A method for predicting the battery life of new energy heavy-duty trucks based on convolutional neural networks

By using a convolutional neural network-based method, battery degradation characteristics are automatically learned and a lightweight network is constructed, which solves the problems of accuracy and real-time performance in predicting the battery life of new energy heavy trucks. This enables high-precision battery life prediction and real-time health management, supporting accurate decision-making for battery replacement and maintenance.

CN123115633APending Publication Date: 2026-11-17BEIBEN TRUCKS GRP
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611054080.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-15
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies cannot effectively utilize multi-source heterogeneous battery data and are difficult to automatically extract battery degradation characteristics, resulting in low accuracy and poor generalization in predicting the battery life of new energy heavy trucks. Furthermore, existing methods have high computational complexity, making it difficult to meet the real-time requirements of on-board edge devices.

Method used

By employing a convolutional neural network-based approach, through data acquisition, preprocessing, feature engineering, and model training, the system automatically learns the spatiotemporal characteristics of battery degradation, constructs a lightweight network structure, achieves high-precision prediction of remaining battery life, and uploads the prediction results in real time via edge deployment.

Benefits of technology

It improves the accuracy and generalization ability of predicting the battery life of new energy heavy trucks, supports the real-time performance and interpretability of the battery health management system, promotes the transformation from regular maintenance to predictive maintenance, and is in line with the development trend of vehicle networking.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115633A_ABST
    Figure CN123115633A_ABST
Patent Text Reader

Abstract

This invention relates to a method for predicting the battery life of new energy heavy-duty trucks based on convolutional neural networks, solving the problem of reliance on manual feature engineering in traditional methods. The method includes the following steps: Step 1: Data acquisition; Step 2: Data preprocessing, including: 1) Data filtering; 2) Missing value handling; 3) Noise filtering; 4) Standardization; Step 3: Feature engineering and input construction, including: 1) Temporal alignment; 2) 3D tensor construction; Step 4: Feature engineering and input construction, including: 1) Model input; 2) Core network structure; 3) Output layer; Step 5: Model training and deployment, including: 1) Loss function; 2) Edge deployment. This invention provides interpretable features for battery health management (BMS), supports accurate decision-making for battery replacement and maintenance, promotes the transformation of new energy commercial vehicles from "periodic maintenance" to "predictive maintenance," and aligns with the development trend of vehicle-to-everything (V2X) communication.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to a method for predicting the battery life of new energy heavy-duty trucks based on convolutional neural networks, belonging to the field of new energy heavy-duty trucks. Background Technology

[0002] The lifespan of batteries in new energy heavy-duty trucks is currently affected by multiple factors such as charge-discharge cycles, temperature, and load. Traditional prediction methods (such as empirical models and equivalent circuit models) rely on manual feature extraction, resulting in low accuracy and poor generalization. Existing machine learning methods (such as support vector machines and random forests) are insufficient in capturing features from time-series data and struggle to handle the nonlinear relationships in the battery degradation process. RNN / LSTM-based prediction methods suffer from difficulties in training long sequences and high computational complexity, and do not fully utilize multi-source heterogeneous battery data (voltage, current, temperature, and other sensor data). Current technologies lack a lightweight solution that can automatically extract multi-dimensional features of battery degradation, fuse operating condition data, and achieve high-precision end-to-end lifespan prediction. Summary of the Invention

[0003] This invention aims to address the reliance of traditional methods on manual feature engineering. By using CNN to automatically learn the spatiotemporal features of battery degradation, it improves prediction accuracy, enables early warning of the remaining service life (RUL) of batteries in new energy heavy trucks, and designs a lightweight network structure to meet the real-time requirements of on-board edge devices.

[0004] This invention is achieved through the following technical solutions: A method for predicting the battery life of new energy heavy-duty trucks based on convolutional neural networks includes the following steps: Step 1: Data Acquisition. Data on the vehicle's power battery is collected using the vehicle's built-in sensors, including: 1) Battery parameters: charge / discharge voltage, charge / discharge current, temperature, internal resistance, state of charge (SOC), and state of health (SOH); 2) Vehicle operating parameters: load weight, driving gradient, acceleration, braking frequency, ambient temperature and humidity; 3) Historical attenuation records: cumulative number of cycles, historical fault codes, and maintenance records; Step 2: Data preprocessing, including: 1) Data filtering: Extract data for the charge / discharge range of SOC from 30% to 80%; 2) Missing value handling: Piecewise linear interpolation is used to fill in missing data caused by sensor packet loss. Data points that are missing for more than 5 consecutive time steps are marked as abnormal and removed. 3) Noise filtering: Adaptive Kalman filtering, i.e., dynamic adjustment of process noise covariance, is used to prioritize the processing of high-frequency noise from current and voltage; 4) Standardization: Perform Z-score standardization on the denoised data according to the feature dimension to eliminate the difference in results caused by large differences in the values ​​of different features. The formula is: Z= Where x: the original feature value, which is a specific sensor data after preprocessing, i.e., after denoising and interpolation; μ: The mean of this feature dimension refers to the average value of all values ​​of the same sensor channel in all training data, representing the "central reference point" of this feature; σ: Standard deviation of this feature dimension, which measures the degree of dispersion of the feature data around the mean, i.e., the magnitude of fluctuation. Z: The standardized feature value. After calculation, the overall distribution of this feature becomes a standard normal distribution with a mean of 0 and a standard deviation of 1. Step 3: Feature engineering and input construction, including: 1) Timing alignment: Sensor data with different sampling frequencies (e.g., voltage 1Hz, temperature 0.2Hz) are resampled and unified to 10 seconds / frame through a time window; 2) Construction of three-dimensional tensors: Dimension 1: Time step: A sliding window captures 60 consecutive minutes of data (360 time steps). Dimension 2 sensor channels: including 8 feature channels such as voltage, current, temperature, load weight, driving slope, acceleration, braking frequency, and ambient temperature and humidity; Dimension 3: Cycle count: Stacked according to the number of battery charge and discharge cycles, forming a three-dimensional matrix of time sequence-channel-cycle; Step 4: Feature engineering and input construction, including: 1) Model input: preprocessed 3D tensor and vehicle operating parameters (encoded as vectors through an embedding layer); 2) Core network structure: Multi-branch convolutional layers: Branch 1: 1D convolution, kernel_size=3, stride=1, extracts local temporal features; Branch 2: Dilated convolution, dilation_rate=2, to capture long-period dependencies; Feature fusion layer: concatenates the outputs of multiple branches along the channel dimension and reduces dimensionality through 1×1 convolution; CBAM module: Apply attention weights to sensor channels and key time steps; 3) Output layer: SOH prediction: The Sigmoid activation function outputs a health status between 0 and 1; RUL prediction: Linear layer outputs the remaining number of iterations; Step 5: Model training and deployment, including: 1) Loss function: Where α=0.7, β=0.3; λ: L2 regularization coefficient, which is a hyperparameter, initially set to 0.0001 and adjusted according to the training situation to control the strength of the regularization term; L2: L2 regularization, also known as weight decay, is the sum of squares of all trainable parameters of the model. 2) Edge deployment: Data is acquired in real time via CAN bus, and the prediction results are uploaded to the cloud monitoring platform.

[0005] This invention provides interpretable features for battery health management (BMS) (visualizing key sensor channels through Grad-CAM), supports accurate decision-making for battery replacement and maintenance, promotes the transformation of new energy commercial vehicles from "routine maintenance" to "predictive maintenance", and is in line with the development trend of vehicle-to-everything (V2X) communication. Attached Figure Description

[0006] Figure 1 It is a CBAM structure. Detailed Implementation

[0007] This invention is a method for predicting the battery life of new energy heavy-duty trucks based on convolutional neural networks. The specific implementation steps are as follows: (1) Data Acquisition: Data on the vehicle's power battery is collected through the vehicle's built-in sensors, including: ① Battery parameters: charge / discharge voltage, charge / discharge current, temperature, internal resistance, SOC (state of charge), and SOH (state of health).

[0008] ② Vehicle operating parameters: load weight, driving gradient, acceleration, braking frequency, ambient temperature and humidity.

[0009] ③ Historical decay records: cumulative number of cycles, historical fault codes, and maintenance records.

[0010] (2) Data preprocessing: ① Data filtering: Extract data for the charge / discharge range of SOC between 30% and 80%.

[0011] ② Missing value handling: Piecewise linear interpolation is used to fill in missing data caused by sensor packet loss. Data points that are missing for more than 5 consecutive time steps are marked as abnormal and removed.

[0012] ③ Noise filtering: Adaptive Kalman filtering (dynamic adjustment of process noise covariance) is used to prioritize the processing of high-frequency noise from current and voltage.

[0013] ④ Standardization: Perform Z-score standardization on the denoised data according to the feature dimension to eliminate the difference in results caused by large differences in the values ​​of different features. The formula is: = Where μ is the mean and σ is the standard deviation.

[0014] (3) Feature engineering and input construction: ① Timing alignment: Sensor data with different sampling frequencies (such as voltage 1Hz, temperature 0.2Hz) are resampled through a time window to be unified to 10 seconds / frame.

[0015] ② Construction of three-dimensional tensors: Dimension 1 (Time Step): A sliding window captures 60 consecutive minutes of data (360 time steps). Dimension 2 (Sensor Channels): 8 feature channels including voltage, current, and temperature; Dimension 3 (Number of Cycles): Stacked according to the number of battery charge and discharge cycles, forming a three-dimensional matrix of time sequence-channel-cycle.

[0016] (4) Convolutional Neural Network Prediction Model: ① Model input: preprocessed 3D tensor and vehicle operating parameters (encoded as vectors through an embedding layer); ② Core network structure: Multi-branch convolutional layers: Branch 1: 1D convolution (kernel_size=3, stride=1) extracts local temporal features; Branch 2: Dilated convolution (dilation_rate=2) captures long-periodic dependencies; Feature fusion layer: concatenates the outputs of multiple branches along the channel dimension and reduces dimensionality through 1×1 convolution; CBAM module: Apply attention weights to sensor channels and key time steps (e.g., ... Figure 1 (As shown).

[0017] ③ Output layer: SOH prediction: The Sigmoid activation function outputs a health status between 0 and 1; RUL prediction: Linear layer outputs the remaining number of iterations.

[0018] ④ Edge deployment: Data is acquired in real time via the CAN bus, and the prediction results are uploaded to the cloud monitoring platform.

[0019] (5) Model training and deployment: ① Loss function: Where α=0.7, β=0.3.

[0020] ② Edge deployment: Data is acquired in real time via the CAN bus, and the prediction results are uploaded to the cloud monitoring platform.

[0021] This invention jointly models vehicle operating condition data and battery body data, and realizes cross-modal feature interaction through convolution kernel sharing weights; after deployment, the model can be continuously optimized through federated learning, protecting data privacy while adapting to different battery degradation modes of different fleets.

Claims

1. A new energy heavy truck battery life prediction method based on a convolutional neural network, characterized by: Includes the following steps: Step 1: Data Acquisition. Data on the vehicle's power battery is collected using the vehicle's built-in sensors, including: 1) Battery parameters: charge / discharge voltage, charge / discharge current, temperature, internal resistance, state of charge (SOC), and state of health (SOH); 2) Vehicle operating parameters: load weight, driving gradient, acceleration, braking frequency, ambient temperature and humidity; 3) Historical attenuation records: cumulative number of cycles, historical fault codes, and maintenance records; Step 2: Data preprocessing, including: 1) Data filtering: Extract data for the charge / discharge range of SOC from 30% to 80%; 2) Missing value handling: Piecewise linear interpolation is used to fill in missing data caused by sensor packet loss. Data points that are missing for more than 5 consecutive time steps are marked as abnormal and removed. 3) Noise filtering: Adaptive Kalman filtering, i.e., dynamic adjustment of process noise covariance, is used to prioritize the processing of high-frequency noise from current and voltage; 4) Standardization: Perform Z-score standardization on the denoised data according to the feature dimension to eliminate the difference in results caused by large differences in the values ​​of different features. The formula is: Z= Where x: the original feature value, which is a specific sensor data after preprocessing, i.e., after denoising and interpolation; μ: The mean of this feature dimension refers to the average value of all values ​​of the same sensor channel in all training data, representing the "central reference point" of this feature; σ: Standard deviation of this feature dimension, which measures the degree of dispersion of the feature data around the mean, i.e., the magnitude of fluctuation. Z: The standardized feature value. After calculation, the overall distribution of this feature becomes a standard normal distribution with a mean of 0 and a standard deviation of 1. Step 3: Feature engineering and input construction, including: 1) Timing alignment: Sensor data with different sampling frequencies are resampled and unified to 10 seconds / frame through a time window; 2) Construction of three-dimensional tensors: Dimension 1: Time step: A sliding window captures continuous data for m minutes. Dimension 2 sensor channels: including 8 feature channels such as voltage, current, temperature, load weight, driving slope, acceleration, braking frequency, and ambient temperature and humidity; Dimension 3: Cycle count: Stacked according to the number of battery charge and discharge cycles, forming a three-dimensional matrix of time sequence-channel-cycle; Step 4: Feature engineering and input construction, including: 1) Model input: preprocessed 3D tensors and vehicle operating parameters; 2) Core network structure: Multi-branch convolutional layers: Branch 1: 1D convolution, kernel_size=3, stride=1, extracts local temporal features; Branch 2: Dilated convolution, dilation_rate=2, to capture long-period dependencies; Feature fusion layer: concatenates the outputs of multiple branches along the channel dimension and reduces dimensionality through 1×1 convolution; CBAM module: Apply attention weights to sensor channels and key time steps; 3) Output layer: SOH prediction: The Sigmoid activation function outputs a health status between 0 and 1; RUL prediction: Linear layer outputs the remaining number of iterations; Step 5: Model training and deployment, including: 1) Loss function: Where α=0.7, β=0.3; λ: L2 regularization coefficient, which is a hyperparameter, initially set to 0.0001 and adjusted according to the training situation to control the strength of the regularization term; L2: L2 regularization, also known as weight decay, is the sum of squares of all trainable parameters of the model. 2) Edge deployment: Data is acquired in real time via CAN bus, and the prediction results are uploaded to the cloud monitoring platform.

2. The new energy heavy truck battery life prediction method based on a convolutional neural network according to claim 1, characterized in that: In step 3, 1) Timing alignment, sensor data with different sampling frequencies are resampled through a time window to be unified to 10 seconds / frame.

3. The new energy heavy truck battery life prediction method based on a convolutional neural network according to claim 2, characterized in that: Step 3, 2) Dimension 1 Time Step: Use a sliding window to capture 60 consecutive minutes of data, i.e., 360 time steps.