Deep learning-based fast-charging charger life prediction method

Abstract

The invention provides a fast-charging charger service life prediction method based on deep learning, and the method comprises the steps: collecting electrical, thermodynamic and environmental parameters, carrying out the denoising through extended Kalman filtering, employing adaptive quantile normalization and time window weighted interpolation to process missing values, extracting the time sequence and thermodynamic characteristics in a sliding window, constructing a double-flow Transform-LSTM network, and carrying out the prediction of the service life of a fast-charging charger. The method comprises the following steps: respectively processing an electric current flow and a thermodynamic flow, introducing a stage perception attention mechanism, dynamically adjusting a feature weight, integrating double-current output through weighting, designing a physical loss function, finally outputting residual life, calculating a health state through weighting, triggering multi-stage early warning based on a threshold value, and classifying fault types in combination with Softmax. Abnormal data are selected through incremental training, the model is regularly and finely adjusted, physical constraint and deep learning are combined, charging stage characteristics are dynamically adapted, high-precision life prediction and fault diagnosis are achieved, and the method is suitable for real-time health management of fast charging equipment.

Description

Technical Field

[0001] The present invention relates to the technical field of charger life prediction, and in particular to a method for predicting the life of a fast-charging charger based on deep learning. Background Art

[0002] When collecting electrical and thermodynamic parameters, existing technologies often suffer from significant data noise due to sensor drift or environmental interference, and data missing in charging scenarios are not effectively supplemented, resulting in a decrease in the accuracy of feature extraction. Traditional methods usually rely on statistical features such as the mean and variance of a fixed window and are unable to capture the dynamic degradation patterns during the charging process. At the same time, the correlation between thermodynamic parameters and electrical parameters has not been systematically modeled, resulting in limited model generalization capabilities.

[0003] Existing life prediction models mostly use a single flow network, which makes it difficult to simultaneously handle the temporal dependencies of electrical and thermodynamic flows. Traditional models lack a stage-aware mechanism and cannot dynamically assign feature weights. Furthermore, the models fail to incorporate physical constraints such as power conservation and thermodynamic laws, resulting in deviations between prediction results and actual physical laws. Existing methods usually adopt a one-time training strategy, which cannot adapt to the dynamic changes of charging scenarios. In long-term predictions, the failure to integrate new data in a timely manner leads to the accumulation of model prediction deviations and ultimately failure. Summary of the Invention

[0004] The purpose of this invention is to provide a method for predicting the life of a fast charging charger based on deep learning.

[0005] The problem to be solved by the present invention is: to solve the problems of inaccurate life prediction of fast charging chargers in high-power and high-frequency usage scenarios, delayed sudden fault warning, difficulty in multi-source data fusion, and accumulation of long-term model prediction deviations.

[0006] A fast charger life prediction method based on deep learning, the technical solution adopted is as follows: S1: Collects electrical parameters, including current, voltage, and power; collects thermodynamic parameters, including the charger's internal temperature and ambient temperature; and collects environmental data, including humidity, usage frequency, fast charging mode ratio, and input voltage fluctuations; S2: Use extended Kalman filtering to remove noise from the data collected by S1, process the denoised data in S1, including normalization and interpolation, and use cubic spline interpolation combined with time window to fill in missing values; S3: Extract the time series characteristics of the data processed in S2, including the current, voltage mean, variance, power efficiency, power fluctuation rate, and temperature change rate; and extract the thermodynamic characteristics of the data processed in S2, including the heat dissipation efficiency and temperature fluctuation standard deviation; S4: Integrate electrical, thermodynamic, and environmental data into multi-channel time series to form input tensors. Use a two-stream Transformer-LSTM network to process electrical and thermodynamic flows, including processing current, voltage, power efficiency, capturing timing dependencies, processing temperature and heat dissipation efficiency, analyzing thermodynamic dynamics, establishing a stage-aware attention mechanism, dynamically assigning feature weights based on the charging stage, and designing a dynamic weight function based on the charging stage. S5: Build a two-stream Transformer-LSTM network. The fusion layer integrates the two-stream outputs through dynamic weighting and designs a loss function based on physical constraints. S6: Time warp and spatial jitter are performed on the current and voltage curves. The Transformer-LSTM model is pre-trained on a common dataset, and the bottom Transformer-LSTM layer is frozen to remove redundant neurons. S7: Establish a fast charger remaining life prediction model. Based on the model output, predict the remaining life, perform health status monitoring, calculate the health status by weight, output warning signals, and predict the specific failure type based on the fault classification module. S8: Select data points with abnormal information content, including abnormal temperature fluctuations, for incremental training, and regularly use new data to fine-tune the model to prevent the accumulation of prediction bias.

[0007] Furthermore, the electrical parameters, thermodynamic parameters, and environmental data collected in S1 include: Integrates Hall effect sensors and Rogowski coils for contactless current measurement, uses a 6.5-bit digital voltage module for voltage acquisition, and calculates power based on the measured current and voltage; A distributed thermopile sensor array is used to measure the charger temperature, and a digital temperature sensor is used to collect the ambient temperature; A capacitive humidity sensor is used to collect ambient humidity data. The start and end times of each charge are recorded based on the STM32 to calculate the usage frequency. Fast charging mode is defined as a charging current greater than 3A. The fast charging mode ratio is the proportion of fast charging mode time to the total charging time. The standard deviation of the input voltage is calculated. If the voltage standard deviation is greater than 5% of the standardized charging voltage, it indicates that the input voltage is fluctuating.

[0008] Furthermore, the denoised data in S1 is processed in S2, including normalization and interpolation, and missing values are filled using cubic spline interpolation combined with a time window, including: Adaptive quantile normalization is used. , , , ,in is the original data, is the normalized data, is the 5th quantile of the data, is the 95th percentile of the data, To obtain the maximum and minimum values, quantiles are calculated for each data series. The values outside the range are marked as outliers and removed; Using non-uniform cubic spline interpolation, introducing time window weighted interpolation, , ,in is the final weighted interpolation result, the estimated value at time t, is the number of local interpolation results, is the weight of the j-th interpolation result, is the jth local interpolation function, is the absolute time difference between the center time of the jth data window and the current interpolation time t, is the time constant.

[0009] Furthermore, the step S3 extracts the temporal characteristics of the data processed in S2 and extracts the thermodynamic characteristics of the data processed in S2, including: Select a window length of 1 second and divide the time series data into sliding windows with an overlap rate of 50%. Calculate the mean and variance of the current and voltage values, and the temperature change rate based on the temperature value. Based on the selected window length, the instantaneous power of each window is calculated, the power difference between adjacent windows is differentiated to obtain the power fluctuation rate, and the ratio of the output power to the input power of the window length is calculated and recorded as the power efficiency; Based on the selected window length, the difference between the charger internal temperature and the ambient temperature is calculated, and the heat dissipation efficiency is obtained by dividing the difference by the input power; Based on the selected window length, the temperature fluctuation standard deviation is calculated from the sliding window standard deviation combined with the temperature values.

[0010] Furthermore, the S4 uses a dual-stream Transformer to process electrical and thermodynamic flows, establishes a stage-aware attention mechanism, dynamically allocates feature weights according to the charging stage, and designs a dynamic weight function based on the charging stage, including: S41: Processing the timing dependencies of current, voltage, and power efficiency using a self-attention mechanism. ,in is the query, key, and value matrix of the electrical flow channel, is the dimension of the key, As a function, input the electrical flow channel data and generate , calculate the self-attention weight and output the dimensionally compressed feature vector; S42: Processing the thermodynamic dynamics of temperature and heat dissipation efficiency, using the self-attention mechanism, ,in is the query, key, and value matrix of the thermodynamic channel, is the dimension of the key, is the thermodynamic constraint matrix, is a function that satisfies the laws of thermodynamics and is generated by linear projection , add the constraint matrix , calculate the self-attention weight and output the dimensionally compressed feature vector; S43: Based on the current average value described in S3 The fast charge current threshold 3A is divided into stages as described in S1 , ,calculate the current mean in real time, determine the current stage, and generate a stage identifier sequence; S44: Combination stage Design weight function, , ,in is the electrical current weight function, is the thermodynamic flow weight function, are coefficients, which are 0.5, 0.3, is the standard deviation of temperature fluctuation described in S3.

[0011] Furthermore, the fusion layer in S5 integrates the dual-stream outputs through dynamic weighting and designs a loss function based on physical constraints, including: Based on the dynamic weights described in S4, the dual-stream output is weighted and integrated. , ,in is the characteristic after electrical flow fusion, is the characteristic after thermodynamic flow fusion, is the electrical current weight function, is the thermodynamic flow weight function, Output features for the electrical flow Transformer, Output features for the thermodynamic flow Transformer, is the electrical flow LSTM output feature, Output features for thermodynamic flow LSTM; Design the total loss function, ,in is the constraint weight, which is 0.1; ,in is the main task loss function, is the life expectancy prediction weight, is the sample size, is the fault prediction weight, For the predicted charger life, For the actual charger life, is the binary cross entropy loss, is the predicted failure probability, is the actual fault label; Design penalty terms based on power conservation constraints , ,in is the output power, is the input power, is the power efficiency, is the power conservation tolerance threshold, To obtain the maximum value; Design penalty items based on heat dissipation efficiency constraints , , For heat dissipation efficiency, is the device temperature, is the ambient temperature, is the heat dissipation efficiency tolerance threshold; Design penalty terms based on temperature fluctuation constraints , ,in is the standard deviation of temperature fluctuation, is the temperature fluctuation safety threshold.

[0012] Furthermore, S6 performs time warping and spatial jittering on the current and voltage curves, freezes the underlying Transformer-LSTM layer, and removes redundant neurons, including: S61: Based on the characteristics of the current and voltage curves, the time axis is elastically deformed and combined with the periodic perturbation of the sine function to simulate the timing disturbance during the charging process. Cubic spline interpolation is used to fill the non-integer time points after distortion, and physical constrained random perturbations are added to the spatial dimensions of current and voltage. S62: Pre-train the Transformer-LSTM model on the General Power Equipment dataset, freeze the first three encoder layers of the model, disable gradient updates for the frozen layers, and fine-tune the remaining layers with a learning rate of 1. e −5; S63: Remove redundant neurons based on gradient importance, calculate the redundancy of each neuron on the validation set, and gradually remove redundant neurons until the model size is reduced by 50%.

[0013] Furthermore, in S7, based on the model output, the remaining life is predicted, the health status is weightedly calculated, an early warning signal is output, and the specific failure type is predicted based on the fault classification module, including: S71: Construct model input based on the collected parameters, ,in is the power efficiency, is the temperature change rate, For heat dissipation efficiency, For humidity, is the frequency of use, The ratio of fast charging mode is input into the Transformer-LSTM model, and the remaining life is output based on the fully connected layer; S72: Calculate health status based on the weighted parameters in S71 , ,in are the corresponding weights, which are 0.3, 0.2, 0.2, 0.15, and 0.15 respectively. is the temperature change rate safety threshold, is the humidity threshold, is the frequency safety threshold; S73: Triggering multiple levels of warnings based on thresholds, divided into red warning, yellow warning, and normal situation; Red warning: When any of the following conditions are met: remaining life < remaining life critical value, health status < health status critical value, temperature change rate > temperature change rate safety threshold, humidity > humidity threshold; Yellow warning: When any of the following conditions are met: Remaining life critical value ≤ Remaining life < 2 times remaining life critical value, Health status critical value ≤ Health status < 2 times health status critical value, Temperature change rate > 0.7 times the temperature change rate safety threshold, Humidity > 0.7 times the humidity threshold; Normal situation: The normal situation is when the remaining conditions are met; S74: Outputting a fault type based on the parameters in S71, including an overheating fault, a thermal runaway fault, and an overload fault, and outputting a probability of the fault type based on a Softmax layer; The overheating fault trigger condition is that the temperature change rate is greater than the temperature change rate safety threshold, the thermal runaway fault trigger condition is that the humidity is greater than the humidity threshold and the heat dissipation efficiency is less than 0.5, and the overload fault trigger condition is that the usage frequency is greater than the usage frequency safety threshold.

[0014] The beneficial effects of the present invention are: through multi-dimensional data collection, dynamic feature modeling, physical constraint-driven model optimization and adaptive incremental training, accurate prediction of the charger health status and failure warning are achieved; Combining extended Kalman filtering with adaptive quantile normalization technology effectively removes noise and handles data missing, ensuring high quality and robustness of input data; The dual-stream Transformer-LSTM architecture automatically adjusts feature weights and performs stage-aware temporal feature modeling, enabling the model to focus on key parameters under different working conditions, improving the ability to capture temporal dependencies, and enhancing the model's adaptability to different working conditions. Through a multi-level early warning system and fault type classification, a closed-loop management of the entire process from remaining life prediction to failure mode identification is achieved, which helps identify risks in advance, reduce maintenance costs and extend equipment life. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a fast charging charger life prediction method based on deep learning. DETAILED DESCRIPTION

[0016] The present invention is further clearly and completely described below, but the protection scope of the present invention is not limited thereto.

[0017] A fast charger life prediction method based on deep learning, the technical solution adopted is as follows: S1: Collects electrical parameters, including current, voltage, and power; collects thermodynamic parameters, including the charger's internal temperature and ambient temperature; and collects environmental data, including humidity, usage frequency, fast charging mode ratio, and input voltage fluctuations; S2: Use extended Kalman filtering to remove noise from the data collected by S1, process the denoised data in S1, including normalization and interpolation, and use cubic spline interpolation combined with time window to fill in missing values; S3: Extract the time series characteristics of the data processed in S2, including the current, voltage mean, variance, power efficiency, power fluctuation rate, and temperature change rate; and extract the thermodynamic characteristics of the data processed in S2, including the heat dissipation efficiency and temperature fluctuation standard deviation; S4: Integrate electrical, thermodynamic, and environmental data into multi-channel time series to form input tensors. Use a two-stream Transformer-LSTM network to process electrical and thermodynamic flows, including processing current, voltage, power efficiency, capturing timing dependencies, processing temperature and heat dissipation efficiency, analyzing thermodynamic dynamics, establishing a stage-aware attention mechanism, dynamically assigning feature weights based on the charging stage, and designing a dynamic weight function based on the charging stage. S5: Build a two-stream Transformer-LSTM network. The fusion layer integrates the two-stream outputs through dynamic weighting and designs a loss function based on physical constraints. S6: Time warp and spatial jitter are performed on the current and voltage curves. The Transformer-LSTM model is pre-trained on a common dataset, and the bottom Transformer-LSTM layer is frozen to remove redundant neurons. S7: Establish a fast charger remaining life prediction model. Based on the model output, predict the remaining life, perform health status monitoring, calculate the health status by weight, output warning signals, and predict the specific failure type based on the fault classification module. S8: Select data points with abnormal information content, including abnormal temperature fluctuations, for incremental training, and regularly use new data to fine-tune the model to prevent the accumulation of prediction bias.

[0018] refer to Figure 1 As shown in the figure, a fast charging charger life prediction method based on deep learning is proposed.

[0019] Furthermore, the electrical parameters, thermodynamic parameters, and environmental data collected in S1 include: Integrates Hall effect sensors and Rogowski coils for contactless current measurement, uses a 6.5-bit digital voltage module for voltage acquisition, and calculates power based on the measured current and voltage; A distributed thermopile sensor array is used to measure the charger temperature, and a digital temperature sensor is used to collect the ambient temperature; A capacitive humidity sensor is used to collect ambient humidity data. The start and end times of each charge are recorded based on the STM32 to calculate the usage frequency. Fast charging mode is defined as a charging current greater than 3A. The fast charging mode ratio is the proportion of fast charging mode time to the total charging time. The standard deviation of the input voltage is calculated. If the voltage standard deviation is greater than 5% of the standardized charging voltage, it indicates that the input voltage is fluctuating.

[0020] Furthermore, the denoised data in S1 is processed in S2, including normalization and interpolation, and missing values are filled using cubic spline interpolation combined with a time window, including: Adaptive quantile normalization is used. , , , ,in is the original data, is the normalized data, is the 5th quantile of the data, is the 95th percentile of the data, To obtain the maximum and minimum values, quantiles are calculated for each data series. The values outside the range are marked as outliers and removed; Using non-uniform cubic spline interpolation, introducing time window weighted interpolation, , ,in is the final weighted interpolation result, the estimated value at time t, is the number of local interpolation results, is the weight of the j-th interpolation result, is the jth local interpolation function, is the absolute time difference between the center time of the jth data window and the current interpolation time t, is the time constant.

[0021] Furthermore, the step S3 extracts the temporal characteristics of the data processed in S2 and extracts the thermodynamic characteristics of the data processed in S2, including: Select a window length of 1 second and divide the time series data into sliding windows with an overlap rate of 50%. Calculate the mean and variance of the current and voltage values, and the temperature change rate based on the temperature value. Based on the selected window length, the instantaneous power of each window is calculated, the power difference between adjacent windows is differentiated to obtain the power fluctuation rate, and the ratio of the output power to the input power of the window length is calculated and recorded as the power efficiency; Based on the selected window length, the difference between the charger internal temperature and the ambient temperature is calculated, and the heat dissipation efficiency is obtained by dividing the difference by the input power; Based on the selected window length, the temperature fluctuation standard deviation is calculated from the sliding window standard deviation combined with the temperature values.

[0022] Furthermore, the S4 uses a dual-stream Transformer to process electrical and thermodynamic flows, establishes a stage-aware attention mechanism, dynamically allocates feature weights according to the charging stage, and designs a dynamic weight function based on the charging stage, including: S41: Processing the timing dependencies of current, voltage, and power efficiency using a self-attention mechanism. ,in is the query, key, and value matrix of the electrical flow channel, is the dimension of the key, As a function, input the electrical flow channel data and generate , calculate the self-attention weight and output the dimensionally compressed feature vector; S42: Processing the thermodynamic dynamics of temperature and heat dissipation efficiency, using the self-attention mechanism, ,in is the query, key, and value matrix of the thermodynamic channel, is the dimension of the key, is the thermodynamic constraint matrix, is a function that satisfies the laws of thermodynamics and is generated by linear projection , add the constraint matrix , calculate the self-attention weight and output the dimensionally compressed feature vector; S43: Based on the current average value described in S3 The fast charge current threshold 3A is divided into stages as described in S1 , ,calculate the current mean in real time, determine the current stage, and generate a stage identifier sequence; S44: Combination stage Design weight function, , ,in is the electrical current weight function, is the thermodynamic flow weight function, are coefficients, which are 0.5, 0.3, is the standard deviation of temperature fluctuation described in S3.

[0023] Furthermore, the fusion layer in S5 integrates the dual-stream outputs through dynamic weighting and designs a loss function based on physical constraints, including: Based on the dynamic weights described in S4, the dual-stream output is weighted and integrated. , ,in is the characteristic after electrical flow fusion, is the characteristic after thermodynamic flow fusion, is the electrical current weight function, is the thermodynamic flow weight function, Output features for the electrical flow Transformer, Output features for the thermodynamic flow Transformer, is the electrical flow LSTM output feature, Output features for thermodynamic flow LSTM; Design the total loss function, ,in is the constraint weight, which is 0.1; ,in is the main task loss function, is the life expectancy prediction weight, is the sample size, is the fault prediction weight, For the predicted charger life, For the actual charger life, is the binary cross entropy loss, is the predicted failure probability, is the actual fault label; Design penalty terms based on power conservation constraints , ,in is the output power, is the input power, is the power efficiency, is the power conservation tolerance threshold, To obtain the maximum value; Design penalty items based on heat dissipation efficiency constraints , , For heat dissipation efficiency, is the device temperature, is the ambient temperature, is the heat dissipation efficiency tolerance threshold; Design penalty terms based on temperature fluctuation constraints , ,in is the standard deviation of temperature fluctuation, is the temperature fluctuation safety threshold.

[0024] Furthermore, S6 performs time warping and spatial jittering on the current and voltage curves, freezes the underlying Transformer-LSTM layer, and removes redundant neurons, including: S61: Based on the characteristics of the current and voltage curves, the time axis is elastically deformed and combined with the periodic perturbation of the sine function to simulate the timing disturbance during the charging process. Cubic spline interpolation is used to fill the non-integer time points after distortion, and physical constrained random perturbations are added to the spatial dimensions of current and voltage. S62: Pre-train the Transformer-LSTM model on the General Power Equipment dataset, freeze the first three encoder layers of the model, disable gradient updates for the frozen layers, and fine-tune the remaining layers with a learning rate of 1. e −5; S63: Remove redundant neurons based on gradient importance, calculate the redundancy of each neuron on the validation set, and gradually remove redundant neurons until the model size is reduced by 50%.

[0025] Furthermore, in S7, based on the model output, the remaining life is predicted, the health status is weightedly calculated, an early warning signal is output, and the specific failure type is predicted based on the fault classification module, including: S71: Construct model input based on the collected parameters, ,in is the power efficiency, is the temperature change rate, For heat dissipation efficiency, For humidity, is the frequency of use, The ratio of fast charging mode is input into the Transformer-LSTM model, and the remaining life is output based on the fully connected layer; S72: Calculate health status based on the weighted parameters in S71 , ,in are the corresponding weights, which are 0.3, 0.2, 0.2, 0.15, and 0.15 respectively. is the temperature change rate safety threshold, is the humidity threshold, is the frequency safety threshold; S73: Triggering multiple levels of warnings based on thresholds, divided into red warning, yellow warning, and normal situation; Red warning: When any of the following conditions are met: remaining life < remaining life critical value, health status < health status critical value, temperature change rate > temperature change rate safety threshold, humidity > humidity threshold; Yellow warning: When any of the following conditions are met: Remaining life critical value ≤ Remaining life < 2 times remaining life critical value, Health status critical value ≤ Health status < 2 times health status critical value, Temperature change rate > 0.7 times the temperature change rate safety threshold, Humidity > 0.7 times the humidity threshold; Normal situation: The normal situation is when the remaining conditions are met; S74: Outputting a fault type based on the parameters in S71, including an overheating fault, a thermal runaway fault, and an overload fault, and outputting a probability of the fault type based on a Softmax layer; The overheating fault trigger condition is that the temperature change rate is greater than the temperature change rate safety threshold, the thermal runaway fault trigger condition is that the humidity is greater than the humidity threshold and the heat dissipation efficiency is less than 0.5, and the overload fault trigger condition is that the usage frequency is greater than the usage frequency safety threshold.

[0026] The present invention provides a fast charging charger life prediction method based on deep learning. The method collects electrical, thermodynamic and environmental parameters, performs denoising through extended Kalman filtering, adopts adaptive quantile normalization and time window weighted interpolation to process missing values, extracts timing and thermodynamic features within the sliding window, constructs a dual-stream Transformer-LSTM network to process electrical flow and thermodynamic flow respectively, introduces a stage-aware attention mechanism, dynamically adjusts feature weights, integrates dual-stream outputs through weighted integration, designs a physical loss function, and finally outputs the remaining life. The health status is weightedly calculated, and multi-level warnings are triggered based on thresholds. Combined with Softmax classification of fault types, abnormal data is selected through incremental training and the model is regularly fine-tuned. Combining physical constraints with deep learning, the method dynamically adapts to charging stage characteristics to achieve high-precision life prediction and fault diagnosis, which is suitable for real-time health management of fast charging equipment.

Claims

1. A fast charging charger life prediction method based on deep learning, characterized in that: include: S1: Collects electrical parameters, including current, voltage, and power; collects thermodynamic parameters, including the charger's internal temperature and ambient temperature; and collects environmental data, including humidity, usage frequency, fast charging mode ratio, and input voltage fluctuations; S2: Use extended Kalman filtering to remove noise from the data collected by S1, process the denoised data in S1, including normalization and interpolation, and use cubic spline interpolation combined with time window to fill in missing values; S3: Extract the time series characteristics of the data processed in S2, including the current, voltage mean, variance, power efficiency, power fluctuation rate, and temperature change rate; and extract the thermodynamic characteristics of the data processed in S2, including the heat dissipation efficiency and temperature fluctuation standard deviation; S4: Integrate electrical, thermodynamic, and environmental data into multi-channel time series to form input tensors. Use a two-stream Transformer-LSTM network to process electrical and thermodynamic flows, including processing current, voltage, power efficiency, capturing timing dependencies, processing temperature and heat dissipation efficiency, analyzing thermodynamic dynamics, establishing a stage-aware attention mechanism, dynamically assigning feature weights based on the charging stage, and designing a dynamic weight function based on the charging stage. S5: Build a two-stream Transformer-LSTM network. The fusion layer integrates the two-stream outputs through dynamic weighting and designs a loss function based on physical constraints. S6: Time warp and spatial jitter are performed on the current and voltage curves. The Transformer-LSTM model is pre-trained on a common dataset, and the bottom Transformer-LSTM layer is frozen to remove redundant neurons. S7: Establish a fast charger remaining life prediction model. Based on the model output, predict the remaining life, perform health status monitoring, calculate the health status by weight, output warning signals, and predict the specific failure type based on the fault classification module. S8: Select data points with abnormal information content, including abnormal temperature fluctuations, for incremental training, and regularly use new data to fine-tune the model to prevent the accumulation of prediction bias.

2. A method for predicting the life of a fast charging charger based on deep learning according to claim 1, characterized in that: The electrical parameters, thermodynamic parameters and environmental data collected in S1 include: Integrates Hall effect sensors and Rogowski coils for contactless current measurement, uses a 6.5-bit digital voltage module for voltage acquisition, and calculates power based on the measured current and voltage; A distributed thermopile sensor array is used to measure the charger temperature, and a digital temperature sensor is used to collect the ambient temperature; A capacitive humidity sensor is used to collect ambient humidity data. The start and end times of each charge are recorded based on the STM32 to calculate the usage frequency. Fast charging mode is defined as a charging current greater than 3A. The fast charging mode ratio is the proportion of fast charging mode time to the total charging time. The standard deviation of the input voltage is calculated. If the voltage standard deviation is greater than 5% of the standardized charging voltage, it indicates that the input voltage is fluctuating.

3. The method for predicting the life of a fast charging charger based on deep learning according to claim 1, characterized in that: In S2, the denoised data in S1 is processed, including normalization and interpolation, and missing values are filled using cubic spline interpolation combined with a time window, including: Adaptive quantile normalization is used. , , , ,in is the original data, is the normalized data, is the 5th quantile of the data, is the 95th percentile of the data, To obtain the maximum and minimum values, quantiles are calculated for each data series. The values outside the range are marked as outliers and removed; Using non-uniform cubic spline interpolation, introducing time window weighted interpolation, , ,in is the final weighted interpolation result, the estimated value at time t, is the number of local interpolation results, is the weight of the j-th interpolation result, is the jth local interpolation function, is the absolute time difference between the center time of the jth data window and the current interpolation time t, is the time constant.

4. The method for predicting the life of a fast charger based on deep learning according to claim 1, characterized in that: The step S3 extracts the temporal features of the data processed in S2 and extracts the thermodynamic features of the data processed in S2, including: Select a window length of 1 second and divide the time series data into sliding windows with an overlap rate of 50%. Calculate the mean and variance of the current and voltage values, and the temperature change rate based on the temperature value. Based on the selected window length, the instantaneous power of each window is calculated, the power difference between adjacent windows is differentiated to obtain the power fluctuation rate, and the ratio of the output power to the input power of the window length is calculated and recorded as the power efficiency; Based on the selected window length, the difference between the charger internal temperature and the ambient temperature is calculated, and the heat dissipation efficiency is obtained by dividing the difference by the input power; Based on the selected window length, the temperature fluctuation standard deviation is calculated from the sliding window standard deviation combined with the temperature values.

5. The method for predicting the life of a fast charger based on deep learning according to claim 1, characterized in that: S4 uses a dual-stream Transformer to process electrical and thermodynamic flows, establishes a stage-aware attention mechanism, dynamically allocates feature weights according to the charging stage, and designs a dynamic weight function based on the charging stage, including: S41: Processing the timing dependencies of current, voltage, and power efficiency using a self-attention mechanism. ,in is the query, key, and value matrix of the electrical flow channel, is the dimension of the key, As a function, input the electrical flow channel data and generate , calculate the self-attention weight and output the dimensionally compressed feature vector; S42: Processing the thermodynamic dynamics of temperature and heat dissipation efficiency, using the self-attention mechanism, ,in is the query, key, and value matrix of the thermodynamic channel, is the dimension of the key, is the thermodynamic constraint matrix, is a function that satisfies the laws of thermodynamics and is generated by linear projection , add the constraint matrix , calculate the self-attention weight and output the dimensionally compressed feature vector; S43: Based on the current average value described in S3 The fast charge current threshold 3A is divided into stages as described in S1 , ,calculate the current mean in real time, determine the current stage, and generate a stage identifier sequence; S44: Combination stage Design weight function, , ,in is the electrical current weight function, is the thermodynamic flow weight function, are coefficients, which are 0.5, 0.3, is the standard deviation of temperature fluctuation described in S3.

6. The method for predicting the life of a fast charger based on deep learning according to claim 1, characterized in that: The fusion layer in S5 integrates the dual-stream outputs through dynamic weighting and designs a loss function based on physical constraints, including: Based on the dynamic weights described in S4, the dual-stream output is weighted and integrated. , ,in is the characteristic after electrical flow fusion, is the characteristic after thermodynamic flow fusion, is the electrical current weight function, is the thermodynamic flow weight function, Output features for the electrical flow Transformer, Output features for the thermodynamic flow Transformer, is the electrical flow LSTM output feature, Output features for thermodynamic flow LSTM; Design the total loss function, ,in is the constraint weight, which is 0.1; ,in is the main task loss function, is the life expectancy prediction weight, is the sample size, is the fault prediction weight, For the predicted charger life, For the actual charger life, is the binary cross entropy loss, is the predicted failure probability, is the actual fault label; Design penalty terms based on power conservation constraints , ,in is the output power, is the input power, is the power efficiency, is the power conservation tolerance threshold, To obtain the maximum value; Design penalty items based on heat dissipation efficiency constraints , , For heat dissipation efficiency, is the device temperature, is the ambient temperature, is the heat dissipation efficiency tolerance threshold; Design penalty terms based on temperature fluctuation constraints , ,in is the standard deviation of temperature fluctuation, is the temperature fluctuation safety threshold.

7. The method for predicting the life of a fast charger based on deep learning according to claim 1, characterized in that: In S6, the current and voltage curves are subjected to time warping and spatial dithering, the underlying Transformer-LSTM layer is frozen, and redundant neurons are removed, including: S61: Based on the characteristics of the current and voltage curves, the time axis is elastically deformed and combined with the periodic perturbation of the sine function to simulate the timing perturbation during the charging process. Cubic spline interpolation is used to fill the non-integer time points after distortion, and physical constrained random perturbations are added to the spatial dimensions of current and voltage. S62: Pre-train the Transformer-LSTM model on the General Power Equipment dataset, freeze the first three encoder layers of the model, disable gradient updates for the frozen layers, and fine-tune the remaining layers with a learning rate of 1. e −5; S63: Remove redundant neurons based on gradient importance, calculate the redundancy of each neuron on the validation set, and gradually remove redundant neurons until the model size is reduced by 50%.

8. The method for predicting the life of a fast charger based on deep learning according to claim 1, characterized in that: The S7 predicts the remaining life based on the model output, calculates the health status by weight, outputs the warning signal, and predicts the specific failure type based on the fault classification module, including: S71: Construct model input based on the collected parameters, ,in is the power efficiency, is the temperature change rate, For heat dissipation efficiency, For humidity, is the frequency of use, The ratio of fast charging mode is input into the Transformer-LSTM model, and the remaining life is output based on the fully connected layer; S72: Calculate health status based on the weighted parameters in S71 , ,in are the corresponding weights, which are 0.3, 0.2, 0.2, 0.15, and 0.15 respectively. is the temperature change rate safety threshold, is the humidity threshold, is the frequency safety threshold; S73: Triggering multiple levels of warnings based on thresholds, divided into red warning, yellow warning, and normal situation; Red warning: When any of the following conditions are met: remaining life < remaining life critical value, health status < health status critical value, temperature change rate > temperature change rate safety threshold, humidity > humidity threshold; Yellow warning: When any of the following conditions are met: Remaining life critical value ≤ Remaining life < 2 times remaining life critical value, Health status critical value ≤ Health status < 2 times health status critical value, Temperature change rate > 0.7 times the temperature change rate safety threshold, Humidity > 0.7 times the humidity threshold; Normal situation: The normal situation is when the remaining conditions are met; S74: Outputting a fault type based on the parameters in S71, including an overheating fault, a thermal runaway fault, and an overload fault, and outputting a probability of the fault type based on a Softmax layer; The overheating fault trigger condition is that the temperature change rate is greater than the temperature change rate safety threshold, the thermal runaway fault trigger condition is that the humidity is greater than the humidity threshold and the heat dissipation efficiency is less than 0.5, and the overload fault trigger condition is that the usage frequency is greater than the usage frequency safety threshold.

Images