A control method for a charging pile device

Abstract

The present invention relates to the technical field of electric vehicle charging pile regulation. Specifically, the present invention relates to a control method for charging pile equipment. The present invention constructs a multi-modal perception data fusion model, integrates data from electrochemical impedance spectroscopy, ultrasonic, and infrared thermal imaging sensors, deeply mines the battery aging condition, and predicts the aging trend with a deep belief network. According to the deep Q-network algorithm, a dynamic optimal power strategy is generated to improve the charging efficiency and extend the battery service life. For single-cell battery management, relying on the electrochemistry-thermal-mechanical coupling microenvironment model and the finite element and data-driven fusion strategy, the microenvironment of the single-cell battery is closely monitored and finely balanced and regulated, and the temperature and voltage differences between single cells are strictly controlled within a very small range, greatly improving the performance consistency and stability of the battery pack and enhancing the charging efficiency of the single-cell battery of the electric vehicle.

Description

Technical Field

[0001] The present invention relates to the technical field of electric vehicle charging pile regulation, and more specifically, to a control method for charging pile equipment. Background Art

[0002] The regulation of electric vehicle charging piles is an important technology. With the booming development of the electric vehicle industry, charging piles, as key supporting facilities, their control technology has a profound impact on battery performance and lifespan. Traditional charging pile control methods have significant drawbacks when dealing with battery aging problems.

[0003] When traditional charging piles adjust power, they mainly rely on basic parameters such as voltage and current on the battery surface, lacking in-depth insight into the complex electrochemical mechanisms behind battery aging. Battery aging causes changes in the structure of electrode materials and attenuation of electrolyte performance. Traditional methods cannot accurately quantify the power requirements resulting from these changes. This rough power regulation leads to inaccurate charging power for electric vehicle batteries. If it is too high, it is likely to cause irreversible damage such as lithium plating and overheating inside the battery, accelerating aging. If it is too low, it will significantly extend the charging time and reduce the usage efficiency. To solve this technical problem, we thus provide a control method for charging pile equipment. Summary of the Invention

[0004] The purpose of the present invention is to provide a control method for charging pile equipment to solve the problems raised in the above background art.

[0005] To achieve the above purpose, one of the objectives of the present invention is to provide a control method for charging pile equipment, including the following steps:

[0006] S1. Deploy an electrochemical impedance spectroscopy sensor, an ultrasonic sensor array, and an infrared thermal imaging sensor network. The electrochemical impedance spectroscopy sensor is used to obtain data reflecting the internal electrochemical state of the battery. The ultrasonic sensor array is used to detect the deformation, delamination, and electrolyte distribution of the battery plates. The infrared thermal imaging sensor is used to obtain the surface temperature distribution of the battery.

[0007] S2. Build a multi-modal data fusion model based on a deep learning framework, perform spatio-temporal synchronization and feature-level fusion on the data collected by the electrochemical impedance spectroscopy sensor, the data collected by the ultrasonic sensor array, and the data monitored by the infrared thermal imaging sensor, and output a feature vector comprehensively reflecting the battery aging state.

[0008] S3. Collect multi-modal perception data of electric vehicle batteries under different charge and discharge rates, ambient temperatures, and humidity conditions, as well as the corresponding battery remaining life and capacity attenuation curve data. Train a battery aging prediction model based on a deep belief network, predict the aging evolution trend of the battery during the current charging process according to real-time multi-modal perception data, and combine the aging evolution trend output by the battery aging prediction model with the current state of charge, temperature distribution of the battery, and the real-time power capacity of the charging pile. Use the deep Q-network algorithm to generate an optimal power adjustment strategy;

[0009] S4. Establish an electrochemical-thermal-mechanical coupled microenvironment model for each single battery in the battery pack. By arranging a micro temperature, pressure, and voltage sensor network in the battery pack, collect the changes in the microenvironment parameters of each single battery, and use a method combining finite element analysis and data-driven to model and real-time simulate the internal current distribution, temperature field distribution, and stress distribution of the single battery, and adjust the charging current path and power distribution of each single battery by the charging pile.

[0010] As a further improvement of this technical solution, the method for spatio-temporal synchronization of multi-modal data in S2 is as follows:

[0011] Equip the sensor network with a clock synchronization module based on GPS timing. Each sensor is built-in with a GPS receiving chip to receive satellite signals to obtain timestamps, which are used as the reference for the starting time of data acquisition. When all sensors receive a specific synchronization pulse in the GPS signal, synchronously trigger the data acquisition circuit, and record the electrochemical impedance spectroscopy data, ultrasonic data, and infrared thermal imaging data from this moment;

[0012] After completing data acquisition, establish a time deviation correction model, analyze the transmission delay of data from the sensor to the fusion processing center, collect the sending and receiving timestamps of data packets by embedding a time monitoring module at key nodes of the transmission link, calculate the transmission delay, use the least squares method to fit the relationship model between the transmission delay and data traffic and link load, predict the transmission delay of each data packet and compensate and correct it;

[0013] At the same time, regularly collect the standard clock source signal to calibrate the local clock deviation of the sensor, and correct the data acquisition timestamp by linear interpolation method.

[0014] As a further improvement of this technical solution, the method for feature-level fusion of multi-modal data in S2 is as follows:

[0015] Normalize the electrochemical impedance spectroscopy data, ultrasonic data, and infrared thermal imaging data respectively to obtain the energy characteristics of each frequency band, marginal spectrum characteristics, and the temperature mean, variance, and temperature gradient characteristics of each region;

[0016] Construct a deep convolutional neural network fusion architecture, which includes an input layer, a convolutional layer, a pooling layer, and a fully connected layer. The input layer receives the normalized multi-modal feature vectors. The convolutional layer is used to extract features. The pooling layer is used to reduce the dimension of the extracted features. The fully connected layer is used to output a feature vector that comprehensively reflects the probability distribution of the battery aging degree.

[0017] As a further improvement of this technical solution, the sample preprocessing and feature optimization method for training the deep belief network in S3 is as follows:

[0018] Normalize the feature vector that comprehensively reflects the probability distribution of the battery aging degree, the collected multi-modal perception data, and the battery life-related data, calculate the data covariance matrix, solve the eigenvalues and eigenvectors, select the main components according to the eigenvalue magnitudes, and project the high-dimensional data into the low-dimensional space;

[0019] Calculate the mutual information values between each feature and the battery remaining life and capacity decay curve data, and screen the key features according to the mutual information value magnitudes. The mutual information value is used to measure the correlation strength between the feature and the target variable.

[0020] As a further improvement of this technical solution, the environment modeling and policy optimization method of the deep Q-network algorithm in S3 is as follows:

[0021] Abstract the battery charging process into a Markov decision process environment model, define the state space, action space, and transition probability. The state space includes the battery state of charge, temperature distribution, and aging degree. The action space is the charging pile power adjustment strategy. The transition probability is used to describe the probability of transferring from one state to another after performing an action in a certain state. Then, obtain the reward function according to the historical data and the model calculation;

[0022] Divide the deep Q-network into a main network and a target network. The main network is used to select the action evaluation value, and the target network is used to calculate the target Q value. Then, set priorities for the historical multi-modal perception data. The priorities are calculated by a preset fixed formula. Perform replay training on the sample data whose priorities exceed the preset priority threshold to complete the optimization.

[0023] As a further improvement of this technical solution, the multi-model fusion prediction method for the battery aging evolution trend in S3 is as follows:

[0024] Fuse the long short-term memory network and the deep belief network. The long short-term memory network is used to process the multi-modal perception data time series to capture dynamic features, and the deep belief network is used to learn the abstract features of the data and the aging relationship;

[0025] During fusion, the output sequence features of the long short-term memory network and the output of some hidden layers of the deep belief network are concatenated into a joint feature vector and input into the subsequent layer. Based on the prediction of the aging trend by the deep belief network, a support vector machine regression model is introduced to learn and predict the residual correction. One-third of the data is used to train the deep belief network to obtain a preliminary prediction value, and the residual is calculated. Finally, the support vector machine regression model is trained with the residual to obtain a correction function, and the final prediction value is obtained accordingly.

[0026] As a further improvement of this technical solution, the parameter adaptive identification method of the electrochemistry-thermal-mechanical coupling microenvironment model in S4 is as follows:

[0027] The extended Kalman filter algorithm is used to identify the key parameters of the model in real time, and the state equation and measurement equation are constructed. Then, the state estimation and covariance are initialized. The state equation is used to predict the state of the system at the next moment, and the measurement equation is used to describe the relationship between the sensor and the system state;

[0028] At each sampling moment, the prediction step is performed to calculate the prior estimate and covariance, and the update step calculates the Kalman gain according to the measured value to update the state estimate and covariance.

[0029] As a further improvement of this technical solution, the modeling and simulation control method of the finite element analysis and data-driven fusion in S4 is as follows:

[0030] A three-dimensional solid model is constructed according to the battery geometry and material properties, the mesh is divided, and the control equations are set. The control equations include the heat conduction equation, the electrical conduction equation, and the mechanical equilibrium equation;

[0031] Then, the boundary conditions are defined, and the battery microenvironment parameters and charge-discharge performance data under multiple working conditions are collected. The microenvironment parameters include temperature, pressure, and voltage. An artificial neural network is constructed, with the sensor data as the input parameter and the predicted value of the node temperature of the temperature field as the output. The predicted value is used to calibrate the parameters of the finite element model;

[0032] During the charging process, the model input is updated according to the sensor data collected in real time, and the finite element equation and the data-driven model prediction are solved in real time to obtain the dynamic current, temperature, and stress distribution. A multi-objective optimization function is constructed, and the genetic algorithm is used to calculate the optimal charging current path and power distribution strategy to minimize the multi-objective optimization function while satisfying the battery safety constraints.

[0033] Compared with the prior art, the beneficial effects of the present invention are:

[0034] In a control method of a charging pile device, in terms of power regulation, by deploying multi-type sensors to construct a multi-modal perception data fusion system, the aging state of the battery can be keenly captured. The deep neural network model accurately calculates the aging trend, and the deep Q-network algorithm generates a dynamic power strategy, significantly improving the adaptability of the charging power. Based on the collaborative equalization control of the micro-environment of single cells, by coupling the micro-environment model with intelligent regulation means, the temperature and voltage differences between single cells are successfully and strictly controlled within a very small range, ensuring that the charging states of each single cell are uniform, and the consistency of the battery pack is improved to a new level, thus enhancing the charging efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is the overall working flowchart of the present invention. SPECIFIC EMBODIMENTS

[0036] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0037] Please refer to Figure 1 As shown, this embodiment provides a control method for a charging pile device, including the following steps:

[0038] S1. Deploy an electrochemical impedance spectroscopy sensor, an ultrasonic sensor array, and an infrared thermal imaging sensor network. The electrochemical impedance spectroscopy sensor is used to obtain data reflecting the internal electrochemical state of the battery. The ultrasonic sensor array is used to detect the deformation, delamination, and electrolyte distribution of the battery plates. The infrared thermal imaging sensor is used to obtain the surface temperature distribution of the battery.

[0039] S2. Build a multi-modal data fusion model based on a deep learning framework, perform spatio-temporal synchronization and feature-level fusion on the data collected by the electrochemical impedance spectroscopy sensor, the data collected by the ultrasonic sensor array, and the data monitored by the infrared thermal imaging sensor, and output a feature vector comprehensively reflecting the aging state of the battery.

[0040] The method for spatio-temporal synchronization of multi-modal data in S2 is specifically as follows:

[0041] Equip the sensor network with a clock synchronization module based on GPS time synchronization. Each sensor is built-in with a GPS receiving chip to receive satellite signals and obtain high-precision timestamps, which are used as the reference for the starting time of data acquisition, ensuring that all sensors start sampling in the same absolute time coordinate system. The GPS satellite system can provide a globally unified and high-precision time standard, eliminating the clock source differences between sensors, guaranteeing the consistency of the starting time of data from the root cause, avoiding synchronization errors caused by the accumulation of clock deviations, providing a solid time basis for subsequent precise fusion, and greatly improving the accuracy and consistency of the system's real-time monitoring of the battery state.

[0042] At a certain moment , when all sensors receive a specific synchronization pulse in the GPS signal, the data acquisition circuit is synchronously triggered, and the electrochemical impedance spectroscopy data , ultrasonic data and infrared thermal imaging data are recorded from this moment. The collaborative analysis of multi-modal data requires a common time reference. Otherwise, it is impossible to accurately correlate the changes of different physical quantities, providing an accurate time correspondence for subsequent complex fusion analysis and avoiding data mismatching.

[0043] In the post-processing stage of data acquisition, a time deviation correction model is established. According to the characteristics of the data transmission protocol and the network topology structure, analyze the distribution law of the transmission delay of data from the sensor to the fusion processing center. By embedding a time monitoring module at the key nodes of the transmission link, collect the sending and receiving timestamps of data packets and , calculate the transmission delay ; This can accurately quantify the delay situation of data in network transmission. Different transmission paths and network loads will cause fluctuations in transmission delay, which need to be dynamically monitored to provide accurate delay data for subsequent precise compensation, avoid errors in fixed compensation values, ensure that the impact of transmission delay is considered during fusion processing, and then use the least squares method to fit the relationship model between transmission delay and data traffic and link load ; where is the data traffic, is the link load, are the fitting coefficients, predict the transmission delay of each data packet and compensate and correct it. At the same time, for the inherent frequency drift phenomenon of the sensor clock oscillator, regularly collect the standard clock source signal to calibrate the local clock deviation of the sensor , and correct the acquisition data timestamp by linear interpolation method ; where is the acquisition data timestamp, further improving the time synchronization accuracy of multi-modal data to below the microsecond level, ensuring the spatio-temporal consistency of data during fusion processing, avoiding misjudgment of the battery state caused by time misalignment, strengthening the system's ability to capture battery dynamic changes, and optimizing the reliability and timeliness of intelligent charging decisions.

[0044] The method of feature-level fusion of multimodal data in S2 is as follows:

[0045] For electrochemical impedance spectroscopy data, discrete wavelet transform is used to decompose signals in different frequency bands and extract the energy characteristics of each frequency band. ; and linearly normalized to interval, highlighting the relative difference of energy at different frequencies, mining impedance data to hide aging information, and processing ultrasonic data through Hilbert-Huang transform to obtain marginal spectrum characteristics ; Standard normalization is used to highlight the changes in weak fault characteristics, and the infrared thermal imaging data is first segmented to extract the temperature mean, variance and temperature gradient characteristics of each region ; Use robust normalization to balance data distribution, enhance feature stability and noise resistance, lay a standardized data foundation for deep fusion, enable different modal features to participate in fusion at the same dimensional scale, and improve the fairness and accuracy of the fusion effect.

[0046] Construct a deep convolutional neural network fusion architecture. The input layer receives the normalized multimodal feature vector, and the convolution layer uses Convolution kernel arrangement, set 3 convolution layers to extract high-dimensional abstract features, and use the corrected linear unit as the activation function to enhance nonlinear expression. The formula is: ,in The number of convolutional layers is 1, and the pooling layer reduces the dimension and reduces the amount of calculation while retaining the key features. The attention mechanism is introduced before the fully connected layer to adaptively weight the features of each modality according to the importance of the features and enhance the contribution of key aging features. The fully connected layer integrates the features and outputs them through the Softmax function. ;in The output of the fully connected layer corresponds to The aging status score is calculated and the output is a feature vector that comprehensively reflects the probability distribution of the battery aging degree. This realizes the deep fusion and intelligent judgment of multimodal features, accurately describes the battery aging degree and state evolution trend, provides core data basis for subsequent charging strategy optimization, and improves the intelligent and refined level of battery health management of charging piles.

[0047] S3. Collect multimodal perception data of electric vehicle batteries under different charge and discharge rates, ambient temperature and humidity conditions, as well as the corresponding battery remaining life and capacity attenuation curve data. Train a battery aging prediction model based on a deep belief network. Predict the aging evolution trend of the battery during the current charging process based on real-time multimodal perception data. Based on the aging evolution trend output by the battery aging prediction model, combined with the current state of charge, temperature distribution and real-time power capacity of the charging pile, use a deep Q network algorithm to generate the optimal power adjustment strategy.

[0048] The sample preprocessing and feature optimization methods for deep belief network training in S3 are as follows:

[0049] For the eigenvector comprehensively reflecting the probability distribution of battery aging degree and the collected data, min-max normalization is adopted. For the eigenvector elements perform normalization on them, and reduce them to , construct a matrix for all data, and calculate the covariance matrix ; is the number of samples, solve the eigenvalues and eigenvectors , arrange them in descending order according to , select the first k principal components to construct the projection matrix , where the cumulative contribution rate exceeds 85% to determine the value, the low-dimensional data . The different scales of data dimensions affect the training of the deep belief network. Normalization balances the data, and PCA dimensionality reduction reduces redundancy and focuses on the key points, improving the training efficiency and stability, highlighting the key aging features, avoiding model overfitting and computational complexity, and creating a simple and efficient data input for the deep belief network.

[0050] Calculate the mutual information values between each feature and the data of the remaining battery life and capacity attenuation curve, that is:

[0051] ;

[0052] For the eigenvector elements and other data features and the life data , statistically calculate the joint probability and marginal probability calculate , set the threshold , retain > features to construct a new feature set. Since multi-modal data contains redundancy, the mutual information value is the key to accurately measure the correlation between features and life, optimizing the feature set can improve the model learning efficiency and predictive power, focus on the core aging factors and exclude interference, and enhance the ability of the deep belief network to capture the essence of aging.

[0053] The environment modeling and policy optimization method of the deep Q-network algorithm in S3 is as follows:

[0054] Accurately determine the state space, covering the state of charge, temperature distribution and aging degree of the battery, and achieve accurate quantification through multi-sensor data fusion. These factors are closely intertwined and affect the battery performance and charging effect, and none of them can be missing. Such a comprehensive definition constructs a complete battery state perception system for the intelligent agent, enabling it to comprehensively understand the battery working conditions.

[0055] The action space is clearly defined as the power adjustment strategy of the charging pile. According to the rated power of the battery management system and the charging protocol, the power adjustment step and the adjustable range are set. According to the real-time status of the battery, the intelligent agent selects the best action from the discrete action set for implementation. Because power directly affects the charging speed and the battery aging process, it is the core control factor. This setting provides the intelligent agent with precise and clear decision boundaries and operation options, focusing on key decision variables, flexibly and accurately outputting power, accurately balancing charging efficiency and battery life, and reducing energy loss and the risk of battery performance degradation.

[0056] Calculate the transfer probability based on massive historical charging data mining and electrochemical principle driven modeling First, the data is clustered and grouped according to battery type and environmental conditions, and the state transition pattern is mined using a deep neural network. For example, using a deep neural network, the input state ,action Prediction processed by multi-layer perceptron and activation function Probability distribution and accurate transfer probability are the core basis for intelligent agent strategy optimization, which determines the direction of decision-making, gives the intelligent agent powerful predictive ability, helps it to optimize the strategy path by anticipating the consequences of actions, and avoids the risk of adverse state transfer in advance.

[0057] The reward function is constructed by comprehensively considering multiple objectives such as charging time, battery life fluctuation, energy efficiency, etc. ,set up ;in is the rated and actual charging time, is the lifetime attenuation, is the rated life, For energy efficiency, The weights are adjusted dynamically according to the charging scenario and user needs. After repeated experiments and simulation optimization, the reward drives the intelligent agent to explore the optimization strategy, which meets the essential needs of reinforcement learning and guides the intelligent agent to independently seek the best, balance multiple objectives to achieve the optimal decision-making balance, accurately lock the best power adjustment strategy, and improve charging efficiency.

[0058] The deep Q network is subdivided into the main network and the target network. The main network parameters After random initialization, the gradient descent algorithm is used to update the optimization action evaluation value calculation capability and the target network parameters during training. Initialize to The replica is regularly copied and updated from the main network to stabilize the target Q value calculation. This division is to alleviate the problem of overestimation and fluctuation of Q value in traditional Q network training, improve learning stability, and build a robust dual-core architecture for deep Q network strategy learning. The main network explores innovation and the target network navigates stably to avoid strategy learning falling into local optimality, ensuring that strategy exploration is accurate and efficient, steadily upgraded, and accelerating convergence to the optimal power adjustment strategy.

[0059] Set priorities for historical multi-modal perception data and calculate according to the formula where is the temporal difference error of the sample ; is a small positive number to prevent zero, ; is the reward of the sample ; is the discount factor, which measures the importance of the sample, is the state of the sample at the next moment, is the action, is the target network parameter, is the state of the sample at the current moment, is the action selected at the current moment for the sample is the target network parameter After initialization, the copies are different because different samples contribute differently to policy learning. Key samples can accelerate convergence and optimization. This mechanism intelligently focuses on key samples, avoids wasting resources on learning low-quality samples, accelerates convergence to the optimal policy, improves the speed of the model's quick response and precise decision-making in complex and changeable charging scenarios, ensures the efficient and safe charging of the battery, and optimizes the response efficiency of the intelligent control of the charging pile.

[0060] Sample replay training is performed according to the priority. Set the priority threshold . Samples exceeding the threshold are included in the replay buffer. During training, samples are drawn according to the probability to update the main network parameter , and small-batch gradient descent is used for updating ; is the learning rate, is the loss function of the sample . In this process, the deep Q-network with reinforcement learns and remembers key scenarios and optimizes policies, consolidates and improves the network's decision-making accuracy and adaptability to complex key charging scenarios, ensures the rapid and accurate generation of the optimal power policy under extreme or complex working conditions, and guarantees the safe and stable operation of the battery and efficient and energy-saving charging.

[0061] The multi-model fusion prediction method for the battery aging evolution trend in S3 is as follows:

[0062] Build a long short-term memory network, whose core modules include an input gate, a forget gate, an output gate, and a memory unit. Input the time series of multi-modal perception data, and calculate the output sequence features through the gating mechanism to capture dynamic changes. The deep belief network is constructed by stacking restricted Boltzmann machines. The bottom restricted Boltzmann machine receives multi-modal data to learn abstract features and optimizes the weights through the contrastive divergence algorithm. Since the long short-term memory network is good at processing time series data to mine dynamic features, and the deep belief network can abstract feature relationships, integrate the advantages of intelligence, capture battery aging features in all aspects, improve the model's representation ability for complex aging processes, and create a powerful feature extraction engine for battery aging prediction to accurately capture the subtle change trends of aging.

[0063] Concatenate the output sequence features of the long short-term memory network and the output of some hidden layers of the restricted Boltzmann machine along the dimension to form a joint feature vector. Let the output dimension of the long short-term memory network be , and the dimension of the hidden layer of the restricted Boltzmann machine be . Concatenate them into ; where is represented as the restricted Boltzmann machine. Train the subsequent prediction layer with the joint vector, and optimize the weights according to the backpropagation algorithm. This concatenation integrates multi-dimensional features to enrich information, synergistically improves the prediction ability, strengthens the model's understanding of aging features, reduces the limitations of a single network, enables the model to learn a more comprehensive and in-depth aging pattern, and improves the prediction accuracy under complex working conditions.

[0064] Divide the data. Use one-third of the data to train the DBN to obtain the preliminary prediction value , and calculate the residual ; is the true life value. Build a support vector machine regression model, and select the radial basis function as the kernel function ; is a parameter. Train the support vector machine regression model according to the sequential minimal optimization algorithm to learn the residual pattern to obtain the correction function , and the final prediction value ; Due to the errors in the DBN, the SVR can accurately correct the residuals to improve the accuracy. The SVR is a support vector machine regression model, which can effectively correct the prediction errors, enhance the overall prediction reliability, provide highly accurate aging data for the intelligent regulation of charging piles, and improve the charging service quality and battery health management level.

[0065] S4. Establish an electrochemical-thermal-mechanical coupling microenvironment model for each single battery in the battery pack. By arranging a network of micro temperature, pressure, and voltage sensors in the battery pack, collect the changes in the microenvironment parameters of each single battery, and use a method combining finite element analysis and data-driven to model and real-time simulate the internal current distribution, temperature field distribution, and stress distribution of the single battery, and adjust the charging current path and power distribution of the charging pile for each single battery.

[0066] The parameter adaptive identification method for the electrochemistry-thermal-mechanical coupling microenvironment model in S4 is as follows:

[0067] The extended Kalman filter (EKF) algorithm is used to construct the core of the parameter adaptive identification system and establish the state equation ; where is a vector covering key parameters such as the concentration of battery electrochemically active substances, conductivity, and thermal conductivity, as well as state variables such as battery temperature, voltage, and current, is the input vector such as charging current and ambient temperature, is the process noise characterizing system random interference and model uncertainty. The battery operation is affected by multiple factors, and the state equation can quantitatively describe complex causality, ensure that the model dynamically tracks the actual situation, provide a theoretical framework for system dynamic prediction, accurately capture the continuous change trajectory of the battery state, improve the simulation fidelity of the model to the real-time working conditions of the battery, lay a theoretical foundation for precise charging control, and help the charging pile customize the charging strategy according to the "personality" of the battery.

[0068] At the same time, the measurement equation is established ; is the vector of the collected values of the micro temperature, pressure, and voltage sensor network, is the function that maps the state variables to the measured values, is the measurement noise reflecting sensor errors and environmental interference. This equation converts the complex internal state of the battery into observable quantities. The sensor data is the key input to the model. The measurement equation builds a bridge between the two, ensures that the data is effectively incorporated into the model analysis, accurately correlates the theoretical model with the measured data, provides a direct basis for parameter identification, enhances the sensitivity of the model to the real battery state perception, and realizes the coordination of theoretical prediction and actual monitoring.

[0069] Initialize the state estimate and the covariance , and set the initial value of according to the battery model, initial temperature, state of charge, and empirical data or laboratory test results is a diagonal matrix, and the diagonal element values are estimated according to the parameter measurement accuracy and uncertainty, giving an initial state perception and uncertainty quantification description of the model. This reasonable initial value and covariance of the initialization help the EKF algorithm quickly converge to the true parameter value, avoid the initial iteration oscillation or divergence of the algorithm, accelerate the parameter identification process, ensure that the model accurately approximates the true state of the battery from the beginning, and improve the overall computational efficiency and parameter identification timeliness.

[0070] For each sampling time k, the prediction step calculates the prior estimate according to the state equation; this formula represents the prediction of the state at the current time , that is, based on the state estimate at the previous time ​ and the control input , through the state transition function calculate the prior state estimate ; predict the state at the next moment according to the system dynamic model, provide a reference benchmark for updating, and calculate the covariance , the formula calculates the prior state covariance matrix , which is used to quantify the uncertainty of the predicted state. Among them, is the state transition matrix, defined as , that is, the state transition function partial derivative of the state variable, which describes the impact of state changes on error propagation is the state estimate covariance matrix at the previous moment , which describes the uncertainty of the state estimate is the process noise covariance matrix, which describes the random error of the system during the state transition. This prediction is based on the current state and the model to predict the future, prepares for the subsequent measurement data fusion, improves the forward-looking of parameter estimation, reduces the impact of measurement lag, anticipates the battery state change trend in advance, enhances the model's response ability to sudden changes in working conditions, reserves a time margin for the pre-adjustment of the charging strategy, and improves the control stability of the charging pile in a dynamic environment.

[0071] The update step depends on the measured value calculate the Kalman gain , among which, is the observation matrix, which describes the relationship between the measured value and the state is the measurement noise covariance, which describes the uncertainty of the measurement device represents the total uncertainty of the measurement, including the uncertainty of the prediction and the measurement error , the function of this formula is to calculate a weight factor , which is used to balance the prediction and measurement information to make the final estimate as accurate as possible, and then update the state estimate , this formula uses the measured value to correct the prior state estimate to obtain the updated state estimate , among which, represents the error between the measured value and the predicted value (i.e., the observation residual) By adjusting the weight of the observation residual, the state estimate is made as close as possible to the true value.

[0072] Finally, fuse the measured value to correct the prediction deviation and update the covariance , among which, represents the factor for correcting the prior uncertainty, multiplied by​ After that, the prediction error is reduced and the estimation uncertainty is adjusted. This update integrates measurement data to correct the prediction error, optimize the parameter estimation, calibrate the parameters in real-time dynamically, accurately track the changes in battery parameters, improve the adaptability of the model to battery aging and operating condition fluctuations, and ensure that the charging strategy of the charging pile is optimized in real-time according to the battery state.

[0073] The modeling and simulation control method of the finite element analysis and data-driven fusion in S4 is as follows:

[0074] Construct a three-dimensional solid model according to the precise geometric structure and material properties of the battery. When dividing the grid, for the key parts of the battery, such as the electrode-electrolyte interface and the regions with large temperature gradients, use fine tetrahedral grids, and moderately coarsen other parts, taking into account both the calculation accuracy and resource consumption. Set the control equations, the heat conduction equation ; Characterize the heat transfer, where is the material density that determines the size of the heat capacity, the specific heat capacity affects the rate of temperature change, the thermal conductivity affects the speed of heat diffusion, the heat source term covers the electrochemical reaction heat, the electric conduction equation ; Describe the current conduction path, is the electrical conductivity to control the current, is expressed as the electric potential to describe the charge and discharge current distribution of the battery, the mechanical equilibrium equation ; Used to control the internal stress balance of the battery, includes the self-weight of the battery and the internal pressure. Define the boundary conditions, such as the precise current density on the electrode surface is set according to the charging strategy, and the heat convection and heat dissipation conditions of the battery shell are determined according to the ambient temperature and the heat dissipation coefficient. The precise physical model needs to comprehensively consider the structure, characteristics, physical processes and boundary interactions, providing a basis for subsequent accurate simulation and control.

[0075] Comprehensively collect the battery microenvironment parameters and charge and discharge performance data under different charge and discharge rates, ambient temperature and humidity, and battery aging degrees. Construct an artificial neural network. The input layer receives the sensor data and the charge and discharge rate information, which is processed by the hidden layer neurons, and the output layer outputs the predicted values of the node temperatures in the temperature field. Optimize the network weights according to the large historical experience data by the backpropagation algorithm to minimize the prediction error ; is the true temperature value, is the predicted value. Use the network predicted values to calibrate the parameters of the finite element model, such as adjusting the material thermal conductivity according to the predicted temperature deviation. Due to the actual battery manufacturing tolerances, material aging and the complexity of operating conditions, there are deviations in the finite element model, and the data-driven calibration can effectively correct them to ensure that the simulation closely approximates the actual operating conditions.

[0076] During the charging process, the model input is updated at fixed time intervals according to the real-time collected sensor data. The finite element equations and data-driven model predictions are solved in real time, and the dynamic current, temperature, and stress distributions are quickly obtained with the help of a high-performance computing platform. A multi-objective optimization function is constructed, such as ; where is the maximum temperature change, is the maximum current change, is the maximum stress change, is the rated value, is the weight. Using the genetic algorithm, the chromosome encoding of the charging current path and power distribution strategy is randomly generated by initializing the population. The fitness of each individual is evaluated according to the fitness function, and the optimal charging current path and power distribution strategy are calculated through iterative optimization by selection, crossover, and mutation operations, so that is minimized while satisfying the battery safety constraints. In this way, the charging is regulated in real time according to the battery optimization, realizing precise and intelligent regulation throughout the charging process, improving the consistency of battery performance, and enhancing the adaptability of the charging pile to complex working conditions.

[0077] The above control method of a charging pile device integrates the data of electrochemical impedance spectroscopy, ultrasonic, and infrared thermal imaging sensors by constructing a multi-modal perception data fusion model, deeply excavates the battery aging condition, and predicts the aging trend with the help of a deep belief network. According to the deep Q-network algorithm, a dynamic optimal power strategy is generated to improve the charging efficiency and extend the battery service life. For the management of single cells, relying on the electrochemistry-thermal-mechanical coupling microenvironment model and the finite element and data-driven fusion strategy, the microenvironment of single cells is closely monitored and finely balanced and regulated, and the temperature and voltage differences between single cells are strictly controlled within a very small range, greatly improving the consistency and stability of the battery pack performance and the charging efficiency of single cells in electric vehicles.

[0078] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and descriptions in the specification are only preferred examples of the present invention and do not limit the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements fall within the scope of the present invention claimed. The scope of the present invention claimed is defined by the appended claims and their equivalents.

Claims

1. A control method for a charging pile device, characterized in that: The following steps are involved: S1. Deploy electrochemical impedance spectroscopy sensors, ultrasonic sensor arrays, and infrared thermal imaging sensor networks. Electrochemical impedance spectroscopy sensors are used to obtain data reflecting the internal electrochemical state of the battery. Ultrasonic sensor arrays are used to detect battery plate deformation, stratification, and electrolyte distribution. Infrared thermal imaging sensors are used to obtain battery surface temperature distribution. S2. Build a multimodal data fusion model based on a deep learning framework, synchronize the data collected by the electrochemical impedance spectroscopy sensor, the data collected by the ultrasonic sensor array, and the infrared thermal imaging sensor monitoring data in time and space, and fuse them at the feature level, and output a feature vector that comprehensively reflects the battery aging status; S3. Collect multimodal sensing data of electric vehicle batteries under different charge and discharge rates, ambient temperature and humidity conditions and the corresponding battery remaining life and capacity decay curve data, train a battery aging prediction model based on a deep belief network, predict the aging evolution trend of the battery during the current charging process based on real-time multimodal sensing data, and generate the optimal power adjustment strategy using a deep Q network algorithm based on the aging evolution trend output by the battery aging prediction model, combined with the current state of charge, temperature distribution and real-time power capacity of the charging pile; S4. Establish an electrochemical-thermal-mechanical coupling microenvironment model for each single cell in the battery pack. By arranging a network of micro temperature, pressure, and voltage sensors in the battery pack, collect the changes in the microenvironment parameters of each single cell. Use a combination of finite element analysis and data-driven methods to model and simulate the internal current distribution, temperature field distribution, and stress distribution of the single cell in real time, and adjust the charging current path and power distribution of the charging pile to each single cell.

2. A control method for a charging pile device according to claim 1, characterized in that: The method for spatiotemporal synchronization of multimodal data in S2 is specifically as follows: The sensor network is equipped with a clock synchronization module based on GPS timing. Each sensor has a built-in GPS receiving chip to receive satellite signals to obtain a timestamp, which is used as the starting time reference for data collection. When all sensors receive a specific synchronization pulse in the GPS signal, the data acquisition circuit is triggered synchronously to record electrochemical impedance spectroscopy data, ultrasonic data, and infrared thermal imaging data from that moment on. After completing data collection, a time deviation correction model is established to analyze the transmission delay of data from the sensor to the fusion processing center. By embedding a time monitoring module in the key nodes of the transmission link, the sending and receiving timestamps of the data packets are collected, the transmission delay is calculated, and the least squares method is used to fit the relationship model between the transmission delay and the data flow and link load, and the transmission delay of each data packet is predicted and compensated. At the same time, the standard clock source signal is regularly collected to calibrate the local clock deviation of the sensor, and the timestamp of the collected data is corrected by linear interpolation.

3. A control method for a charging pile device according to claim 2, characterized in that: The method for feature-level fusion of multimodal data in S2 is specifically as follows: The electrochemical impedance spectroscopy data, ultrasonic data and infrared thermal imaging data were normalized to obtain the energy characteristics of each frequency band, marginal spectrum characteristics, and the temperature mean, variance and temperature gradient characteristics of each region; A deep convolutional neural network fusion architecture is constructed, which includes an input layer, a convolutional layer, a pooling layer and a fully connected layer. The input layer receives a normalized multimodal feature vector, the convolutional layer is used to extract features, the pooling layer is used to reduce the dimension of the extracted features, and the fully connected layer is used to output a feature vector that comprehensively reflects the probability distribution of battery aging.

4. A control method for a charging pile device according to claim 3, characterized in that: The sample preprocessing and feature optimization method for deep belief network training in S3 is as follows: Normalize the eigenvector that comprehensively reflects the probability distribution of battery aging degree with the collected multimodal sensing data and battery life-related data, calculate the data covariance matrix and solve the eigenvalues ​​and eigenvectors, select the main components according to the eigenvalues, and project the high-dimensional data into a low-dimensional space; The mutual information value between each feature and the remaining battery life and capacity decay curve data is calculated, and the key features are selected according to the mutual information value. The mutual information value is used to measure the correlation between the feature and the target variable.

5. A control method for a charging pile device according to claim 4, characterized in that: The environment modeling and strategy optimization method of the deep Q network algorithm in S3 is as follows: The battery charging process is abstracted into a Markov decision process environment model, and the state space, action space and transition probability are defined. The state space includes the battery state of charge, temperature distribution and aging degree. The action space is the charging pile power adjustment strategy. The transition probability is used to describe the probability of transferring to another state after performing an action in a certain state. Then, the reward function is obtained based on historical data and model calculation. The deep Q network is divided into a main network and a target network. The main network is used to select action evaluation values, and the target network is used to calculate target Q values. Priorities are then set for historical multimodal perception data. The priorities are calculated by a preset fixed formula. Sample data whose priorities exceed the preset priority threshold are replayed for training to complete the optimization.

6. A control method for a charging pile device according to claim 5, characterized in that: The multi-model fusion prediction method for battery aging evolution trend in S3 is as follows: Fusion of long short-term memory network and deep belief network, wherein the long short-term memory network is used to process multimodal perception data time series to capture dynamic features, and the deep belief network is used to learn data abstract features and aging relationships; During fusion, the output sequence features of the long short-term memory network and some hidden layer outputs of the deep belief network are spliced ​​into a joint feature vector and input into the subsequent layer. Then, based on the deep belief network's prediction of the aging trend, the support vector machine regression model is introduced to learn the prediction residual correction. The deep belief network is trained with one-third of the data to obtain the preliminary prediction value, and the residual is calculated. Finally, the support vector machine regression model is trained with the residual to obtain the correction function, and the final prediction value is obtained based on this.

7. The control method of a charging pile device according to claim 1, characterized in that: The parameter adaptive identification method of the electrochemical-thermal-mechanical coupled microenvironment model in S4 is as follows: The extended Kalman filter algorithm is used to identify the key parameters of the model in real time, and the state equation and measurement equation are constructed, and the state estimation and covariance are initialized. The state equation is used to predict the state of the system at the next moment, and the measurement equation is used to describe the relationship between the sensor and the system state. At each sampling moment, the prediction step calculates the prior estimate and covariance, and the update step calculates the Kalman gain based on the measurement value and updates the state estimate and covariance.

8. The control method of a charging pile device according to claim 1, characterized in that: The modeling and simulation control method of the finite element analysis and data-driven integration in S4 is as follows: Constructing a three-dimensional solid model according to the battery geometry and material properties, dividing the grid, and setting control equations, wherein the control equations include heat conduction equations, electrical conduction equations, and mechanical equilibrium equations; Redefine the boundary conditions, collect the battery microenvironment parameters and charge and discharge performance data under multiple working conditions, the microenvironment parameters include temperature, pressure and voltage, build an artificial neural network, the input parameters are sensor data, the output is the temperature prediction value of the temperature field node, the prediction value is used to calibrate the finite element model parameters; During the charging process, the model input is updated based on the real-time collected sensor data, the finite element equations and data-driven model predictions are solved in real time, the dynamic current, temperature, and stress distributions are obtained, a multi-objective optimization function is constructed, and a genetic algorithm is used to calculate the optimal charging current path and power allocation strategy, so that the multi-objective optimization function is minimized while satisfying the battery safety constraints.

Images