A truck battery health assessment method based on multi-source data

Abstract

The application discloses a truck battery health evaluation method based on multi-source data, comprising the following steps: S1, collecting multi-source data and preprocessing, and constructing a multi-source time series data set; S2, converting and weighting the information of the multi-source data set to form a working condition stress feature sequence; S3, calculating a health-related feature vector; S4, modeling the long-term trend and short-term fluctuation of battery capacity degradation by using a time series self-attention mechanism, and outputting a battery health state result; S5, inputting the battery health state and the working condition stress feature sequence into an improved Autoformer model, extracting long-term trend and short-term fluctuation components, predicting battery capacity and determining a risk level; and S6, integrating the battery capacity prediction value, the risk level and the early warning result, generating a battery health evaluation result package and displaying it in real time. Through multi-source data fusion and advanced model prediction, the application significantly improves the accuracy and practicality of battery health evaluation.

Description

Technical Field

[0001] This invention relates to the field of multi-source data processing and analysis and time series forecasting, and in particular to a method for assessing the health of truck batteries based on multi-source data. Background Technology

[0002] In recent years, battery health assessment methods based on multi-source data have gradually become an important tool for electric truck battery management, especially in the accurate monitoring and prediction of battery degradation. However, existing technologies still face some key challenges and shortcomings, limiting their widespread adoption and efficiency in practical applications.

[0003] First, most existing battery health assessment methods rely on single-modal data, making comprehensive assessment and health monitoring of battery status difficult. Especially under dynamic operating conditions, battery health evolves rapidly and complexly, and traditional methods fail to fully utilize multi-source data, limiting the accuracy and comprehensiveness of health assessment models. Second, existing battery health prediction methods largely depend on static models, lacking adaptive adjustment mechanisms. For example, most methods fail to adequately consider instantaneous changes and long-term trends during battery degradation, ignoring the dynamic impact of factors such as battery internal resistance, capacity degradation, and temperature stability, resulting in poor adaptability of assessment results to different operating environments and usage conditions.

[0004] Finally, existing battery health assessment methods often lack the embedding of physical knowledge, leading to model predictions that may not conform to actual physical laws. For example, the relationship between battery internal resistance and capacity degradation, as well as temperature fluctuations, is not accurately constrained by physical models, which may reduce the credibility and accuracy of assessment results. For complex battery systems, models lacking physical constraints struggle to effectively capture the battery's true performance and degradation characteristics.

[0005] Therefore, how to provide a truck battery health assessment method based on multi-source data is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a truck battery health assessment method based on multi-source data. This invention constructs an intelligent battery health assessment process with multi-dimensional feature fusion, dynamic prediction, and adaptive adjustment. By introducing an improved Autoformer model, this invention can accurately extract the long-term trend and short-term fluctuation characteristics of battery capacity degradation, extrapolate and predict the battery's capacity evolution trajectory in future periods, and output the remaining usable lifespan. The improved Autoformer model employs a time-series data decomposition mechanism, maintaining efficient modeling while incorporating the physical characteristics of the battery, avoiding the problem of insufficient understanding of battery dynamic behavior in traditional methods. Furthermore, this invention considers multi-dimensional features such as changes in battery internal resistance and temperature fluctuations during the health assessment process, and improves the accuracy and stability of the model under different operating conditions through an adaptive adjustment mechanism. This method enables real-time monitoring of battery health status and provides early warnings based on prediction results, helping to extend battery lifespan and improve battery management efficiency.

[0007] A truck battery health assessment method based on multi-source data according to an embodiment of the present invention includes the following steps: S1. Collect multi-source data from truck batteries, perform preprocessing, and construct a multi-source time-series dataset; S2. Convert the temperature change amplitude, discharge rate, vehicle load and road slope in the multi-source time series dataset into working condition stress factors with unified dimensions, and weighted accumulate to form a working condition stress feature sequence. S3. Based on multi-source time-series datasets, calculate dynamic voltage change rate, estimated equivalent internal resistance, individual cell voltage difference, individual cell temperature difference and discharge curve slope to form a health-related feature vector. S4. Using a time-series self-attention mechanism, the health-related feature vectors are modeled to depict the long-term trend and short-term fluctuation characteristics of battery capacity degradation, resulting in capacity decay values, internal resistance change rate, and temperature stability indicators, and outputting the battery health status results. S5. Input the battery health status results and the working condition stress characteristic sequence into the improved Autoformer model, extract the long-term trend component and the short-term fluctuation residual component representing the battery capacity degradation respectively, output the battery capacity prediction value, and determine the risk event, output the battery risk level and warning result. S6. Integrate the predicted battery capacity, battery risk level, and early warning results to generate a battery health assessment result package and output it for real-time display on the user interface.

[0008] Optionally, S1 specifically includes: S11. Read the raw data output of the multi-source sensors deployed on the truck to obtain multi-source data of the truck's power battery during operation; S12. Match the collected multi-source data according to the timestamp to construct a unified time axis, and perform linear interpolation on the data with different sampling frequencies; S13. Using a sliding window detection strategy and threshold determination method, identify and remove mutation points or multi-source data that exceed a preset reasonable threshold in multi-source data, perform outlier removal and data filtering, perform Kalman filter interpolation on data segments with missing values, and continuously fill in the values ​​at the breakpoints. S14. Arrange the multi-source data after time synchronization, anomaly removal and interpolation repair in chronological order to construct a multi-source time series dataset.

[0009] Optionally, S2 specifically includes: S21. Read the data sequences corresponding to the temperature change amplitude, discharge rate, vehicle load and road slope in the multi-source time series dataset, and extract the temperature data segment, discharge rate data segment, load data segment and slope data segment corresponding to the battery operating status. S22. Dimensionless normalization is performed on the temperature change range, discharge rate, vehicle load and road slope respectively, and the numerical range of each data is linearly mapped to a unified numerical range in proportion. S23. Preset corresponding weight parameters according to the degree of influence of temperature change on battery performance, the degree of influence of discharge rate on battery load pressure, the degree of influence of vehicle load on battery mechanical stress, and the degree of influence of road slope on battery energy output demand. S24. The normalized temperature change amplitude, discharge rate, vehicle load and road slope are weighted and summed according to the weight parameters to obtain the working condition stress factor value for a single time step. The working condition stress factors of consecutive time steps are accumulated and summed in chronological order to form a working condition stress characteristic sequence.

[0010] Optionally, S3 specifically includes: S31. Read the voltage data sequence from the multi-source time series dataset, perform differential operation on the voltage value changes between consecutive time points, calculate the voltage change value per unit time, and arrange all the change values ​​in time order to generate a dynamic voltage change rate sequence. S32. By inputting the battery's voltage, current data, and open-circuit voltage, the parameters are fitted using a pre-trained RC equivalent circuit model. The relationship between the battery's equivalent internal resistance and current-voltage characteristics is fitted using the least squares method to obtain the internal resistance value at each time step and generate the battery internal resistance time sequence. S33. Calculate the difference between the temperature and voltage values ​​of all individual cells at the same time point, obtain the temperature difference and pressure difference between adjacent cells, and form a cell temperature difference sequence and a cell pressure difference sequence according to the time series. S34. Extract the voltage and state of charge (SOC) data of the SOC drop interval that exceeds the continuous time threshold during the discharge phase. Select the corresponding voltage and SOC values ​​at set intervals within each SOC drop interval. Calculate the slope of voltage change with respect to SOC using the least squares linear fitting method. Use the slope values ​​of each segment as the voltage change rate index during the discharge process and form a discharge curve slope sequence in time order. S35. The dynamic voltage change rate sequence, the equivalent internal resistance estimate sequence, the single cell voltage difference sequence, the single cell temperature difference sequence, and the discharge curve slope value are standardized and spliced ​​together to form a health-related feature vector characterizing the battery's operating status.

[0011] Optionally, S4 specifically includes: S41. Arrange the vectors of all time steps of the health-related feature vectors into an input matrix in order, and apply position encoding to each row of the input matrix to write the time sequence information into the input matrix to generate a feature matrix with position encoding. S42. Input the feature matrix into the multi-head self-attention layer, calculate the dot product correlation coefficient of any two row vectors in the input matrix, and perform Softmax normalization. Multiply the normalized correlation coefficients with the corresponding row vectors element by element and sum them to obtain multiple attention output matrices. Then, concatenate them column by column and input them into two fully connected layers to perform element-wise linear calculation and ReLU nonlinear calculation to generate the temporal feature matrix. S43. Input the time-series feature matrix into the output mapping layer, perform preset linear matrix multiplication and vector addition on each row, and output three results: capacity decay value, internal resistance change rate and temperature stability index. Arrange the three sets of values ​​in time order to obtain the battery health status result sequence. S44. Subtract the capacity decay value from the corresponding actual capacity decay value, square the difference and sum them up to obtain the capacity loss value. Subtract the internal resistance change rate from the corresponding actual internal resistance change rate, square the difference and sum them up to obtain the internal resistance loss value. Subtract the temperature stability index from the corresponding actual temperature stability index, square the difference and sum them up to obtain the temperature loss value. S45. Substitute the output internal resistance, current and initial voltage into the equivalent circuit formula in the pre-trained RC equivalent circuit model, calculate the difference between the theoretical voltage value and the predicted voltage value step by step, square the difference and sum them up to obtain the circuit constraint loss value, calculate the derivative of adjacent differences of capacity output at each time step, sum the derivative values ​​at the positions where the derivative is greater than zero to obtain the capacity monotonic constraint loss value, compare the temperature stability index of each time step with the preset maximum allowable temperature difference, sum the differences exceeding the preset value to obtain the temperature difference constraint loss value. S46. Multiply the capacity loss value, internal resistance loss value, temperature loss value, circuit constraint loss value, capacity monotonic constraint loss value, and temperature difference constraint loss value by their respective preset weights and add them together to obtain the total loss value. Backpropagate the total loss value to calculate the gradient of each layer's parameters. Multiply the gradient by the learning rate as the parameter update amount. Perform subtraction update operations on all weight parameters and bias parameters one by one to obtain the updated parameters.

[0012] Optionally, S5 specifically includes: S51. The battery health status results and the working condition stress characteristic sequence are combined into a joint input sequence and fed into the improved Autoformer model. The trend component representing the long-term decline trend of battery capacity and the residual component representing short-term fluctuations are extracted to extend and predict the capacity evolution trajectory of the battery in the future period and output the predicted value of battery capacity. S52. Extract the trend component of the long-term battery capacity degradation trend output by the Autoformer model, calculate the average change slope within a set time window, and determine the risk of rapid capacity degradation when the trend slope is greater than the battery capacity degradation threshold. S53. Perform cumulative summation on the changes in the internal resistance of individual cells within the corresponding time window to obtain the cumulative increase in internal resistance. When the increase exceeds the internal resistance degradation threshold, it is determined to be a risk of internal resistance degradation. S54. Calculate the temperature difference of each unit at each time step and count the fluctuation range. When the temperature difference fluctuation range exceeds the temperature difference threshold, it is judged as a risk of temperature control imbalance. S55. Divide the above three risks into low risk, medium risk and high risk according to the number of judgments, and output the corresponding battery risk level and the warning result generated by combining the trigger risk.

[0013] Optionally, the improved Autoformer model specifically includes: After aligning the battery health status results with the working condition stress feature sequence by timestamp, they are spliced ​​together to form a joint input sequence. The joint input sequence is then standardized by using Z-Score standardization to calculate the mean and standard deviation of each dimension. Each value in the input sequence is then normalized to zero mean and unit variance according to the corresponding dimension to generate a standardized input matrix. An improved relative position encoding method is adopted, which maps the time step index value to a position vector through sine and cosine functions and adds it element by element to the standardized input matrix to form an input sequence matrix with position information and input it into the improved Autoformer model. A trend-residual dual-path modeling mechanism and a dynamic sliding window decomposition mechanism are introduced to capture the long-term degradation trend and short-term fluctuation anomaly in battery capacity change, respectively. A dynamic sliding window decomposition is performed by setting a time window length. The stationarity of the subsequence formed by each time window length and time step is detected and the spectrum is analyzed. Subsequences whose mean slope and frequency energy density reach the preset slope threshold and density threshold within the window are extracted as trend components, otherwise they are residual components. The operation is repeated to complete the trend-residual decomposition of the entire input sequence matrix and output the trend component sequence and residual component sequence. The trend component sequence is input into the main time series prediction path to obtain the time series input matrix, where each row represents the trend feature of a time step. The matrix is ​​standardized so that the mean of each column is 0 and the standard deviation is 1. The matrix is ​​then input into a multi-layer decoupled self-attention mechanism. The dot product similarity between time steps is calculated in each layer. The weighting coefficient is obtained by Softmax normalization and multiplied by the corresponding trend feature to obtain the weighted feature representation. The weighted feature representations at each time step are subjected to stacked convolutions with pre-defined kernel sizes and strides to extract local features and pass them to the next layer. The output of each layer is compared with the true target value, the mean squared error residual is calculated, the gradient is calculated, and the weights of each layer are updated through backpropagation. Through the self-attention mechanism and convolution operation of all layers, the dot product similarity between time steps is calculated at each layer and normalized by Softmax to obtain the weighting coefficients. The pre-weighted summation of local features is used to output the battery capacity degradation trend prediction result for future periods. The residual component sequences are aligned by time steps to form an input matrix, representing the short-term fluctuation characteristics of each time step. The extracted internal resistance change rate and temperature stability index are introduced and weighted and summed with preset weights to obtain the perturbation factor of each time step. The residual component sequences of each time step are weighted and superimposed with the corresponding perturbation factor to generate the weighted residual features. The weighted residual features are input into a feedforward neural network with residual connections. The short-term disturbances in the residuals are compensated by the ReLU activation function and the weighted summation operation with preset weights. The disturbance prediction results of each layer are accumulated layer by layer to obtain the disturbance correction result of the battery under the current operating conditions. The disturbance correction results of all layers are added to the trend prediction result to generate the corrected battery capacity prediction value. The capacity degradation trend prediction result and the residual disturbance prediction result are summed element by element over time step to form the capacity evolution prediction trajectory of the battery in the future period. Based on this prediction trajectory curve, the time required to decrease to the preset capacity threshold is calculated, and the remaining usable life prediction result of the battery at the current moment is output.

[0014] Optionally, S6 specifically includes: aligning the battery capacity prediction value, battery risk level and warning result according to the timestamp, splicing them into a structured data packet, encoding it according to a preset format, generating a battery health assessment result packet, and outputting the battery health assessment result packet through the data transmission interface for real-time display on the user interface.

[0015] The beneficial effects of this invention are: This invention, by combining an improved Autoformer model with multi-source data analysis technology, significantly enhances the accuracy and predictive capability of truck battery health assessment. During battery health assessment, the improved Autoformer model, through its unique time-series data decomposition mechanism, accurately extracts the long-term trend and short-term fluctuation characteristics of battery capacity degradation, achieving comprehensive perception and dynamic prediction of battery health status. It not only captures the long-term trend of battery capacity degradation but also effectively identifies instantaneous fluctuations and abnormal changes, improving the prediction accuracy of remaining battery life. By introducing physical constraints, such as internal resistance and temperature stability, this invention avoids the problem of neglecting physical laws in traditional methods, ensuring that the prediction results, while meeting the accuracy requirements of data-driven modeling, always conform to the actual physical characteristics of the battery.

[0016] Furthermore, this invention innovatively employs a multi-source data fusion strategy, combining multi-dimensional information such as battery voltage, current, internal resistance, and temperature with operating stress characteristics to provide a comprehensive health assessment framework. This method enhances the model's adaptability to battery behavior under complex operating conditions by integrating multi-dimensional features such as battery health status, internal resistance changes, and temperature fluctuations into a unified health-related feature vector.

[0017] In summary, this invention, by integrating an improved Autoformer model with multi-source data fusion technology, enables accurate monitoring and dynamic optimization prediction of battery health status, significantly improving the accuracy and stability of battery assessment. It also solves the problems of inaccurate battery health prediction and poor adaptability in existing methods, providing an intelligent, efficient, and reliable solution for electric truck battery management. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a truck battery health assessment method based on multi-source data proposed in this invention; Figure 2 This is a flowchart of the truck battery health status result generation method based on multi-source data proposed in this invention; Figure 3 This is a flowchart of the battery capacity degradation prediction and remaining life assessment based on the improved Autoformer model proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figures 1-3 A method for assessing truck battery health based on multi-source data includes the following steps: S1. Collect multi-source data on truck battery voltage, current, cell temperature, state of charge, charge and discharge capacity, vehicle load, road slope and ambient temperature and humidity. Perform preprocessing on the multi-source data, including time synchronization, outlier removal and interpolation repair, and construct a cleaned multi-source time-series dataset. S2. Convert the temperature change amplitude, discharge rate, vehicle load and road slope in the multi-source time series dataset into working condition stress factors with unified dimensions, and weighted accumulate to form a working condition stress feature sequence. S3. Based on multi-source time-series datasets, calculate dynamic voltage change rate, estimated equivalent internal resistance, individual cell voltage difference, individual cell temperature difference and discharge curve slope to form a health-related feature vector. S4. Input the health-related feature vectors into the time-series self-attention mechanism, and obtain the capacity decay value, internal resistance change rate and temperature stability index by modeling the long-term trend and short-term fluctuation characteristics of battery capacity degradation, and output the battery health status result. S5. Input the battery health status results and the working condition stress characteristic sequence into the improved Autoformer model, extract the long-term trend component and the short-term fluctuation residual component representing the battery capacity degradation respectively, output the battery capacity prediction value, and determine the risk event, output the battery risk level and warning result. S6. Integrate the predicted battery capacity, battery risk level, and early warning results to generate a battery health assessment result package and output it for real-time display on the user interface.

[0021] This implementation significantly improves the accuracy and response speed of truck battery health assessment through multi-source data fusion and an improved Autoformer model. First, multi-source time-series data is collected and preprocessed to construct the dataset required for battery health assessment. Then, a time-series self-attention mechanism is used to model the long-term trend and short-term fluctuations of battery capacity degradation, accurately predicting the battery health status. The Autoformer model is used to extract the trend and residual components of capacity degradation, further improving the accuracy of battery capacity prediction and risk assessment capabilities. By real-time assessment of battery health status and operating stress characteristics, this method achieves accurate determination of battery capacity prediction, internal resistance changes, and temperature control imbalances, providing early warnings of potential risks. Finally, the predicted battery capacity value and risk assessment results are integrated, and the health assessment results are output in real time and displayed via a user interface. Through dynamic optimization and multi-dimensional assessment, this implementation achieves intelligent battery health monitoring and risk warning under complex operating conditions, effectively improving the intelligence level and operational efficiency of battery management.

[0022] In this embodiment, S1 specifically includes: S11. Read the raw data output of the multi-source sensors deployed on the truck to obtain multi-source data of the truck's power battery during operation; S12. Match the voltage, current, cell temperature, state of charge, charge and discharge capacity, vehicle load, road slope and ambient temperature in the collected multi-source data according to the timestamp to construct a unified time axis, and perform linear interpolation processing on the data with different sampling frequencies. S13. Using a sliding window detection strategy and threshold determination method, identify and remove mutation points or multi-source data that exceed a preset reasonable threshold in multi-source data, perform outlier removal and data filtering, perform Kalman filter interpolation on data segments with missing values, and continuously fill in the values ​​at the breakpoints. S14. Arrange the multi-source data of voltage, current, unit temperature, state of charge, charge and discharge capacity, vehicle load, road slope and ambient temperature in chronological order after time synchronization, anomaly removal and interpolation repair, and construct a unified and continuous multi-source time series dataset.

[0023] In this embodiment, S2 specifically includes: S21. Read the data sequences corresponding to the temperature change amplitude, discharge rate, vehicle load and road slope in the multi-source time series dataset, and extract the temperature data segment, discharge rate data segment, load data segment and slope data segment corresponding to the battery operating status. S22. Dimensionless normalization is performed on the temperature change range, discharge rate, vehicle load and road slope respectively, and the numerical range of each data is linearly mapped to a unified numerical range in proportion. S23. Preset corresponding weight parameters according to the degree of influence of temperature change on battery performance, the degree of influence of discharge rate on battery load pressure, the degree of influence of vehicle load on battery mechanical stress, and the degree of influence of road slope on battery energy output demand. S24. The normalized temperature change amplitude, discharge rate, vehicle load and road slope are weighted and summed according to the weight parameters to obtain the working condition stress factor value for a single time step. The working condition stress factors of consecutive time steps are accumulated and summed in chronological order to form a working condition stress characteristic sequence.

[0024] In this embodiment, S3 specifically includes: S31. Read the voltage data sequence from the multi-source time series dataset, perform differential operation on the voltage value changes between consecutive time points, calculate the voltage change value per unit time, and arrange all the change values ​​in time order to generate a dynamic voltage change rate sequence. S32. By inputting the battery's voltage, current data, and open-circuit voltage, the parameters are fitted using a pre-trained RC equivalent circuit model. The relationship between the battery's equivalent internal resistance and current-voltage characteristics is fitted using the least squares method to obtain the internal resistance value at each time step and generate the battery internal resistance time sequence. S33. Calculate the difference between the temperature and voltage values ​​of all individual cells at the same time point, obtain the temperature difference and pressure difference between adjacent cells, and form a cell temperature difference sequence and a cell pressure difference sequence according to the time series. S34. Extract the voltage and state of charge (SOC) data of the SOC drop interval that exceeds the continuous time threshold during the discharge phase. Select the corresponding voltage and SOC values ​​at set intervals within each SOC drop interval. Calculate the slope of voltage change with respect to SOC using the least squares linear fitting method. Use the slope values ​​of each segment as the voltage change rate index during the discharge process and form a discharge curve slope sequence in time order. S35. The dynamic voltage change rate sequence, the equivalent internal resistance estimate sequence, the single cell voltage difference sequence, the single cell temperature difference sequence, and the discharge curve slope value are standardized and spliced ​​together to form a health-related feature vector characterizing the battery's operating status.

[0025] In this embodiment, S4 specifically includes: S41. Arrange the vectors of all time steps of the health-related feature vectors into an input matrix in order, and apply position encoding to each row of the input matrix to write the time sequence information into the input matrix to generate a feature matrix with position encoding. S42. Input the feature matrix into the multi-head self-attention layer, calculate the dot product correlation coefficient of any two row vectors in the input matrix, and perform Softmax normalization. Multiply the normalized correlation coefficients with the corresponding row vectors element by element and sum them to obtain multiple attention output matrices. Then, concatenate them column by column and input them into two fully connected layers to perform element-wise linear calculation and ReLU nonlinear calculation to generate the temporal feature matrix. S43. Input the time-series feature matrix into the output mapping layer, perform preset linear matrix multiplication and vector addition on each row, and output three results: capacity decay value, internal resistance change rate and temperature stability index. Arrange the three sets of values ​​in time order to obtain the battery health status result sequence. S44. Subtract the capacity decay value from the corresponding actual capacity decay value, square the difference and sum them up to obtain the capacity loss value. Subtract the internal resistance change rate from the corresponding actual internal resistance change rate, square the difference and sum them up to obtain the internal resistance loss value. Subtract the temperature stability index from the corresponding actual temperature stability index, square the difference and sum them up to obtain the temperature loss value. S45. Substitute the output internal resistance, current and initial voltage into the equivalent circuit formula in the pre-trained RC equivalent circuit model, calculate the difference between the theoretical voltage value and the predicted voltage value step by step, square the difference and sum them up to obtain the circuit constraint loss value, calculate the derivative of adjacent differences of capacity output at each time step, sum the derivative values ​​at the positions where the derivative is greater than zero to obtain the capacity monotonic constraint loss value, compare the temperature stability index of each time step with the preset maximum allowable temperature difference, sum the differences exceeding the preset value to obtain the temperature difference constraint loss value. S46. Multiply the capacity loss value, internal resistance loss value, temperature loss value, circuit constraint loss value, capacity monotonic constraint loss value, and temperature difference constraint loss value by their respective preset weights and add them together to obtain the total loss value. Backpropagate the total loss value to calculate the gradient of each layer's parameters. Multiply the gradient by the learning rate as the parameter update amount. Perform subtraction update operations on all weight parameters and bias parameters one by one to obtain the updated parameters.

[0026] In this embodiment, S5 specifically includes: S51. The battery health status results and the working condition stress characteristic sequence are combined into a joint input sequence and fed into the improved Autoformer model. The trend component representing the long-term decline trend of battery capacity and the residual component representing short-term fluctuations are extracted to extend and predict the capacity evolution trajectory of the battery in the future period and output the predicted value of battery capacity. S52. Extract the trend component of the long-term battery capacity degradation trend output by the Autoformer model, calculate the average change slope within a set time window, and determine the risk of rapid capacity degradation when the trend slope is greater than the battery capacity degradation threshold. S53. Perform cumulative summation on the changes in the internal resistance of individual cells within the corresponding time window to obtain the cumulative increase in internal resistance. When the increase exceeds the internal resistance degradation threshold, it is determined to be a risk of internal resistance degradation. S54. Calculate the temperature difference of each unit at each time step and count the fluctuation range. When the temperature difference fluctuation range exceeds the temperature difference threshold, it is judged as a risk of temperature control imbalance. S55. Divide the above three risks into low risk, medium risk and high risk according to the number of judgments, and output the corresponding battery risk level and the warning result generated by the trigger risk. For example, if the internal resistance increases and the capacity deteriorates, it is recommended to check the cooling system and the battery aging status.

[0027] In this embodiment, the improved Autoformer model specifically includes: After aligning the battery health status results with the working condition stress feature sequence by timestamp, they are spliced ​​together to form a joint input sequence. The joint input sequence is then standardized by using Z-Score standardization to calculate the mean and standard deviation of each dimension. Each value in the input sequence is then normalized to zero mean and unit variance according to the corresponding dimension to generate a standardized input matrix. An improved relative position encoding method is adopted, which maps the time step index value to a position vector through sine and cosine functions and adds it element by element to the standardized input matrix to form an input sequence matrix with position information and input it into the improved Autoformer model. A trend-residual dual-path modeling mechanism and a dynamic sliding window decomposition mechanism are introduced to capture the long-term degradation trend and short-term fluctuation anomaly in battery capacity change, respectively. A dynamic sliding window decomposition is performed by setting a time window length. The stationarity of the subsequence formed by each time window length and time step is detected and the spectrum is analyzed. Subsequences whose mean slope and frequency energy density reach the preset slope threshold and density threshold within the window are extracted as trend components, otherwise they are residual components. The operation is repeated to complete the trend-residual decomposition of the entire input sequence matrix and output the trend component sequence and residual component sequence. The trend component sequence is input into the main time series prediction path to obtain the time series input matrix, where each row represents the trend feature of a time step. The matrix is ​​standardized so that the mean of each column is 0 and the standard deviation is 1. The matrix is ​​then input into a multi-layer decoupled self-attention mechanism. The dot product similarity between time steps is calculated in each layer. The weighting coefficient is obtained by Softmax normalization and multiplied by the corresponding trend feature to obtain the weighted feature representation. The weighted feature representations at each time step are subjected to stacked convolutions with pre-defined kernel sizes and strides to extract local features and pass them to the next layer. The output of each layer is compared with the true target value, the mean squared error residual is calculated, the gradient is calculated, and the weights of each layer are updated through backpropagation. Through the self-attention mechanism and convolution operation of all layers, the dot product similarity between time steps is calculated at each layer and normalized by Softmax to obtain the weighting coefficients. The pre-weighted summation of local features is used to output the battery capacity degradation trend prediction result for future periods. The residual component sequences are aligned by time steps to form an input matrix, representing the short-term fluctuation characteristics of each time step. The extracted internal resistance change rate and temperature stability index are introduced and weighted and summed with preset weights to obtain the perturbation factor of each time step. The residual component sequences of each time step are weighted and superimposed with the corresponding perturbation factor to generate the weighted residual features. The weighted residual features are input into a feedforward neural network with residual connections. The short-term disturbances in the residuals are compensated by the ReLU activation function and the weighted summation operation with preset weights. The disturbance prediction results of each layer are accumulated layer by layer to obtain the disturbance correction result of the battery under the current operating conditions. The disturbance correction results of all layers are added to the trend prediction result to generate the corrected battery capacity prediction value. The capacity degradation trend prediction result and the residual disturbance prediction result are summed element by element over time step to form the capacity evolution prediction trajectory of the battery in the future period. Based on this prediction trajectory curve, the time required to decrease to the preset capacity threshold is calculated, and the remaining usable life prediction result of the battery at the current moment is output.

[0028] This implementation significantly improves the accuracy and predictive capability of battery health assessment through an improved Autoformer model and multi-source data fusion. First, multi-source data is collected and preprocessed to construct health-related feature vectors, which, combined with operating stress characteristics, are input into the Autoformer model. This model models the long-term trend and short-term fluctuations of battery capacity degradation using a time-series self-attention mechanism, accurately predicting battery capacity degradation. Utilizing a trend-residual dual-path modeling mechanism, the long-term trend components and short-term fluctuation anomalies of battery capacity changes are extracted, outputting the predicted battery capacity and remaining usable life. Simultaneously, by combining internal resistance changes and temperature stability indicators, multi-layer convolution operations and weighted residual features are used to correct for short-term disturbances, enhancing the stability and accuracy of the prediction. This invention's method can reflect battery health status in real time, accurately assess battery remaining life, significantly improve the intelligence level of battery management and risk warning, and has broad application prospects.

[0029] In this embodiment, S6 specifically includes: aligning the battery capacity prediction value, battery risk level and warning result according to the timestamp, splicing them into a structured data packet, encoding them according to a preset format, generating a battery health assessment result packet, and outputting the battery health assessment result packet through the data transmission interface for real-time display on the user interface.

[0030] Example 1: To verify the feasibility of applying this invention to truck battery health assessment, the method was deployed at a large logistics company to monitor the battery health status of all electric trucks in the company's fleet in real time. This logistics company currently operates over 1,000 electric trucks, with an average daily driving time exceeding 14 hours, making battery management a crucial aspect of fleet operation.

[0031] In practical applications, this invention collects multi-source data, such as truck battery voltage, current, cell temperature, state of charge, charge / discharge capacity, vehicle load, road slope, and ambient temperature, and generates a unified multi-source time-series dataset through efficient data synchronization and processing. This data not only provides real-time battery health information but also incorporates external environmental factors to accurately model the battery's operating status.

[0032] By introducing an improved Autoformer model, this invention successfully extracts the long-term trend and short-term fluctuation characteristics of battery capacity degradation. Specifically, the model first standardizes all health-related feature vectors and models degradation-dependent features through a temporal self-attention mechanism, outputting the battery's capacity degradation value, internal resistance change rate, and temperature stability index, thereby accurately assessing the battery's health status. Based on these outputs, the system can further predict the battery's remaining usable lifespan and determine whether there are any battery risks.

[0033] During implementation, comparative data on battery health assessments using traditional rule-based methods and the method of this invention were collected over the past two months. The data covers several typical battery failure types, such as capacity degradation, internal resistance changes, and temperature fluctuations. The results are shown in the table below: Table 1. Performance Comparison of the Invention Method and Traditional Rule-Based Methods

[0034] Based on the comparative data shown in Table 1, the truck battery health assessment method based on multi-source data proposed in this invention outperforms traditional methods in several key indicators. In particular, the method of this invention achieves significant improvements in the accuracy of battery health prediction, response time, assessment level matching degree, and false alarm control.

[0035] First, regarding accuracy, the method of this invention achieves an accuracy of 96.6% in the capacity degradation scenario, an improvement of 23.1 percentage points compared to the 73.5% of the traditional method. In scenarios involving internal resistance changes and temperature instability, this invention also performs excellently, with accuracies of 97.5% and 96.2%, respectively. Traditional methods rely on fixed rules and are easily affected by data noise, while this invention dynamically adapts to changes in battery state through an improved Autoformer model, thereby improving prediction accuracy.

[0036] Secondly, the response time is significantly reduced. Through deep Q-networks and reinforcement learning feedback mechanisms, the average response time of the method in this invention is reduced from 36 seconds in traditional methods to 11.6 seconds, an improvement of nearly three times, ensuring rapid intervention in battery health issues. In terms of risk level matching accuracy, this invention maintains a rate above 88%, far exceeding the approximately 60% of traditional methods. The dynamically evolving battery health status map and historical trend comparison mechanism make the level judgment more refined and accurate.

[0037] Finally, the false alarm rate and false negative rate are effectively controlled. Compared with the traditional method's 13% false alarm rate and nearly 10% false negative rate, the false alarm rate of this invention is controlled to within 4%, and the false negative rate is about 2%, significantly reducing redundant warnings and omissions.

[0038] In summary, this invention demonstrates higher accuracy, faster response speed, better risk assessment matching, and lower false alarm / false negative rate in battery health assessment, and has broad application prospects.

[0039] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for assessing the health of truck batteries based on multi-source data, characterized in that, Includes the following steps: S1. Collect multi-source data from truck batteries, perform preprocessing, and construct a multi-source time-series dataset; S2. Convert the temperature change amplitude, discharge rate, vehicle load and road slope in the multi-source time series dataset into working condition stress factors with unified dimensions, and weighted accumulate to form a working condition stress feature sequence. S3. Based on multi-source time-series datasets, calculate dynamic voltage change rate, estimated equivalent internal resistance, individual cell voltage difference, individual cell temperature difference and discharge curve slope to form a health-related feature vector. S4. Using a time-series self-attention mechanism, the health-related feature vectors are modeled to depict the long-term trend and short-term fluctuation characteristics of battery capacity degradation, resulting in capacity decay values, internal resistance change rate, and temperature stability indicators, and outputting the battery health status results. S5. Input the battery health status results and the working condition stress characteristic sequence into the improved Autoformer model, extract the long-term trend component and the short-term fluctuation residual component representing the battery capacity degradation respectively, output the battery capacity prediction value, and determine the risk event, output the battery risk level and warning result. S6. Integrate the predicted battery capacity, battery risk level, and early warning results to generate a battery health assessment result package and output it for real-time display on the user interface.

2. The truck battery health assessment method based on multi-source data according to claim 1, characterized in that, S1 specifically includes: S11. Read the raw data output of the multi-source sensors deployed on the truck to obtain multi-source data of the truck's power battery during operation; S12. Match the collected multi-source data according to the timestamp to construct a unified time axis, and perform linear interpolation processing on the data with different sampling frequencies; S13. Using a sliding window detection strategy and threshold determination method, identify and remove mutation points or multi-source data that exceed a preset reasonable threshold in multi-source data, perform outlier removal and data filtering, perform Kalman filter interpolation on data segments with missing values, and continuously fill in the values ​​at the breakpoints. S14. Arrange the multi-source data after time synchronization, anomaly removal and interpolation repair in chronological order to construct a multi-source time series dataset.

3. The truck battery health assessment method based on multi-source data according to claim 1, characterized in that, S2 specifically includes: S21. Read the data sequences corresponding to the temperature change amplitude, discharge rate, vehicle load and road slope in the multi-source time series dataset, and extract the temperature data segment, discharge rate data segment, load data segment and slope data segment corresponding to the battery operating status. S22. Perform dimensionless normalization on the temperature change range, discharge rate, vehicle load and road slope respectively, and linearly map the numerical range of each data to a unified numerical range in proportion. S23. Preset corresponding weight parameters according to the degree of influence of temperature change on battery performance, the degree of influence of discharge rate on battery load pressure, the degree of influence of vehicle load on battery mechanical stress, and the degree of influence of road slope on battery energy output demand. S24. The normalized temperature change amplitude, discharge rate, vehicle load and road slope are weighted and summed according to the weight parameters to obtain the working condition stress factor value for a single time step. The working condition stress factors of consecutive time steps are accumulated and summed in chronological order to form a working condition stress characteristic sequence.

4. The truck battery health assessment method based on multi-source data according to claim 1, characterized in that, S3 specifically includes: S31. Read the voltage data sequence from the multi-source time series dataset, perform differential operation on the voltage value changes between consecutive time points, calculate the voltage change value per unit time, and arrange all the change values ​​in time order to generate a dynamic voltage change rate sequence. S32. By inputting the battery's voltage, current data, and open-circuit voltage, the parameters are fitted using a pre-trained RC equivalent circuit model. The relationship between the battery's equivalent internal resistance and current-voltage characteristics is fitted using the least squares method to obtain the internal resistance value at each time step and generate the battery internal resistance time sequence. S33. Calculate the difference between the temperature and voltage values ​​of all individual cells at the same time point, obtain the temperature difference and pressure difference between adjacent cells, and form a cell temperature difference sequence and a cell pressure difference sequence according to the time series. S34. Extract the voltage and state of charge (SOC) data of the SOC drop interval that exceeds the continuous time threshold during the discharge phase. Select the corresponding voltage and SOC values ​​at set intervals within each SOC drop interval. Calculate the slope of voltage change with respect to SOC using the least squares linear fitting method. Use the slope values ​​of each segment as the voltage change rate index during the discharge process and form a discharge curve slope sequence in time order. S35. The dynamic voltage change rate sequence, the equivalent internal resistance estimate sequence, the single cell voltage difference sequence, the single cell temperature difference sequence, and the discharge curve slope value are standardized and spliced ​​together to form a health-related feature vector characterizing the battery's operating status.

5. The truck battery health assessment method based on multi-source data according to claim 1, characterized in that, S4 specifically includes: S41. Arrange the vectors of all time steps of the health-related feature vectors into an input matrix in order, and apply position encoding to each row of the input matrix to write the time sequence information into the input matrix to generate a feature matrix with position encoding. S42. Input the feature matrix into the multi-head self-attention layer, calculate the dot product correlation coefficient of any two row vectors in the input matrix, and perform Softmax normalization. Multiply the normalized correlation coefficients with the corresponding row vectors element by element and sum them to obtain multiple attention output matrices. Then, concatenate them column by column and input them into two fully connected layers to perform element-wise linear calculation and ReLU nonlinear calculation to generate the temporal feature matrix. S43. Input the time-series feature matrix into the output mapping layer, perform preset linear matrix multiplication and vector addition on each row, and output three results: capacity decay value, internal resistance change rate and temperature stability index. Arrange the three sets of values ​​in time order to obtain the battery health status result sequence. S44. Subtract the capacity decay value from the corresponding actual capacity decay value, square the difference and sum them up to obtain the capacity loss value. Subtract the internal resistance change rate from the corresponding actual internal resistance change rate, square the difference and sum them up to obtain the internal resistance loss value. Subtract the temperature stability index from the corresponding actual temperature stability index, square the difference and sum them up to obtain the temperature loss value. S45. Substitute the output internal resistance, current and initial voltage into the equivalent circuit formula in the pre-trained RC equivalent circuit model, calculate the difference between the theoretical voltage value and the predicted voltage value step by step, square the difference and sum them up to obtain the circuit constraint loss value, calculate the derivative of adjacent differences of capacity output at each time step, sum the derivative values ​​at the positions where the derivative is greater than zero to obtain the capacity monotonic constraint loss value, compare the temperature stability index of each time step with the preset maximum allowable temperature difference, sum the differences exceeding the preset value to obtain the temperature difference constraint loss value. S46. Multiply the capacity loss value, internal resistance loss value, temperature loss value, circuit constraint loss value, capacity monotonic constraint loss value, and temperature difference constraint loss value by their respective preset weights and add them together to obtain the total loss value. Backpropagate the total loss value to calculate the gradient of each layer's parameters. Multiply the gradient by the learning rate as the parameter update amount. Perform subtraction update operations on all weight parameters and bias parameters one by one to obtain the updated parameters.

6. The truck battery health assessment method based on multi-source data according to claim 1, characterized in that, S5 specifically includes: S51. The battery health status results and the working condition stress characteristic sequence are combined into a joint input sequence and fed into the improved Autoformer model. The trend component representing the long-term decline trend of battery capacity and the residual component representing short-term fluctuations are extracted to extend and predict the capacity evolution trajectory of the battery in the future period and output the predicted value of battery capacity. S52. Extract the trend component of the long-term battery capacity degradation trend output by the Autoformer model, calculate the average change slope within a set time window, and determine the risk of rapid capacity degradation when the trend slope is greater than the battery capacity degradation threshold. S53. Perform cumulative summation on the changes in the internal resistance of individual cells within the corresponding time window to obtain the cumulative increase in internal resistance. When the increase exceeds the internal resistance degradation threshold, it is determined to be a risk of internal resistance degradation. S54. Calculate the temperature difference of each unit at each time step and count the fluctuation range. When the temperature difference fluctuation range exceeds the temperature difference threshold, it is judged as a risk of temperature control imbalance. S55. Divide the above three risks into low risk, medium risk and high risk according to the number of judgments, and output the corresponding battery risk level and the warning result generated by combining the trigger risk.

7. The truck battery health assessment method based on multi-source data according to claim 6, characterized in that, The improved Autoformer model specifically includes: After aligning the battery health status results with the working condition stress feature sequence by timestamp, they are spliced ​​together to form a joint input sequence. The joint input sequence is then standardized by using Z-Score standardization to calculate the mean and standard deviation of each dimension. Each value in the input sequence is then normalized to zero mean and unit variance according to the corresponding dimension to generate a standardized input matrix. An improved relative position encoding method is adopted, which maps the time step index value to a position vector through sine and cosine functions and adds it element by element to the standardized input matrix to form an input sequence matrix with position information and input it into the improved Autoformer model. A trend-residual dual-path modeling mechanism and a dynamic sliding window decomposition mechanism are introduced to capture the long-term degradation trend and short-term fluctuation anomaly in battery capacity change, respectively. A dynamic sliding window decomposition is performed by setting a time window length. The stationarity of the subsequence formed by each time window length and time step is detected and the spectrum is analyzed. Subsequences whose mean slope and frequency energy density reach the preset slope threshold and density threshold within the window are extracted as trend components, otherwise they are residual components. The operation is repeated to complete the trend-residual decomposition of the entire input sequence matrix and output the trend component sequence and residual component sequence. The trend component sequence is input into the main time series prediction path to obtain the time series input matrix, where each row represents the trend feature of a time step. The matrix is ​​standardized so that the mean of each column is 0 and the standard deviation is 1. The matrix is ​​then input into a multi-layer decoupled self-attention mechanism. The dot product similarity between time steps is calculated in each layer. The weighting coefficient is obtained by Softmax normalization and multiplied by the corresponding trend feature to obtain the weighted feature representation. The weighted feature representations at each time step are subjected to stacked convolutions with pre-defined kernel sizes and strides to extract local features and pass them to the next layer. The output of each layer is compared with the true target value, the mean squared error residual is calculated, the gradient is calculated, and the weights of each layer are updated through backpropagation. Through the self-attention mechanism and convolution operation of all layers, the dot product similarity between time steps is calculated at each layer and normalized by Softmax to obtain the weighting coefficients. The pre-weighted summation of local features is used to output the battery capacity degradation trend prediction result for future periods. The residual component sequences are aligned by time steps to form an input matrix, representing the short-term fluctuation characteristics of each time step. The extracted internal resistance change rate and temperature stability index are introduced and weighted and summed with preset weights to obtain the perturbation factor of each time step. The residual component sequences of each time step are weighted and superimposed with the corresponding perturbation factor to generate the weighted residual features. The weighted residual features are input into a feedforward neural network with residual connections. The short-term disturbances in the residuals are compensated by the ReLU activation function and the weighted summation operation with preset weights. The disturbance prediction results of each layer are accumulated layer by layer to obtain the disturbance correction result of the battery under the current operating conditions. The disturbance correction results of all layers are added to the trend prediction result to generate the corrected battery capacity prediction value. The capacity degradation trend prediction result and the residual disturbance prediction result are summed element by element over time step to form the capacity evolution prediction trajectory of the battery in the future period. Based on this prediction trajectory curve, the time required to decrease to the preset capacity threshold is calculated, and the remaining usable life prediction result of the battery at the current moment is output.

8. The truck battery health assessment method based on multi-source data according to claim 1, characterized in that, S6 specifically includes: aligning the battery capacity prediction value, battery risk level and warning result according to the timestamp, splicing them into a structured data packet, encoding them according to a preset format, generating a battery health assessment result packet, outputting the battery health assessment result packet through the data transmission interface, and displaying it in real time on the user interface.

Images