used to validate the model, and the hyperparameters of the model, such as the learning rate and Dropout probability, are adjusted according to the validation results to prevent the model from overfitting.

[0165] Step 5: Model update and application

[0166] The on-vehicle battery management system of the electric vehicle sends newly collected battery data, vehicle driving data, and environmental data to the cloud server every 1 hour. The cloud server adopts an incremental learning algorithm to gradually integrate the new data into the model training. Specifically, the new data is merged with the historical training data, the features are recalculated, and then the hybrid model is fine-tuned and trained, and only some parameters of the model are updated to reduce the computational amount and training time. The updated model parameters are then sent down to the on-vehicle battery management system to achieve online update of the model.

[0167] Step 6: The on-vehicle battery management system estimates the SOH of the battery in real time according to the updated model and displays the estimation result on the vehicle dashboard. When the SOH value is lower than 80%, the system issues a yellow warning; when the SOH value is lower than 60%, the system issues a red alarm to remind the user to perform battery maintenance or replacement in time. At the same time, the energy management system dynamically adjusts the vehicle's power output strategy according to the SOH estimation result. Among them, when the SOH is low, the maximum power output of the vehicle is limited to avoid over-discharging of the battery; during charging, according to the SOH and the current state of the battery, the charging current and voltage are optimized, and a suitable charging algorithm is adopted to extend the battery life.

[0168] Specifically, calculate the charging power limit coefficient: Set the charging power limit coefficient β according to the SOH to protect the battery.

[0169] When SOH ≥ 90%, β = 1, and the charging power is not restricted.

[0170] When 80% ≤ SOH < 90%, β = 0.7, and the charging power is reduced to 70% of the original.

[0171] When SOH < 80%, β = 0.5, and the charging power is reduced to 50% of the original.

[0172] Determine the target charging time: According to the user's needs and itinerary arrangements, determine the target charging time t target .

[0173] Calculate the actual allowable charging power: Let the maximum charging power of the battery be P max , the actual allowable charging power P allow The calculation formula is:

[0174] P allow = β * P max ;

[0175] Calculate the required charging duration: According to the current remaining battery power Q remaining and the actual allowable charging power P allow , calculate the required charging duration t charge , the formula is:

[0176] t charge =(Q rated -Q remaining ) / P allow ;

[0177] Adjust the charging plan: If t charge > t target , then it is necessary to start charging in advance or find a charging pile with a higher power; if t charge < t target Then the charging power can be appropriately reduced to extend the battery life. Through the above specific implementation methods, the battery health status can be accurately and real-time evaluated, providing a strong guarantee for the safe and efficient operation of electric vehicles.

[0178] In this specification, each embodiment is described in a progressive manner. The key point of each embodiment is to illustrate the differences from other embodiments. For the same and similar parts among the embodiments, reference can be made to each other. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple. For the relevant parts, reference can be made to the description in the method section.

[0179] The foregoing description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for predicting the health status of an on-vehicle battery of an electric vehicle, characterized in that, Including: Collecting multi-source data and preprocessing the collected multi-source data; Extracting features from the preprocessed data and performing screening; Constructing a hybrid model that combines deep learning and reinforcement learning based on the screened features; Collecting historical data of electric vehicle batteries across different brands and models under various working conditions to train the hybrid model to obtain an optimal state of health (SOH) prediction model; Collecting real-time data and inputting it into the optimal SOH prediction model to obtain SOH estimation and prediction results.

2. The method for predicting the health of an in-vehicle battery of an electric vehicle according to claim 1, wherein The multi-source data includes: battery voltage, current, temperature, charge and discharge times, cumulative driving mileage, acceleration, speed change, number of hard brakes, duration of different driving modes, as well as environmental temperature and humidity.

3. A method for predicting the health degree of an in-vehicle battery of an electric vehicle according to claim 1, characterized in that, The preprocessing of the collected multi-source data includes: using an adaptive filtering algorithm to remove noise interference, using an interpolation algorithm to fill in missing data values, and for abnormal data, identifying and correcting it through an anomaly detection model based on statistical analysis and machine learning.

4. A method for predicting the health degree of an in-vehicle battery of an electric vehicle according to claim 1, characterized in that, The extraction of features from the preprocessed data includes: extracting features such as voltage and current change rate from the preprocessed data, the correlation feature between hard acceleration frequency and battery current impact, the battery power consumption feature under different driving modes, and the influence feature of environmental temperature on battery internal resistance change.

5. A method for predicting the health of an in-vehicle battery of an electric vehicle according to claim 1, characterized in that, The screening includes: using a method that combines principal component analysis and mutual information analysis to perform dimensionality reduction and screening on the extracted features, removing redundant features, and retaining key features that have a significant impact on SOH.

6. A method for predicting the health of an in-vehicle battery of an electric vehicle according to claim 1, characterized in that, The construction of a hybrid model that combines deep learning and reinforcement learning based on the screened features includes: the deep learning part uses an improved convolutional long short-term memory network, automatically learning local features in the battery data through a convolutional neural network, and then capturing the time series features of the data through a long short-term memory network; the reinforcement learning part introduces a deep Q network, taking the current state and operation of the battery as input, and through continuous interaction with the environment, learning the optimal SOH estimation and prediction strategy and dynamically adjusting the model parameters.

7. A method for predicting the health of an in-vehicle battery of an electric vehicle according to claim 6, characterized in that, Also including: Introducing a Dropout layer and Batch Normalization technology into the improved convolutional long short-term memory network, and at the same time using a regularization method to optimize the loss function of the deep Q network to optimize the overfitting problem in model training.

8. A method for predicting the health of an in-vehicle battery of an electric vehicle according to claim 1, characterized in that, The collection of historical data of electric vehicle batteries across different brands and models under various working conditions to train the hybrid model includes: collecting historical data of electric vehicle batteries of different brands and models under various working conditions, including normal use data and fault data, to construct a training data set; Using the training data set to perform supervised training on the hybrid model, using the actually measured SOH value as a label, minimizing the error between the model prediction value and the true value, and continuously updating the model parameters through the backpropagation algorithm to enable the model to learn the relationship between battery health and various input features.

9. A method for predicting the health of an in-vehicle battery of an electric vehicle according to claim 8, characterized in that, Also including: When new battery data is received, using an incremental learning algorithm to gradually integrate the new data into model training and dynamically adjust the model parameters.

10. An on-vehicle battery health prediction system for an electric vehicle, characterized in that, Including: A collection and processing module: used to collect multi-source data and preprocess the collected multi-source data; Screening module: used to extract features from the preprocessed data and perform screening; Model construction module: used to construct a hybrid model that combines deep learning and reinforcement learning based on the screened features; Training module: used to collect historical data of electric vehicle batteries across different brands and models under various working conditions to train the hybrid model to obtain an optimal state of health prediction model; Prediction module: used to collect real-time data and input it into the optimal state of health prediction model to obtain the SOH estimation and prediction results.

Images