Electric vehicle battery state real-time monitoring system based on deep learning

Abstract

The invention relates to the technical field of battery monitoring, in particular to an electric vehicle battery state real-time monitoring system based on deep learning. Comprising a battery data acquisition module, a feature extraction module, a battery state prediction module, a fault monitoring module and a safety regulation and control module. The battery data acquisition module acquires battery parameters in real time, the feature extraction module extracts key features through deep learning, and the battery state prediction module predicts battery capacity, health state and performance degradation based on a neural network and a differential equation. The fault monitoring module adopts integrated learning to detect abnormity and output an alarm signal, and the safety regulation and control module optimizes a charging and discharging strategy through reinforcement learning and fuzzy control to guarantee safe and stable operation of the battery. According to the invention, accurate prediction and intelligent monitoring of the battery state are realized through the deep learning technology, comprehensive state evaluation and fault early warning are provided, and the accuracy and stability of battery state monitoring are effectively improved, so that the operation safety of the electric vehicle is improved, and the service life of the battery is prolonged.

Description

Technical Field

[0001] The present invention relates to the technical field of battery monitoring, and particularly to a real-time monitoring system for the state of an electric vehicle battery based on deep learning. Background Art

[0002] With the rapid development of electric transportation, electric vehicles have gradually become one of the main means of urban transportation due to their environmental protection, energy conservation and other advantages. The core component of an electric vehicle is the power battery, and its performance and state directly affect the endurance and running safety of the electric vehicle. However, due to problems such as capacity attenuation, overcharging and over-discharging, and overheating that occur during the long-term use of the battery, the running state of the electric vehicle becomes unstable, and even safety accidents may be caused. Therefore, it is of great significance to monitor the state of the electric vehicle battery in real time. There are still the following problems at present: existing systems are difficult to accurately reflect the non-linear characteristics of the battery under complex working conditions, especially in high and low temperature environments, aging states, and large current discharge situations, and the accuracy significantly decreases; existing systems mainly rely on mathematical models or shallow algorithms, and it is difficult to effectively cope with the long-term evolution process of the battery state, lacking the comprehensive prediction ability for the future battery health state; the fault identification of existing technologies based on threshold detection or single algorithms is vulnerable to noise interference, and the recognition accuracy and stability for multiple fault modes are relatively low. Summary of the Invention

[0003] To solve the above problems, the present invention provides a real-time monitoring system for the state of an electric vehicle battery based on deep learning, which solves the problems of how to accurately monitor the state of the electric vehicle battery under complex working conditions, accurately reflect its non-linear characteristics, effectively predict the long-term evolution process, and identify multiple fault modes under noise interference, thereby effectively improving the accuracy, stability and safety of battery state monitoring.

[0004] To achieve the above object, the technical solution adopted by the present invention is:

[0005] A real-time monitoring system for the state of an electric vehicle battery based on deep learning, comprising a battery data acquisition module, a feature extraction module, a battery state prediction module, a fault monitoring module and a safety regulation module that are communicatively connected in sequence;

[0006] The battery data acquisition module is used to collect the real-time operating parameters of the battery in real time through a distributed sensor array;

[0007] The feature extraction module is used to extract key features from the operating parameters based on a deep feature extraction model of a convolutional neural network and multi-modal data fusion technology;

[0008] The battery state prediction module is used to construct a battery state prediction model based on the key features, using a neural network algorithm and differential equations to predict the state evolution of the battery and generate an output result, where the output result includes the remaining capacity, health state, performance degradation rate, and future charge and discharge efficiency of the battery;

[0009] The fault monitoring module is used to classify and identify potential fault modes according to the output result, using an anomaly detection algorithm based on ensemble learning, and output a hierarchical alarm signal; the fault modes include overheating, overcharging, over-discharging, cell imbalance, and short-circuit risk;

[0010] The safety regulation module is used to dynamically adjust the charge and discharge strategy and operating parameters of the battery based on the alarm signal, using a reinforcement learning algorithm and a fuzzy logic control algorithm, and execute safety protection measures, including adjusting the charging current, equalizing the cell voltage, starting active cooling, or isolating abnormal cells.

[0011] Further, the operating parameters include the current of the battery, cell voltage, total voltage output, charge and discharge current, temperature gradient distribution, and internal resistance parameters.

[0012] Further, the operation process of the feature extraction module includes the following steps:

[0013] Perform preprocessing operations on the operating parameters and construct a three-dimensional feature matrix, including a temperature space topology matrix, a voltage frequency domain matrix, and a current-internal resistance time-varying matrix;

[0014] Design a multi-branch convolutional neural network architecture, construct a feature extraction model based on the convolutional neural network, and generate deep feature maps of temperature distribution patterns, internal resistance change rates, and voltage fluctuation spectra;

[0015] Based on the deep feature maps, use tensor contraction operations to fuse multi-branch outputs, perform frequency band optimization on cross-modal features through a spectral attention mechanism, calculate the energy entropy of each frequency band, and perform weighted screening with convolutional features to screen key features;

[0016] Use deep residual compression to reduce the dimension of the key feature subset, eliminate redundant noise components through a differentiable threshold function, and output key features; the key features include temperature distribution patterns, internal resistance change rates, and voltage fluctuation spectra.

[0017] Even further, the multi-branch convolutional neural network architecture specifically includes that the first branch uses an atrous convolutional kernel to extract long-range dependence features of the temperature space topology, the second branch uses temporal causal convolution to capture the non-linear dynamic characteristics of the internal resistance parameters, and the third branch analyzes the phase-amplitude coupling relationship of the voltage frequency domain matrix through a complex-valued convolutional layer.

[0018] Furthermore, the formula of the feature extraction model is as follows:

[0019] ;

[0020] Among them, represents the finally extracted key feature vector; represents the input matrix, that is, the input data of the i-th branch; represents the convolutional kernel weight matrix; represents the bias term; represents the activation function; represents the adaptive weighting coefficient; represents the variational weight coefficient; represents the tensor compression operation; represents the variational feature extraction function; represents the non-linear fusion function; represents the number of convolutional layers in each branch.

[0021] Further, the operation process of the battery state prediction module includes the following steps:

[0022] Based on the key features, a battery state prediction model is constructed using neural network algorithms and differential equations;

[0023] Perform phase space embedding on the key features, use a variational encoder to extract latent variables to represent the initial state of the battery degradation trajectory, and fuse real-time working condition labels and environmental noise to construct a multi-dimensional state input;

[0024] Based on the multi-dimensional state input, solve the battery state prediction model using neural differential equations, calculate the gradient using automatic differentiation methods, and achieve feature reconstruction and error correction through residual connections to obtain the battery state prediction value;

[0025] Through an adaptive error correction mechanism, combine the residual between the model prediction value and the measured value to dynamically update the model parameters.

[0026] Furthermore, the formula of the battery state prediction model is as follows:

[0027] ;

[0028] Among them, represents the battery state prediction value; represents the input feature vector; represents the set of model parameters, including convolutional kernel weight matrix, adaptive weighting coefficient, variational weight coefficient, and neural differential equation parameters; represents the static mapping between the input features and the output results; represents the battery state evolution function; Represents the instantaneous change rate of the input feature over time; Represents the time gradient adjustment factor; Represents the adaptive residual correction term; Represents the time variable.

[0029] Furthermore, the fault monitoring module classifies and identifies the fault feature space based on the anomaly detection algorithm of ensemble learning, extracts the potential fault mode features through the random forest, gradient boosting tree and isolation forest algorithms, and fuses the detection results based on the weighted voting mechanism to generate a comprehensive evaluation of the fault probability and severity level, and outputs multi-level alarm signals for early warning, general faults and serious faults.

[0030] Even further, the formula of the weighted voting mechanism is as follows:

[0031] ;

[0032] ;

[0033] Among them, Represents the final fault probability of the comprehensive evaluation; m represents the number of the ensemble learning algorithm; Represents the initial weight in the weighted voting mechanism; Represents the adaptive weight adjustment coefficient; Represents the temperature feature extracted by the m-th algorithm at time t; Represents the influence degree of the control voltage feature on the comprehensive evaluation; Represents the amplitude of the voltage change; Represents the output feature classification function of the m-th algorithm; Represents the normalized probability distribution function; C represents the confidence factor matrix.

[0034] Furthermore, the safety regulation module includes a hierarchical fault response mechanism: dynamic parameter adjustment and voltage balancing are started during early warning; the charging mode is forced to be switched and the liquid cooling system is activated during general faults; the fuse protection device is triggered and the faulty battery cell is isolated during serious faults, and the complete operation data in the 60 seconds before the fault is recorded at the same time.

[0035] The beneficial effects of the present invention are as follows:

[0036] Through the distributed sensor array in the battery data acquisition module, the present invention can collect the operating parameters of the electric vehicle battery in real time, ensuring the real-time and comprehensiveness of the data. Combining the convolutional neural network in deep learning and the multi-modal data fusion technology, key features are extracted from a large amount of data, improving the accuracy of battery state monitoring. The battery state prediction module, based on the battery state prediction model constructed by the neural network algorithm and differential equations, can predict the multi-dimensional state of the battery, including the remaining capacity, health state, performance decay rate, and future charge and discharge efficiency. Compared with traditional methods, the prediction dimensions are more comprehensive, providing more detailed and reliable references for battery management. The potential fault modes are classified and identified through the anomaly detection algorithm of ensemble learning, realizing the intelligent monitoring of the battery operating state. The system can automatically identify risks such as overheating, overcharging, over-discharging, cell imbalance, and short circuit, and output graded alarm signals, effectively improving the safety of the battery system. By combining the reinforcement learning algorithm and the fuzzy logic control algorithm, the charge and discharge strategy and operating parameters of the battery can be dynamically adjusted according to the alarm signal. Specifically, measures such as adjusting the charging current, equalizing the monomer voltage, actively cooling, or isolating abnormal cells are included, further enhancing the operating safety and stability of the battery. Brief Description of the Drawings

[0037] Figure 1 is a schematic diagram of the modules of a real-time monitoring system for the state of an electric vehicle battery based on deep learning according to the present invention.

[0038] Figure 2 is a schematic flowchart of the operation process of the feature extraction module provided by an embodiment of the present invention.

[0039] Figure 3 is a schematic flowchart of the operation process of the battery state prediction module provided by an embodiment of the present invention. Detailed Description of the Embodiment

[0040] Please refer to Figures 1-3 as shown, the present invention relates to a real-time monitoring system for the state of an electric vehicle battery based on deep learning.

[0041] Embodiment

[0042] A real-time monitoring system for the state of an electric vehicle battery based on deep learning includes a battery data acquisition module, a feature extraction module, a battery state prediction module, a fault monitoring module, and a safety regulation module that are communicatively connected in sequence;

[0043] The battery data acquisition module is used to collect the real-time operating parameters of the battery in real time through a distributed sensor array; the operating parameters include the current, monomer voltage, total voltage output, charge and discharge current, temperature gradient distribution, and internal resistance parameters of the battery; the sensor array includes a high-precision micro-current sensor and a distributed thermocouple array;

[0044] In one embodiment, the battery data acquisition module collects the operating parameters of the battery in real time through a distributed sensor array to ensure the accuracy and real-time nature of the monitoring data.

[0045] The specific types of sensors are as follows:

[0046] High-precision micro-current sensor: Adopting a Hall effect or shunt-type micro-current sensor, it can accurately measure currents from milliamperes to hundreds of amperes, with the error controlled within 0.1%.

[0047] Single-cell voltage sensor: Using a high-precision ADC chip (such as ADS1299), the resolution can reach 24 bits, and the error is less than 0.05%.

[0048] Distributed thermocouple array: Selecting a K-type thermocouple, with a temperature measurement range of -200°C to +1250°C, and cooperating with a high-resolution amplifier (such as MAX31855) for data acquisition to ensure accurate measurement of the temperature gradient.

[0049] Internal resistance measurement: Using the AC injection method, calculating the internal resistance through the excitation current and voltage response, with a frequency range of 1Hz to 1kHz to ensure accurate capture of the internal resistance characteristics at different frequencies.

[0050] The operating parameters of the battery are collected in real time through a distributed sensor array, including:

[0051] Battery current: Collected through a high-precision micro-current sensor, with a measurement range of 0.1mA to 100A, and a sampling frequency of up to 1000Hz to ensure accurate capture of transient currents;

[0052] Single-cell voltage and total voltage output: Detected using a high-precision voltage sensor module, with the measurement accuracy of the single-cell voltage reaching ±0.01V and the detection accuracy of the total voltage being ±0.1V;

[0053] Charge and discharge current: Achieved through a Hall current sensor, with high sensitivity and anti-interference ability;

[0054] Temperature gradient distribution: Using a distributed thermocouple array to cover the battery surface and key nodes inside the battery cells, with the temperature measurement accuracy of each node being ±0.1°C;

[0055] Internal resistance parameter: Measured through the AC impedance method, and the internal resistance of the battery is dynamically detected using a frequency range of 20Hz to 10kHz.

[0056] The feature extraction module is used to extract key features from the operating parameters based on a deep feature extraction model of a convolutional neural network and multi-modal data fusion technology;

[0057] Among them, the operation process of the feature extraction module includes the following steps:

[0058] Preprocess the operating parameters and construct a three-dimensional feature matrix, including a temperature space topology matrix, a voltage frequency domain matrix, and a current-internal resistance time-varying matrix;

[0059] Specifically, the three-dimensional feature matrix is constructed as follows:

[0060] Temperature Space Topology Matrix (TSM): Extract temperature data from the thermocouple array and construct a matrix based on the spatial distribution of the sensors. The temperature matrix is used to characterize the temperature distribution pattern of the battery, including the hot spot location and the overall temperature distribution pattern.

[0061] Voltage Frequency Domain Matrix (VFM): Perform a frequency domain transformation on the voltage signal and extract its response characteristics at different frequencies. The matrix contains information in multiple frequency bands, and each unit represents the voltage amplitude or phase at a certain frequency band.

[0062] Current-Internal Resistance Time-Varying Matrix (CIM): Record the changes in current and internal resistance at different time points, used to capture the impact of current fluctuations on the internal resistance and the dynamic change characteristics of the internal resistance.

[0063] Design a multi-branch convolutional neural network architecture, construct a feature extraction model based on the convolutional neural network, and generate depth feature maps of the temperature distribution pattern, the internal resistance change rate, and the voltage fluctuation spectrum; the multi-branch convolutional neural network architecture specifically includes that the first branch uses an atrous convolutional kernel to extract the long-range dependence features of the temperature space topology, the second branch uses temporal causal convolution to capture the non-linear dynamic characteristics of the internal resistance parameters, and the third branch analyzes the phase-amplitude coupling relationship of the voltage frequency domain matrix through a complex-valued convolutional layer;

[0064] Specifically, the first branch: the temperature branch (atrous convolutional network)

[0065] Adopt a four-level residual structure, each level contains an atrous convolutional layer to capture the long-range dependence of heat conduction between battery cells; spatial pyramid pooling to aggregate temperature gradient features at different scales; finally output a depth feature map representing the temperature distribution pattern, which can effectively identify the areas of abnormal temperature rise inside or on the surface of the battery.

[0066] The second branch: the internal resistance branch (temporal causal convolution)

[0067] Temporal causal convolution is used to process sequence data and extract non-linear dynamic characteristics. Network structure: contains multiple layers of causal convolutional networks, combines time series information to extract the law of internal resistance change with current. Finally, output a depth feature sequence for extracting the internal resistance change rate, used to judge the changes in the electrochemical characteristics inside the battery.

[0068] The third branch: the voltage branch (complex-valued convolutional network)

[0069] Design a complex convolutional layer to process the real and imaginary parts in the frequency domain: the kernel weights are complex numbers, and the phase relationship in the frequency domain is maintained during the convolution process; cascade complex batch normalization and ReLU activation; introduce cross-attention in the frequency domain dimension to calculate the amplitude-phase joint weights; finally, output a deep feature matrix representing the voltage fluctuation spectrum for identifying voltage anomalies at different frequencies.

[0070] Based on the deep feature map, use tensor contraction operations to fuse the multi-branch outputs, perform frequency band optimization on the cross-modal features through the spectral attention mechanism, calculate the energy entropy of each frequency band, and perform weighted screening with the convolutional features to screen out key features.

[0071] Specifically, obtain from the outputs of three feature extraction branches:

[0072] Temperature distribution pattern feature map: representing the distribution structure and relative position relationship of temperature in space.

[0073] Internal resistance change rate feature map: representing the dynamic change trend and change rate of the battery internal resistance over time.

[0074] Voltage fluctuation spectrum feature map: representing the change law and important frequency components of the voltage signal in the frequency domain.

[0075] Use convolution operations to compress the fused feature tensor, reduce the number of feature channels, and at the same time reduce information redundancy. The generated compressed feature matrix retains important information, simplifies the calculation process, and improves the efficiency of feature extraction.

[0076] The optimization of the spectral attention mechanism specifically includes: regarding each channel in the fused feature tensor as an independent frequency band. Calculate the energy of each channel feature, and the energy value reflects the importance of the information in this channel. Normalize the energy of each channel to calculate the uniformity and importance of its distribution. The larger the energy entropy, the more useful information this channel contains. Based on the calculation result of the energy entropy, generate corresponding weight values for each channel. The size of the weight value represents the relative importance of this channel, and high weight values correspond to more significant features. Multiply each feature channel by its corresponding weight to form a weighted feature matrix. The optimized feature matrix highlights the feature information of the key frequency bands more prominently and suppresses the influence of unimportant or interfering information.

[0077] Compare the optimized weighted feature matrix with the convolutional features extracted from the original three branches, and calculate the similarity between them. According to the result of the similarity calculation, eliminate the feature channels with low correlation or unimportance. The features with high correlation are extracted and combined to form a key feature subset.

[0078] The feature set includes:

[0079] Temperature distribution pattern: Extract the spatial distribution of temperature in the battery pack to detect phenomena such as abnormal temperature rise or uneven heat dissipation.

[0080] Rate of change of internal resistance: Extract the characteristics of the change of the battery internal resistance over time to judge the states of battery aging, short circuit or damage, etc.

[0081] Voltage fluctuation spectrum: Extract the fluctuation characteristics of the voltage signal in the frequency domain to identify abnormalities during load changes or battery discharge processes.

[0082] Use deep residual compression to reduce the dimension of the key feature subset, eliminate redundant noise components through a differentiable threshold function, and output key features; the key features include temperature distribution pattern, rate of change of internal resistance, and voltage fluctuation spectrum.

[0083] Furthermore, the formula of the feature extraction model is as follows:

[0084]

[0085] Among them, represents the finally extracted key feature vector; represents the input matrix, that is, the input data of the i-th branch; represents the convolution kernel weight matrix; represents the bias term; represents the activation function; represents the adaptive weighting coefficient; represents the variational weight coefficient, which regulates the fusion intensity of convolution features and variational features; represents the tensor compression operation, which uses CP decomposition or Tucker decomposition to reduce the dimension of the multi-branch output features; represents the variational feature extraction function, which is used to model the probability distribution of each input matrix and extract latent space features; represents the non-linear fusion function, which weights and fuses features of different modalities and optimizes them; represents the number of convolutional layers in each branch.

[0086] The battery state prediction module is used to construct a battery state prediction model based on the key features, using neural network algorithms and differential equations, predict the state evolution of the battery, and generate an output result, and the output result includes the remaining capacity, health state, performance decay rate, and future charge and discharge efficiency of the battery;

[0087] Among them, the operation process of the battery state prediction module includes the following steps:

[0088] Based on the key features, construct a battery state prediction model using neural network algorithms and differential equations;

[0089] Specifically, by combining deep neural networks with time series analysis techniques, a multi-layer network architecture is constructed for battery state prediction. Based on the electrochemical characteristics and internal kinetic mechanisms of the battery, a mathematical model describing the battery state change is constructed. This model is used to describe the evolution process of battery state variables (such as voltage, temperature, SOC, SOH, etc.) at different time points. The core of the model is to express the dynamic change relationship of the battery in the form of differential equations, so that the prediction process not only depends on data, but also includes the simulation of physical processes. By combining the traditional neural network architecture with the differential equation solving process, a neural differential equation model is formed. The introduction of neural differential equations enables the model to dynamically capture the change law of the battery state through a numerical solver. During the training process, an adaptive integration method is used to improve the prediction accuracy and ensure the stability of long-term prediction. By combining network training and differential equation solving, the model can more accurately perform real-time prediction and adjustment of the battery state. The solving process of neural differential equations is integrated with the forward propagation process of deep neural networks. An automatic differentiation tool is used to calculate the gradients of model parameters, enabling the network training process to be automatically adjusted during continuous iteration. The output of the model not only includes the prediction of the current battery state, but also includes the simulation of the evolution trend of the future state.

[0090]

[0090] Perform phase space embedding on the key features, use a variational autoencoder to extract latent variables to represent the initial state of the battery degradation trajectory, and fuse real-time operating condition labels and environmental noise to construct a multi-dimensional state input;

[0091]

[0091] It should be noted that according to the time series of battery state changes, phase space reconstruction is performed to construct a multi-dimensional state variable space. The delay coordinate method is used to embed the time series data into a high-dimensional space to form a state trajectory. The variational autoencoder (VAE) is used to encode the phase space embedding result, compressing the high-dimensional features into latent variables to represent the initial state of the battery degradation trajectory. The decoder reconstructs the battery state sequence according to the latent variables and calculates the reconstruction error to optimize the model. The extracted latent variables are fused with real-time operating condition labels (such as temperature, charge-discharge rate) and environmental noise variables to form a multi-dimensional state input. Through the feature fusion layer, normalization and dimensionality reduction processing are performed to ensure the consistency of the input feature dimensions.

[0092] Based on the multi-dimensional state input, use a neural differential equation to solve the battery state prediction model, use the automatic differentiation method to calculate the gradient, and achieve feature reconstruction and error correction through residual connection to obtain the battery state prediction value;

[0093] ​Specifically, based on the multi-dimensional state input, a neural differential equation solver is used for model deduction to generate predicted values. An adaptive step-size integration method is adopted to improve the solution accuracy and reduce the cumulative error of long-term prediction. The automatic differentiation method is used to calculate the gradient of the battery state predicted value with respect to the input features. Gradient calculation helps to understand the sensitivity of feature changes to the battery state predicted value and provides a basis for model optimization. Through the residual connection mechanism, the input features and the predicted values are weighted and superimposed to achieve feature reconstruction and error compensation. Residual blocks are used for feature correction to avoid gradient vanishing or gradient explosion caused by an increase in model depth.

[0094] Through the adaptive error correction mechanism, the residual between the model predicted value and the measured value is combined to dynamically update the model parameters.

[0095] Specifically, calculate the residual between the model predicted value and the measured value, and analyze the error sources (such as external disturbances or feature loss). Determine the model correction strategy through error analysis, such as adding or reducing network layers or adjusting feature weights. Introduce Kalman filtering or an adaptive mean algorithm to smooth the residual and remove the interference of outliers. Adopt a weighted update mechanism to adjust the model parameters in real time according to the residual weights to ensure the dynamic optimization of prediction accuracy. Use the Adam optimization algorithm for gradient update to dynamically adjust the learning rate and weight decay coefficient. Utilize the rolling window mechanism to continuously track the prediction error and automatically update the model parameters to adapt to different operating conditions.

[0096] Furthermore, the formula of the battery state prediction model is as follows:

[0097]

[0098] Where, represents the battery state predicted value; represents the input feature vector; represents the set of model parameters, including the convolutional kernel weight matrix, adaptive weighting coefficients, variational weighting coefficients, and neural differential equation parameters; represents the static mapping between the input features and the output results; represents the battery state evolution function, which describes the evolution process of the input features based on the neural differential equation; represents the instantaneous change rate of the input features over time; represents the time gradient adjustment factor, which is used to balance the influence of static feature mapping and dynamic feature evolution; when tends to zero, the system pays more attention to the influence of static features; represents the adaptive residual correction term; represents the time variable, indicating the feature evolution process within the time interval [0, t].

[0099] The instantaneous rate of change of the input feature over time is calculated as follows:

[0100] ;

[0101] in, Represents the instantaneous rate of change of the input feature over time, that is, at the time point The dynamic change speed of the following characteristics is used to capture the real-time characteristics of battery state changes, including temperature fluctuations, internal resistance mutations, and voltage responses; Indicates at a point in time The input feature vector at time instant; represents the time increment; Indicates at a point in time The input feature vector at each moment, its value will vary due to the dynamic changes of the battery status.

[0102] The fault monitoring module is used to classify and identify potential fault modes according to the output results using an abnormality detection algorithm based on ensemble learning, and output a graded alarm signal; the fault modes include overheating, overcharging, overdischarging, cell imbalance and short circuit risks;

[0103] Among them, the fault monitoring module classifies and identifies the fault feature space based on the anomaly detection algorithm of ensemble learning, extracts potential fault mode features through random forest, gradient boosting tree and isolation forest algorithms, and fuses the detection results based on the weighted voting mechanism to generate a comprehensive assessment of fault probability and severity level, and outputs multi-level alarm signals for early warning, general fault and severe fault.

[0104] It should be noted that the anomaly detection algorithm of ensemble learning is as follows:

[0105] Random forest model construction: Multiple decision trees are built, and each tree is trained by randomly sampling a subset of the data set. It is used to detect the overall trend of the battery status and general faults (such as overcharging and over-discharging).

[0106] Gradient boosting tree model construction: It uses iterative training, and each training is optimized based on the error of the previous one. It is better at capturing subtle changes in battery status and complex failure modes (such as cell imbalance).

[0107] Isolation forest model construction: Designed specifically for anomaly detection, it determines the degree of anomalies based on the segmentation process of data points. It can quickly locate and identify isolated anomalies, and is particularly suitable for identifying short-circuit risks.

[0108] Divide the labeled dataset into training set, validation set and test set. Use cross-validation methods (such as K-fold cross-validation) to evaluate the accuracy and robustness of each model. Tune the parameters of each model (such as decision tree depth, sample splitting strategy, etc.) to improve the accuracy of classification and recognition.

[0109] Specifically, the design of the weighted voting mechanism: fuse the output results of three models: Random Forest, Gradient Boosting Tree, and Isolation Forest. Assign different weights to each model, and the weights can be adjusted according to the performance on the validation set. Use the weighted voting method to generate the final fault detection result.

[0110] According to the output distribution of different models, calculate the probability value of each type of fault. Compare the historical data with the threshold standard to classify the severity of the fault:

[0111] Warning signal: The fault probability is low, but there are slight abnormal signs.

[0112] General fault: The fault probability is high and needs to be processed in time.

[0113] Severe fault: The fault probability is extremely high, and safety protection measures must be taken immediately.

[0114] Furthermore, the formula of the weighted voting mechanism is as follows:

[0115] ;

[0116] ;

[0117] Where, represents the final fault probability of the comprehensive evaluation (range: 0 to 1); m represents the number of the ensemble learning algorithm (from 1 to 3), corresponding to: m = 1: Random Forest algorithm, m = 2: Gradient Boosting Tree algorithm, m = 3: Isolation Forest algorithm; represents the initial weight in the weighted voting mechanism; represents the adaptive weight adjustment coefficient; represents the temperature feature extracted by the m-th algorithm at time t; represents the influence degree of the control voltage feature on the comprehensive evaluation, used to suppress the influence of excessive voltage changes on weight allocation; represents the amplitude of the voltage change; represents the output feature classification function of the m-th algorithm; represents the normalized probability distribution function, used to normalize the fault detection outputs of each algorithm to ensure that the outputs of all algorithms have the same probability scale; C represents the confidence factor matrix, which is dynamically adjusted according to the historical detection accuracy of each algorithm and the confidence of the current feature.

[0118] The safety control module is used to dynamically adjust the charging and discharging strategies and operating parameters of the battery, and execute safety protection measures, including adjusting the charging current, balancing the cell voltages, starting active cooling or isolating abnormal cells, based on the alarm signal, by using reinforcement learning algorithms and fuzzy logic control algorithms; the safety control module includes a hierarchical fault response mechanism: when a warning is issued, dynamic parameter adjustment and voltage balancing are started; in case of a general fault, the charging mode is forcibly switched and the liquid cooling system is activated; in case of a severe fault, the fuse protection device is triggered and the faulty cell is isolated, and at the same time, the complete operating data for 60 seconds before the fault is recorded.

[0119] Specifically, the implementation steps include: training and optimizing based on deep reinforcement learning algorithms (such as DQN or PPO) according to the real-time data and historical data of the current battery state; through the design of a reward mechanism, enabling the system to adaptively adjust the charging and discharging strategies under different working conditions, and optimizing the usage efficiency and safety of the battery. Combining the fuzzy rule base and membership functions, fuzzifying parameters such as temperature, voltage, and internal resistance; based on the inference system, comprehensively judging the battery state and generating refined control decisions.

[0120] It should be noted that the hierarchical fault response mechanism is as follows:

[0121] Fault characteristics: Slight fluctuations in voltage or current, and the temperature is slightly high but does not exceed the safe range.

[0122] Warning-level response: Dynamically adjust the charging current to reduce the battery stress; activate the voltage balancing system to ensure that the cell voltages are maintained within a reasonable range (such as the difference is less than 0.05V).

[0123] Fault characteristics: Abnormal SOC, rapid decay of SOH, and the temperature exceeds the safety threshold but does not reach the limit.

[0124] General-fault-level response: Forcefully switch to the safe charging mode (such as constant current or constant voltage mode); start the liquid cooling system or air cooling system to reduce the battery temperature; start the current shunt strategy to limit the charging current of the abnormal single cell; use the reinforcement learning algorithm to generate the optimal voltage balancing strategy to avoid faults caused by cell imbalance.

[0125] Fault characteristics: Cell short circuit, extreme temperature exceeding the standard (such as >80°C), severe overcharging or over-discharging.

[0126] Severe-fault-level response: Trigger the fuse protection device to physically isolate the faulty cell; stop all charging and discharging operations, record the complete operating data, including the current, voltage, temperature, and internal resistance data for 60 seconds before the fault; trigger the alarm signal and upload the fault information to the monitoring center or cloud system for subsequent analysis and diagnosis.

[0127] In summary, the present invention collects the current, single-cell voltage, total voltage, charge and discharge current, temperature gradient distribution, and internal resistance parameters of the battery through a distributed sensor array, ensuring the accuracy and real-time nature of data collection. By using high-precision micro-current sensors and a distributed thermocouple array, the operation state of the battery can be effectively monitored, and minute current fluctuations and temperature changes can be accurately captured, ensuring the integrity and accuracy of the monitoring data. The multi-branch architecture based on a convolutional neural network extracts deep features from the temperature, internal resistance, and voltage characteristics, and is optimized through a spectral attention mechanism. The construction of the temperature spatial topology matrix, voltage frequency domain matrix, and current-internal resistance time-varying matrix improves the accuracy of feature extraction, and the generated deep feature map can accurately represent the internal state evolution and abnormal phenomena of the battery.

[0128] The present invention combines a neural network algorithm and a neural differential equation to construct a battery state prediction model. By calculating the gradients of the model parameters through an automatic differentiation tool, the remaining capacity, health state, performance degradation rate, and future charge and discharge efficiency of the battery can be accurately predicted. The adaptive error correction mechanism can optimize the prediction results in real time, improving the robustness and accuracy of the model. An anomaly detection algorithm based on ensemble learning is used to classify and identify potential fault modes through random forests, gradient boosting trees, and isolation forests, and generate hierarchical alarm signals. The weighted voting mechanism combines the output results of multiple models, ensuring the accuracy and comprehensiveness of fault detection, and can effectively identify risks such as overheating, overcharging, over-discharging, cell imbalance, and short circuits. The safety regulation module based on a deep reinforcement learning algorithm and a fuzzy logic control algorithm can adaptively adjust the charge and discharge strategies and operating parameters according to real-time monitoring data. The hierarchical fault response mechanism ensures the safety and efficiency of the battery by dynamically adjusting the current, activating the cooling system, and triggering the fuse protection device, etc., and provides refined protection measures for different levels of faults.

[0129] Through the combination of automatic differentiation, convolutional neural network, and ensemble learning algorithms, the system of the present invention can operate stably under complex working conditions. Through the adaptive step-size integration method and the weighted voting mechanism, the accuracy and stability of battery state monitoring are effectively improved, providing strong support for the intelligent management and safety protection of the battery.

[0130] The above embodiments are only descriptions of the preferred embodiments of the present invention, and do not limit the scope of the present invention. Without departing from the design spirit of the present invention, various deformations and improvements made by those of ordinary skill in the art to the technical solutions of the present invention shall fall within the protection scope determined by the claims of the present invention.

Claims

1. A real-time monitoring system for electric vehicle battery status based on deep learning, characterized in that: It includes a battery data acquisition module, a feature extraction module, a battery status prediction module, a fault monitoring module and a safety control module which are sequentially connected in communication; The battery data acquisition module is used to collect real-time operating parameters of the battery through a distributed sensor array; The feature extraction module is used to extract key features from the operating parameters based on a deep feature extraction model of a convolutional neural network and a multimodal data fusion technology; The battery state prediction module is used to construct a battery state prediction model based on the key features using a neural network algorithm and a differential equation, predict the state evolution of the battery, and generate an output result, wherein the output result includes the remaining capacity, health state, performance decay rate, and future charge and discharge efficiency of the battery; The fault monitoring module is used to classify and identify potential fault modes according to the output results using an abnormality detection algorithm based on ensemble learning, and output a graded alarm signal; the fault modes include overheating, overcharging, overdischarging, cell imbalance and short circuit risks; The safety control module is used to dynamically adjust the battery's charging and discharging strategies and operating parameters based on the alarm signal using a reinforcement learning algorithm and a fuzzy logic control algorithm, and to execute safety protection measures, including adjusting the charging current, balancing the single cell voltage, starting active cooling, or isolating abnormal cells.

2. The electric vehicle battery status real-time monitoring system based on deep learning according to claim 1 is characterized in that: The operating parameters include battery current, single cell voltage, total voltage output, charge and discharge current, temperature gradient distribution and internal resistance parameters.

3. The electric vehicle battery status real-time monitoring system based on deep learning according to claim 1 is characterized in that: The operation process of the feature extraction module includes the following steps: Preprocessing the operating parameters and constructing a three-dimensional characteristic matrix, including a temperature space topological matrix, a voltage frequency domain matrix, and a current-internal resistance time-varying matrix; Design a multi-branch convolutional neural network architecture, build a feature extraction model based on the convolutional neural network, and generate deep feature maps of temperature distribution pattern, internal resistance change rate, and voltage fluctuation spectrum; Based on the deep feature map, the multi-branch outputs are fused using tensor contraction operations, the cross-modal features are optimized by frequency band through the spectral attention mechanism, the energy entropy of each frequency band is calculated, and weighted screening is performed with the convolution features to screen the key features; Deep residual compression is used to reduce the dimension of the key feature subset, and the redundant noise components are eliminated through a differentiable threshold function to output the key features; the key features include temperature distribution pattern, internal resistance change rate and voltage fluctuation spectrum.

4. The electric vehicle battery status real-time monitoring system based on deep learning according to claim 3 is characterized in that: The multi-branch convolutional neural network architecture specifically includes a first branch that uses an expanded convolution kernel to extract the long-range dependence characteristics of the temperature space topology, a second branch that uses temporal causal convolution to capture the nonlinear dynamic characteristics of the internal resistance parameters, and a third branch that analyzes the phase-amplitude coupling relationship of the voltage frequency domain matrix through a complex-valued convolution layer.

5. The electric vehicle battery status real-time monitoring system based on deep learning according to claim 3 is characterized in that: The formula of the feature extraction model is as follows: ； in, Represents the key feature vector finally extracted; Represents the input matrix, that is, the input data of the i-th branch; Represents the convolution kernel weight matrix; represents the bias term; represents the activation function; represents the adaptive weighting coefficient; represents the variational weight coefficient; Represents a tensor compression operation; represents the variational feature extraction function; represents the nonlinear fusion function; Indicates the number of convolutional layers in each branch.

6. The electric vehicle battery status real-time monitoring system based on deep learning according to claim 1 is characterized in that: The operation process of the battery state prediction module includes the following steps: Based on the key features, a battery state prediction model is constructed using a neural network algorithm and differential equations; The key features are embedded in phase space, a variational encoder is used to extract latent variables to represent the initial state of the battery degradation trajectory, and a multi-dimensional state input is constructed by integrating real-time operating condition labels and environmental noise; Based on the multi-dimensional state input, the battery state prediction model is solved by using a neural differential equation, the gradient is calculated by using an automatic differentiation method, and feature reconstruction and error correction are achieved through a residual connection to obtain a battery state prediction value; The model parameters are dynamically updated through an adaptive error correction mechanism, combining the residuals between the model predictions and the measured values.

7. The electric vehicle battery status real-time monitoring system based on deep learning according to claim 6 is characterized in that: The formula of the battery state prediction model is as follows: ； in, Indicates the predicted value of battery status; represents the input feature vector; Represents a set of model parameters, including convolution kernel weight matrix, adaptive weight coefficients, variational weight coefficients, and neural differential equation parameters; Represents a static mapping between input features and output results; represents the battery state evolution function; Represents the instantaneous rate of change of input features over time; represents the time gradient adjustment factor; represents the adaptive residual correction term; Represents a time variable.

8. The electric vehicle battery status real-time monitoring system based on deep learning according to claim 1 is characterized in that: The fault monitoring module classifies and identifies the fault feature space based on the anomaly detection algorithm of ensemble learning, extracts potential fault mode features through random forest, gradient boosting tree and isolation forest algorithms, and fuses the detection results based on a weighted voting mechanism to generate a comprehensive evaluation of fault probability and severity level, and outputs multi-level alarm signals of early warning, general fault and severe fault.

9. The electric vehicle battery status real-time monitoring system based on deep learning according to claim 8 is characterized in that: The formula for the weighted voting mechanism is as follows: ； ； in, represents the final failure probability of comprehensive evaluation; m represents the number of the ensemble learning algorithm; Represents the initial weight in the weighted voting mechanism; Represents the adaptive weight adjustment coefficient; represents the temperature feature extracted by the mth algorithm at time t; Indicates the degree of influence of control voltage characteristics on comprehensive evaluation; Indicates the magnitude of voltage change; represents the output feature classification function of the mth algorithm; represents the normalized probability distribution function; C represents the confidence factor matrix.

10. The electric vehicle battery status real-time monitoring system based on deep learning according to claim 1, characterized in that: The safety control module includes a graded fault response mechanism: dynamic parameter adjustment and voltage balancing are started when an early warning is given; charging mode is forced to switch and the liquid cooling system is activated in case of a general fault; the fuse protection device is triggered and the faulty battery cell is isolated in case of a serious fault, while the complete operating data for 60 seconds before the fault is recorded.

Images