Charging main cell dynamic equalization and health degree prediction method and device based on multi-sensor fusion

Abstract

The invention relates to a charging main cell dynamic equalization and health degree prediction method and device based on multi-sensor fusion, and the method comprises the steps: extracting various health factors including the charging time, the voltage curve area, the temperature change rate and the internal resistance through fusing the data of voltage, current and temperature multi-source sensors; inputting the health factor sequence into a bidirectional long-short-term memory neural network optimized by an improved whale optimization algorithm, and estimating the health state value of each battery cell in real time; and finally, based on the health state difference between the battery cells, dynamically calculating the equalization intensity through a proportional-integral algorithm, and generating a self-adaptive equalization instruction to control the active equalization circuit to perform energy redistribution. According to the method, the equalization basis is improved from the traditional SOC difference to the more essential SOH difference, the capacity attenuation knee point and the knee starting point are recognized through the improved model, accurate health state prediction is achieved through deep learning, conversion from passive equalization to adaptive equalization based on the health state is achieved, and attenuation of the battery pack is effectively delayed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of battery management, in particular to a charging main battery cell dynamic balancing and health degree prediction method and device based on multi-sensor fusion. BACKGROUND

[0002] With the rapid development of electric vehicles and large-scale energy storage industries, the performance and safety of lithium ion batteries as the core power source are crucial. In the battery pack, due to differences in manufacturing process, working temperature and use history, the decay rates of each battery cell are not the same, resulting in inconsistency in the health state. This inconsistency can cause a sharp decrease in the available capacity of the battery pack, shorten the cycle life, and even cause overcharging, overheating and other safety hazards during charging and discharging. Therefore, through the battery management system, the precise monitoring and active balancing of the battery cell state are the key to ensuring the overall performance, safety and life of the system.

[0003] At present, the existing battery management technology usually adopts an equalization strategy based on voltage or state of charge. Specifically, the system monitors the real-time terminal voltage or estimated SOC value of each battery cell in the battery pack, and when the voltage difference or SOC difference between the battery cells exceeds a certain fixed threshold, the equalization circuit is started. This equalization strategy is mostly static, for example, by passive discharge resistor to consume the energy of the high-voltage battery cell, or by active equalization circuit to transfer energy from the high-voltage battery cell to the low-voltage battery cell until the voltage or SOC returns to the set equalization interval.

[0004] For the above technology, the static equalization strategy based on fixed threshold lacks predictability and adaptability, cannot dynamically adjust according to the actual health decline trajectory of the battery cell, leads to lagging equalization action, low efficiency, and cannot prevent the accelerated decay problem caused by sudden performance drop of the battery cell, which is not convenient for optimizing the overall life management of the battery pack.

[0005] Based on this, the present application provides a charging main battery cell dynamic balancing and health degree prediction method and device based on multi-sensor fusion. SUMMARY

[0006] In order to improve the problem that the static equalization strategy based on fixed threshold lacks predictability and adaptability, cannot dynamically adjust according to the actual health decline trajectory of the battery cell, leads to lagging equalization action, low efficiency, and cannot prevent the accelerated decay problem caused by sudden performance drop of the battery cell, which is not convenient for optimizing the overall life management of the battery pack, the present application provides a charging main battery cell dynamic balancing and health degree prediction method and device based on multi-sensor fusion.

[0007] In a first aspect, the present application provides a charging main battery cell dynamic balancing and health degree prediction method based on multi-sensor fusion, which adopts the following technical solution: comprising: During the charging process of the battery pack, voltage, current and temperature data of each battery cell are synchronously collected to form multi-source sensor time series data; Based on the multi-source sensor time series data, a set of health factors capable of representing the aging state of the battery cell are calculated and extracted, the health factors including parameters related to charging time, voltage curve change, temperature change and internal resistance; The health factors are input into a pre-trained health state estimation model to estimate and output the current health state value of each battery cell in real time; The current health state values of the battery cells are compared to calculate the health state difference between the battery cells; when the health state difference exceeds a preset first threshold, a dynamic balancing instruction is generated, and the balancing strength of the dynamic balancing instruction is positively correlated with the size of the health state difference; according to the dynamic balancing instruction, the active balancing circuit is controlled to perform energy redistribution on the corresponding battery cell.

[0008] Preferably, the calculation and extraction of a set of health factors capable of representing the aging state of the battery cell based on the multi-source sensor time series data comprises: The multi-source sensor time series data are preprocessed to eliminate abnormal values and missing values in the data; Based on the current data of the charging process, the constant current charging time and the constant voltage charging time are extracted as the first type of health factors; Based on the voltage data of the charging process, the voltage difference in the same time interval and the area surrounded by the charging voltage curve and the time axis are calculated as the second type of health factors; Based on the temperature data, the maximum temperature, average temperature and temperature rise rate during the charging process are extracted as the third type of health factors; Based on the voltage and current data, the internal resistance value of the battery cell is calculated or obtained through electrochemical impedance spectroscopy analysis as the fourth type of health factors.

[0009] Preferably, before the health factors are input into the pre-trained health state estimation model to estimate and output the current health state value of each battery cell in real time, it further comprises: Obtain historical battery aging data, the historical battery aging data including a sequence of health factors collected in multiple charge and discharge cycle periods and an actual battery capacity corresponding to each cycle period measured; based on the actual battery capacity, calculate the health state true value of each cycle period to generate a labeled training data set; A bidirectional long short-term memory neural network is selected as a basic architecture of the health state estimation model, the basic architecture includes a forward propagation layer and a backward propagation layer, and hyperparameters of the basic architecture are initialized, the hyperparameters include a number of hidden layer nodes, a maximum number of training times, an initial learning rate, and an L2 regularization parameter; The initialized hyperparameters of the basic architecture are globally optimized by using an improved whale optimization algorithm to search for an optimal hyperparameter combination: an initial whale population is generated by using a quasi-backward learning strategy to enhance the diversity of the population; the positions of the whales are updated by introducing a nonlinear convergence factor and an adaptive weight strategy according to the fitness values, so as to balance the global exploration and local exploitation capabilities of the algorithm; A random differential mutation operation is performed on the whales to prevent the algorithm from falling into a local optimal solution too early; the position updating and differential mutation operations are repeatedly performed until a termination condition is met, and the optimal hyperparameter combination that minimizes the model prediction error is obtained; The bidirectional long short-term memory neural network is configured and trained using the optimal hyperparameter combination, and the trained model is solidified as the health state estimation model.

[0010] Preferably, the health factor is input into a pre-trained health state estimation model to estimate and output the current health state value of each battery cell in real time, including: The latest health factor collected and extracted from the current battery cell in real time is combined with recent historical data to form a fixed-length health factor input sequence arranged in chronological order; The health factor input sequence is input into the trained health state estimation model, and forward calculation is performed by the bidirectional long short-term memory neural network inside the model, the forward calculation including: the health factor input sequence is processed in chronological order by the forward propagation layer of the model to capture the dependence from history to the present; at the same time, the health factor input sequence is processed in reverse order by the backward propagation layer to capture the context dependence from the future to the present; the hidden states of the forward propagation layer and the backward propagation layer at each time step are spliced and fused to form a comprehensive feature representation containing complete sequence context information; The comprehensive feature representation is transmitted to the fully connected layer and the output layer at the end of the model, and a high-dimensional feature is mapped to a scalar value between 0 and 1 by linear weighting and a nonlinear Sigmoid activation function, and the scalar value is the estimated current health state value of the battery cell.

[0011] Preferably, the latest health factor collected and extracted from the current battery cell in real time is combined with recent historical data to form a fixed-length health factor input sequence arranged in chronological order, including: In the memory of the battery management system, a first-in-first-out data buffer is maintained for each battery cell for storing all health factors corresponding to the battery cell at the last M sampling time points in chronological order, where M is a positive integer greater than a fixed length N; At each new sampling time point, the latest set of health factors collected and extracted in real time at the current time point is written into the data buffer; From the data buffer, health factor data corresponding to the last N time steps in chronological order from early to late is extracted; The N sets of extracted health factors are arranged in chronological order to construct a two-dimensional matrix of N×K, where N also represents the time step of the sequence, and K represents the number of types of health factors extracted at each time step; The data in the two-dimensional matrix is standardized so that the mean of the data in each health factor dimension is 0 and the variance is 1, and the standardized sequence is used as the health factor input sequence input into the health state estimation model.

[0012] Preferably, the fully connected layer and the output layer at the end of the comprehensive feature representation are connected to the output layer by linear weighting and nonlinear Sigmoid activation function, which maps high-dimensional features to a scalar value between 0 and 1, including: The comprehensive feature representation output by the bidirectional long short-term memory neural network is transmitted to the fully connected layer that follows as an input vector; the fully connected layer performs linear transformation on the input vector through a weight matrix and a bias vector, realizes linear weighted sum of all input features, and generates a preliminary linear prediction value; The preliminary linear prediction value output by the fully connected layer is input into the output layer; the output layer is configured with a Sigmoid activation function, which nonlinearly compresses and maps the preliminary linear prediction value to the interval of 0 to 1 to form the original output value of the model; The original output value is de-normalized to convert and restore it from the theoretical output range of 0 to 1 to the actual estimated value of the current state of health in percentage relative to the rated capacity of the battery, obtaining the scalar value.

[0013] Preferably, the current state of health values of the battery cells are compared to calculate the state of health difference between the battery cells; when the state of health difference exceeds a preset first threshold, a dynamic balancing instruction is generated, including: The standard deviation of the current state of health values of all battery cells in the battery pack is calculated as a quantitative indicator of the state of health difference representing the dispersion degree of the state of health between the battery cells; comparing the state of health difference with a preset first threshold value, when the state of health difference is greater than the first threshold value, determining that the balancing needs to be started, and generating a balancing trigger signal; After the balancing trigger, the cell with the highest and lowest current state of health value in the battery pack is identified and marked as the energy source cell and the energy receiving cell, respectively; Based on the relative size of the state of health difference and the first threshold value, the balancing strength coefficient is calculated by a preset proportional-integral algorithm, wherein the larger the state of health difference, the higher the calculated balancing strength coefficient; According to the balancing strength coefficient, the identification of the energy source cell and the energy receiving cell, the dynamic balancing instruction containing the balancing current size, the balancing duration and the balancing path is generated.

[0014] Preferably, the balancing strength coefficient is calculated based on the relative size of the state of health difference and the first threshold value by a preset proportional-integral algorithm, comprising: The state of health difference at the current time is subtracted from the first threshold value to obtain a result defined as the current error value; The integral value at the last time is added to the product of the current error value and the integral of the sampling time interval to obtain the result as the integral value at the current time; The proportional term obtained by multiplying the current error value by a preset proportional gain coefficient is added to the integral term obtained by multiplying the current integral value by an integral gain coefficient to obtain a sum value as the balancing strength coefficient at the current time, and the proportional gain coefficient and the integral gain coefficient are both positive numbers.

[0015] Preferably, the dynamic balancing instruction containing the balancing current size, the balancing duration and the balancing path is generated according to the balancing strength coefficient, the identification of the energy source cell and the energy receiving cell, comprising: The balancing strength coefficient is multiplied by a preset maximum allowed balancing current to obtain an actual balancing current value for this balancing operation, wherein the actual balancing current value does not exceed the preset upper limit of the safety current; Based on the state of health difference between the cells and the actual balancing current value, the balancing duration is determined by querying a preset balancing time mapping table, and the larger the state of health difference, the longer the required balancing duration; According to the identification of the energy source cell and the energy receiving cell, the corresponding switch control signal sequence is generated to configure the conduction state of the power switching device in the active balancing circuit, and the energy transfer path from the energy source cell to the energy receiving cell is established; The actual balancing current value, the balancing duration and the balancing path configuration information are packaged into a complete dynamic balancing instruction and sent to the balancing execution circuit.

[0016] In a second aspect, the application discloses a device for dynamic balancing and health prediction of a charging main battery cell based on multi-sensor fusion, which adopts the following technical scheme and comprises: a data acquisition module, configured to synchronously acquire voltage, current and temperature data of each battery cell during a charging process of a battery pack, and form multi-source sensor time series data; a health characterization module, configured to calculate and extract a group of health factors capable of representing an aging state of the battery cell based on the multi-source sensor time series data, wherein the health factors include parameters related to charging time, voltage curve change, temperature change and internal resistance; a health model module, configured to input the health factors into a pre-trained health state estimation model, and estimate and output a current health state value of each battery cell in real time; a dynamic balancing module, configured to compare the current health state values of the battery cells, calculate a health state difference between the battery cells, generate a dynamic balancing instruction when the health state difference exceeds a preset first threshold, wherein a balancing strength of the dynamic balancing instruction is positively correlated with a size of the health state difference, and control an active balancing circuit to perform energy redistribution on the corresponding battery cell according to the dynamic balancing instruction.

[0017] In a third aspect, the application further provides a control device, which comprises: a memory and a processor, wherein the memory stores a computer program capable of being loaded and executed by the processor, and the computer program is the method for dynamic balancing and health prediction of a charging main battery cell based on multi-sensor fusion.

[0018] In a fourth aspect, the application further provides a computer readable storage medium, which stores a computer program capable of being loaded and executed by a processor, and the computer program is the method for dynamic balancing and health prediction of a charging main battery cell based on multi-sensor fusion.

[0019] In summary, the system in the present application synchronously collects multi-source sensor data such as voltage, current and temperature of each cell during battery charging, and then extracts key health factors that can represent the aging state; subsequently, these health factors are input into the pre-trained health state estimation model to estimate the accurate health state value of each cell in real time online, and the remaining life and aging inflection point (knee point) can be predicted in the long term by combining models such as Transformer. The system does not rely on the traditional voltage difference, but compares the health state difference between cells to trigger balancing, and based on the difference size, uses a proportional-integral algorithm to adaptively calculate the balancing strength coefficient to dynamically generate instructions including balancing current, duration and path; finally, the active balancing circuit is controlled to realize accurate and adaptive transfer of energy from cells with high health state to cells with low health state. Thus, the balancing control is changed from passive, static strategy based on instantaneous external voltage to active, dynamic management based on internal health state, thereby effectively inhibiting the expansion of battery pack inconsistency, delaying the overall capacity decay of the battery pack, improving energy utilization efficiency, and greatly extending the overall service life of the battery pack. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 FIG. 1 is a flowchart of a charging main cell dynamic balancing and health degree prediction method based on multi-sensor fusion.

[0021] Figure 2 FIG. 3 is a circuit diagram of an equalization execution circuit.

[0022] Figure 3 FIG. 4 is a structural block diagram of a charging main cell dynamic balancing and health degree prediction device based on multi-sensor fusion. DETAILED DESCRIPTION

[0023] The following Figure 1 - Figure 3 The present application is further described in detail.

[0024] To solve the technical problems of battery pack performance decay and safety hazards caused by inconsistent health states of cells in the background art, and the lagging and lack of foresight of existing balancing strategies, the present application proposes a charging main cell dynamic balancing and health degree prediction method based on multi-sensor fusion. The core idea of the present application is to: by fusing voltage, current, temperature and other multi-source sensor information, deeply mine health factors strongly associated with battery aging; then, using an optimized and trained deep learning model, accurately estimate and predict the long-term life of each cell health state online; finally, use the prediction results as the basis for dynamic balancing decision-making to generate adaptive balancing instructions that match the actual decay state of the cells.

[0025] And in the traditional battery management system, the dynamic balancing is mainly based on the SOC difference (or directly represented as voltage difference) between each cell. When the system detects that some cells have high SOC and some have low SOC, it will start balancing to let the power peak shaving. In this application, the trigger of the balancing circuit is based on the SOH difference instead of the traditional SOC difference. By only looking at the SOC, the system cannot identify this internal weak link. Based on SOH balancing, the weak cell can be identified and protected early, thereby achieving preventive and health balancing in a true sense.

[0026] Referring to Figure 1 The embodiments of the application at least include steps S10 to S40.

[0027] S10, in the charging process of the battery pack, the voltage, current and temperature data of each cell are synchronously collected to form multi-source sensor time series data; S20, based on the multi-source sensor time series data, a group of health factors capable of representing the aging state of the cell are calculated and extracted, the health factors including parameters related to charging time, voltage curve change, temperature change and internal resistance; S30, input the health factors into the pre-trained health state estimation model to real-time estimate and output the current health state value of each cell; S40, compare the current health state values of each cell to calculate the health state difference between the cells; when the health state difference exceeds a preset first threshold, generate a dynamic balancing instruction, the balancing strength of the dynamic balancing instruction being positively correlated with the size of the health state difference; according to the dynamic balancing instruction, control the active balancing circuit to perform energy redistribution on the corresponding cell.

[0028] Specifically, by synchronously collecting the voltage, current and temperature data of each cell during charging, and calculating and extracting the health factors directly related to aging, deep perception of the battery state is achieved. Then, the health factors are input into the pre-trained model to output accurate health state values in real time. By comparing the health state difference between the cells to trigger balancing and generating a dynamic balancing instruction positively correlated therewith, the active balancing circuit is finally controlled to complete accurate energy redistribution, upgrading the balancing control from passive response based on voltage to active and adaptive management based on internal health state, thereby effectively delaying the attenuation of the battery pack and improving safety and life.

[0029] In some embodiments, step S20 specifically comprises the following steps: pre-processing the multi-source sensor time series data to eliminate outliers and missing values in the data; based on the current data of the charging process, extracting the constant current charging time and the constant voltage charging time as the first type of health factor; based on the voltage data of the charging process, calculating the voltage difference in the same time interval and the area surrounded by the charging voltage curve and the time axis as the second type of health factor; based on the temperature data, extracting the maximum temperature, average temperature and temperature rise rate during the charging process as the third type of health factor; based on the voltage and current data, calculating or obtaining the internal resistance value of the battery cell through electrochemical impedance spectrum analysis as the fourth type of health factor.

[0030] Specifically, first, the collected voltage, current and temperature time series data are pre-processed to effectively eliminate outliers and fill in missing values, providing a high-quality data basis for subsequent analysis; then, health factors strongly related to battery aging are extracted from four aspects: based on the current data, the constant current and constant voltage charging time are extracted as the first type of factor directly reflecting capacity decay; based on the voltage curve, the voltage difference in the same time interval and the integral area of the curve are calculated as the second type of factor representing electrode process changes; from the temperature data, the maximum temperature, average temperature and temperature rise rate are extracted as the third type of factor reflecting thermal management status and internal side reactions; finally, the internal resistance value is obtained through calculation or electrochemical impedance spectrum analysis as the fourth type of key factor reflecting the degradation of the internal structure of the battery. The multi-dimensional and multi-physical quantity health factor system provides comprehensive and reliable feature input for subsequent precise evaluation of the battery health status.

[0031] It should be understood that the health factor is a comprehensive evaluation system representing the aging state of the battery. The four types of health factors constructed in this embodiment are not isolated, but represent the overall battery degradation mechanism from different physical dimensions. The first type of charging time factor directly reflects the degree of battery capacity decay and the change of lithium ion intercalation / deintercalation kinetics; the second type of voltage curve factor deeply describes the evolution of microelectrochemical properties such as electrode material phase transition process and active material loss; the third type of temperature-related factor relates the internal side reaction rate and thermal runaway risk from the perspective of thermal management, and its change is coupled with the growth of internal resistance and capacity decay; the fourth type of internal resistance factor directly reflects the core nature of battery performance degradation caused by deterioration of the conductive network and thickening of the SEI film.

[0032] Further, to accurately identify the key turning points in the battery aging process, i.e. the knee point marking the start of accelerated decay and the knee point marking the sharp performance degradation, this embodiment uses an improved double Bacon-Watts model to fit the battery capacity degradation curve and extract characteristic points, and the mathematical model is expressed as: ; where Y is the battery capacity, x is the number of battery charge and discharge cycles, and are the change cycle points around the mutation, respectively, identified as the knee start and knee point, are the model parameters, is the parameter for controlling the slope mutation, and Z is the residual. The identification result of the model provides accurate input for subsequent prediction based on the key interval.

[0033] In some embodiments, the steps of constructing and training the health state estimation model specifically include: obtaining historical battery aging data, the historical battery aging data including a health factor sequence collected in a plurality of charge and discharge cycle periods and an actual battery capacity measured in each cycle period; calculating a health state true value of each cycle period based on the actual battery capacity to generate a labeled training data set; selecting a bidirectional long short-term memory neural network as a basic architecture of the health state estimation model, the basic architecture including a forward propagation layer and a backward propagation layer, and initializing hyperparameters of the basic architecture, the hyperparameters including a number of hidden layer nodes, a maximum number of training times, an initial learning rate, and an L2 regularization parameter; globally optimizing the initialized hyperparameters of the basic architecture using an improved whale optimization algorithm to search for an optimal hyperparameter combination; generating an initial whale population using a quasi-backward learning strategy to enhance the diversity of the population; updating the position of a whale individual by introducing a nonlinear convergence factor and an adaptive weight strategy according to a fitness value to balance the global exploration and local development capabilities of the algorithm; performing a random difference mutation operation on the whale individual to prevent the algorithm from prematurely falling into a local optimal solution; repeatedly performing the position updating and difference mutation operations until a termination condition is met to obtain the optimal hyperparameter combination that minimizes the prediction error of the model; and configuring and training the bidirectional long short-term memory neural network using the optimal hyperparameter combination to solidify the trained model as the health state estimation model.

[0034] The health state true value is defined based on the battery capacity, and the specific calculation formula is: is the current actual capacity of the battery, is the rated capacity of the battery.

[0035] wherein the calculation formula of the nonlinear convergence factor is: ; a is the convergence factor, t is the current iteration number, is the maximum iteration number, is the adjustment parameter, which in this embodiment takes values of 1 / 2 and 0, respectively.

[0036] The calculation of the adaptive weight and the position updating formula are: ; wherein ω is the adaptive weight, X(t) is the current individual position, Let A be the current optimal individual position, D be the coefficient vector, and D be the random distance vector. This strategy assigns greater weight to individuals in the later stages of the algorithm, enhancing their local exploitation capabilities.

[0037] Specifically, the construction of the health status estimation model first generates a training dataset containing health factor sequences and their corresponding true capacity values ​​using historical battery aging data. Then, a bidirectional long short-term memory network is selected as the basic architecture, and an improved whale optimization algorithm is used for global hyperparameter optimization: quasi-backward learning enhances population diversity, nonlinear convergence factors and adaptive weight balancing are introduced for exploration and development, and random difference mutation is used to prevent premature convergence, ultimately obtaining the optimal hyperparameter combination. Finally, this combination is used to train the bidirectional long short-term memory network and solidify it into the health status estimation model. This ensures that the model can accurately capture the complex nonlinear relationship between battery aging characteristics and health status.

[0038] In some embodiments, step S30 specifically includes the following steps: combining the latest health factors collected and extracted from the current battery cell in real time with recent historical data to form a fixed-length, time-ordered health factor input sequence; inputting the health factor input sequence into the trained health status estimation model, and performing forward computation through the bidirectional long short-term memory neural network inside the model. The forward computation includes: processing the health factor input sequence in time order through the forward propagation layer of the model to capture the historical-to-current dependency relationship, the hidden state sequence of which is... Simultaneously, through a backpropagation layer, the health factor input sequence is processed in reverse order to capture the contextual dependencies from the future to the present, and its hidden state sequence is... The hidden states of the forward and backward propagation layers at each time step are concatenated and fused to form a comprehensive feature representation containing complete sequence context information. Representing the comprehensive features The fully connected layers and output layer at the end of the model use a linearly weighted and non-linear sigmoid activation function to map high-dimensional features into a scalar value between 0 and 1: ; It is the Sigmoid activation function. and This represents the weights and biases of the prediction layer. This scalar value is the estimated current health status of the battery cell.

[0039] Further, the latest health factors collected and extracted from the current battery cell in real time are combined with the recent historical data to form a fixed-length health factor input sequence arranged in chronological order, which specifically includes: maintaining a first-in-first-out data buffer for each battery cell in the memory of the battery management system, for storing all health factors corresponding to the battery cell at the last M sampling time points in chronological order, where M is a positive integer greater than the fixed length N; at each new sampling time, write the latest set of health factors collected and extracted in real time at the current time into the data buffer; extract the health factor data corresponding to the last N time steps from the data buffer in chronological order from early to late; arrange the extracted N sets of health factors in chronological order to construct a two-dimensional matrix of N x K, where N also represents the time step of the sequence, and K represents the number of types of health factors extracted at each time step; standardize the data in the two-dimensional matrix so that the data on each health factor dimension has a mean of 0 and a variance of 1, and use the standardized sequence as the health factor input sequence input to the health state estimation model.

[0040] Further, the comprehensive feature representation is delivered to the fully connected layer and the output layer at the end of the model, and through linear weighting and nonlinear Sigmoid activation function, the high-dimensional feature is mapped to a scalar value between 0 and 1, which specifically includes: delivering the comprehensive feature representation output by the bidirectional long short-term memory neural network as an input vector to the fully connected layer that follows; the fully connected layer performs linear transformation on the input vector through a weight matrix and a bias vector, realizes linear weighted sum of all input features, and generates a preliminary linear prediction value; inputting the preliminary linear prediction value output by the fully connected layer to the output layer; the output layer is configured with a Sigmoid activation function, which nonlinearly compresses and maps the preliminary linear prediction value to the interval of 0 to 1 through the Sigmoid activation function to form the native output value of the model; performing inverse standardization on the native output value to convert and restore it from the theoretical output range of 0 to 1 to the actual estimated value of the current health state in percentage form relative to the rated capacity of the battery, to obtain a scalar value.

[0041] Specifically, the health factor input sequence of fixed length is constructed by the first-in-first-out data buffer, and is input into the trained health state estimation model after standardization processing. The model uses bidirectional long short-term memory network to simultaneously perform forward and backward processing, respectively capturing the dependence from history to current and from future to current, and splicing and fusing the bidirectional hidden states to form a comprehensive feature representation. After linear weighting by the full connection layer, the feature representation is mapped to the interval of 0-1 by the Sigmoid activation function, and then the health state value based on the percentage of the rated capacity of the battery is finally output after inverse standardization processing. Thus, the end-to-end conversion from multi-dimensional time series data to accurate health state value is realized, greatly increasing the reliability of subsequent dynamic balancing decision.

[0042] In some embodiments, step S40 specifically comprises the following steps: calculating the standard deviation of the current health state values of all battery cells in the battery pack as a quantitative indicator of the health state difference representing the degree of dispersion of the health state among the battery cells; comparing the health state difference with a preset first threshold value, and determining that balancing needs to be started when the health state difference is greater than the first threshold value, and generating a balancing trigger signal; after the balancing trigger, identifying the battery cell with the highest and lowest current health state value in the battery pack, and marking them as the energy source cell and the energy receiving cell, respectively; based on the relative size of the health state difference and the first threshold value, calculating the balancing strength coefficient by a preset proportional-integral algorithm, wherein the larger the health state difference, the higher the balancing strength coefficient calculated; generating a dynamic balancing instruction containing the balancing current size, the balancing duration and the balancing path according to the balancing strength coefficient, the identification of the energy source cell and the energy receiving cell.

[0043] Further, based on the relative size of the health state difference and the first threshold value, the balancing strength coefficient is calculated by a preset proportional-integral algorithm, specifically including: subtracting the first threshold value from the current health state difference to obtain a current error value; adding the integral value of the last time and the product of the current error value and the sampling time interval to obtain the integral value of the current time; adding the proportional term obtained by multiplying the current error value by a preset proportional gain coefficient to the integral term obtained by multiplying the current integral value by an integral gain coefficient to obtain the balancing strength coefficient of the current time, the proportional gain coefficient and the integral gain coefficient are both positive numbers.

[0044] Further, according to the balancing strength coefficient, the energy source battery cell and the energy receiving battery cell, a dynamic balancing instruction containing the balancing current size, the balancing duration and the balancing path is generated, including: multiplying the balancing strength coefficient by the preset maximum allowed balancing current to obtain the actual balancing current value of the current balancing operation, wherein the actual balancing current value does not exceed the preset upper limit of the safety current; based on the state of health difference between the battery cells and the actual balancing current value, the balancing duration is determined by querying the preset balancing time mapping table, the greater the state of health difference, the longer the required balancing duration; according to the identification of the energy source battery cell and the energy receiving battery cell, the corresponding switch control signal sequence is generated to configure the conduction state of the power switch device in the active balancing circuit, and the energy transfer path from the energy source battery cell to the energy receiving battery cell is established; the actual balancing current value, the balancing duration and the balancing path configuration information are packaged into a complete dynamic balancing instruction and sent to the balancing execution circuit.

[0045] wherein the calculation process of the proportional-integral algorithm is defined by the following formula: ; ; ; wherein, is the error at the current time, is the state of health difference, is the first threshold value; is the integral term at the current time, is the integral term at the last time, is the sampling time interval; is the balancing strength coefficient at the current time, is the proportional gain coefficient, is the integral gain coefficient.

[0046] Specifically, first, the battery pack inconsistency degree is quantified by calculating the standard deviation of all battery cell state of health values, and when the value exceeds the preset threshold value, the balancing process is triggered; the system then identifies the battery cell with the highest state of health and the battery cell with the lowest state of health as the energy source and the receiving end respectively, and based on the relative size of the state of health difference and the threshold value, the balancing strength coefficient is dynamically calculated through the proportional-integral control algorithm: the algorithm realizes adaptive adjustment by calculating the weighted sum (proportional term and integral term) of the current error value and the historical error cumulative value, to ensure that the greater the difference, the higher the balancing strength; finally, according to the coefficient, specific balancing parameters are determined, including multiplying the coefficient by the maximum allowed current to obtain the actual balancing current, determining the duration by querying the time mapping table, and generating the corresponding switch control sequence to establish the energy transfer path, and finally packaging these parameters into a complete instruction and sending it to the execution circuit, so as to realize precise adaptive balancing control based on the actual state of health of the battery.

[0047] Further, with reference to Figure 2 , Figure 2 is a dynamic balancing circuit diagram of an embodiment of the present application. The active balancing circuit adopts a modular distributed architecture, and its core is composed of multiple identical balancing sub-modules. Each balancing sub-module in the dashed box is responsible for balancing the adjacent two batteries. For a battery pack containing M batteries, a total of M-1 balancing sub-modules are needed. Among them, L1 is a power inductor, which plays a key role in converting electrical energy and magnetic energy; R1 is a demagnetization resistor, which ensures the safe and reliable operation of the circuit; Q1 and Q2 are power switching devices that receive balancing instructions from the controller and accurately control the start and stop of the balancing process. The advantage of this circuit design is that the balancing bypass is completely independent of the main charging and discharging loop of the battery pack, so that the balancing operation is not affected by the main loop current, and the battery pack can continue to balance in any working state including static, charging, and discharging. In addition, due to the use of a distributed architecture, multiple non-adjacent battery pairs can perform balancing operations simultaneously, which greatly shortens the overall balancing time and improves system efficiency. As the final execution unit of the dynamic balancing instruction of the present application, this circuit, together with the aforementioned health state estimation model and dynamic balancing decision engine, forms a complete system, ensuring that accurate energy management based on the health state of the battery cell can be efficiently implemented.

[0048] The implementation principle of an embodiment of the present application, a charging main cell dynamic balancing and health degree prediction method based on multi-sensor fusion, is as follows: the system synchronously collects multi-source sensor data such as voltage, current, and temperature of each battery cell during battery charging, and then extracts key health factors that can represent the aging state from the data; subsequently, these health factors are input into a pre-trained health state estimation model to estimate the accurate health state value of each battery cell in real time online, and can be combined with models such as Transformer for long-term prediction of remaining life and aging inflection point (knee point). Instead of relying on traditional voltage difference, the system triggers balancing by comparing the health state difference between battery cells, and based on this difference, uses a proportional-integral algorithm to adaptively calculate a balancing strength coefficient to dynamically generate instructions containing balancing current, duration, and path; finally, the active balancing circuit is controlled to achieve accurate and adaptive transfer of energy from battery cells with high health state to battery cells with low health state. Thus, the balancing control is changed from passive, static strategy based on instantaneous external voltage to active, dynamic management based on internal health state, effectively suppressing the expansion of battery pack inconsistency, delaying the capacity decay of the overall battery pack, improving energy utilization efficiency, and greatly extending the overall service life of the battery pack.

[0049] Figure 1 is a flowchart of a charging main cell dynamic balancing and health degree prediction method based on multi-sensor fusion in an embodiment. It should be understood that, although Figure 1The steps in the flowchart are displayed in sequence according to the indication of the arrows, but the steps are not necessarily executed in the order indicated by the arrows; unless otherwise specified herein, the execution of the steps is not strictly limited in sequence, and the steps can be executed in other orders; and Figure 1 At least part of the steps in the flowchart can include multiple sub-steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution order of the sub-steps or stages is not necessarily sequential, but can be alternately executed with at least part of other steps or sub-steps or stages of other steps.

[0050] Based on the same technical concept, referring to Figure 3 The embodiment of the application also provides a charging main electric core dynamic balancing and health degree prediction device based on multi-sensor fusion, which adopts the following technical scheme, and the device comprises: A data acquisition module is configured to synchronously acquire voltage, current and temperature data of each electric core during a charging process of a battery pack, and to form multi-source sensor time series data. A health representation module is configured to calculate and extract a group of health factors capable of representing an aging state of the electric core based on the multi-source sensor time series data, wherein the health factors include parameters related to charging time, voltage curve change, temperature change and internal resistance. A health model module is configured to input the health factors into a pre-trained health state estimation model, and to estimate and output a current health state value of each electric core in real time. A dynamic balancing module is configured to compare the current health state values of the electric cores, to calculate a health state difference between the electric cores, to generate a dynamic balancing instruction when the health state difference exceeds a preset first threshold, and to positively correlate the balancing strength of the dynamic balancing instruction with the size of the health state difference, and to control a positive balancing circuit to perform energy redistribution on the corresponding electric core according to the dynamic balancing instruction.

[0051] In some embodiments, the health representation module is specifically configured to pre-process the multi-source sensor time series data to eliminate abnormal values and missing values in the data. Based on current data of the charging process, constant current charging time and constant voltage charging time are extracted as first-type health factors. Based on voltage data of the charging process, voltage difference values in the same time interval and areas surrounded by the charging voltage curve and the time axis are calculated as second-type health factors. Based on temperature data, the maximum temperature, average temperature and temperature rising rate in the charging process are extracted as third-type health factors. Based on voltage and current data, the internal resistance value of the electric core is calculated or obtained through electrochemical impedance spectrum analysis as fourth-type health factors.

[0052] In some embodiments, the health model module is specifically configured to obtain historical battery aging data, the historical battery aging data including a health factor sequence collected in a plurality of charge-discharge cycle periods and a measured actual battery capacity corresponding to each cycle period; calculate a health state true value of each cycle period based on the actual battery capacity, and generate a labeled training data set; A bidirectional long short-term memory neural network is selected as a basic architecture of the health state estimation model, the basic architecture including a forward propagation layer and a backward propagation layer, and hyperparameters of the basic architecture are initialized, the hyperparameters including a number of hidden layer nodes, a maximum number of training times, an initial learning rate, and an L2 regularization parameter; An improved whale optimization algorithm is used to globally optimize the initialized hyperparameters of the basic architecture to search for an optimal hyperparameter combination: an initial whale population is generated by using a quasi-back propagation learning strategy to enhance the diversity of the population; according to a fitness value, a whale individual position is updated by introducing a nonlinear convergence factor and an adaptive weight strategy to balance the global exploration and local exploitation capabilities of the algorithm; A random differential mutation operation is performed on the whale individual to prevent the algorithm from falling into a local optimal solution too early; the position updating and differential mutation operations are repeatedly performed until a termination condition is met, and the optimal hyperparameter combination that minimizes the model prediction error is obtained; The bidirectional long short-term memory neural network is configured and trained using the optimal hyperparameter combination, and the trained model is solidified as the health state estimation model.

[0053] In some embodiments, the health model module is specifically configured to combine the latest health factor collected and extracted from the current cell in real time with recent historical data to form a fixed-length health factor input sequence arranged in chronological order; The health factor input sequence is input into the trained health state estimation model, and a forward calculation is performed by the bidirectional long short-term memory neural network inside the model, the forward calculation including: processing the health factor input sequence in chronological order by the forward propagation layer of the model to capture the historical-to-current dependency relationship; at the same time, processing the health factor input sequence in reverse order by the backward propagation layer to capture the future-to-current context dependency relationship; and splicing and fusing the hidden states of the forward propagation layer and the backward propagation layer at each time step to form a comprehensive feature representation containing complete sequence context information; The comprehensive feature representation is transmitted to a fully connected layer and an output layer at the end of the model, and a high-dimensional feature is mapped to a scalar value between 0 and 1 by linear weighting and a nonlinear Sigmoid activation function, the scalar value being the estimated current health state value of the cell.

[0054] In some embodiments, the health model module is specifically configured to maintain a first-in-first-out data buffer for each battery cell in the memory of the battery management system, for storing all health factors of the battery cell corresponding to the last M sampling time points in chronological order, where M is a positive integer greater than a fixed length N; At each new sampling time point, the latest set of health factors collected and extracted in real time at the current time point is written into the data buffer; From the data buffer, the health factor data corresponding to the last N time steps in chronological order from early to late is extracted; The N sets of extracted health factors are arranged in chronological order to construct a two-dimensional matrix of N×K, where N also represents the time step of the sequence, and K represents the number of types of health factors extracted at each time step. The data in the two-dimensional matrix is standardized so that the mean of the data in each health factor dimension is 0 and the variance is 1, and the standardized sequence is used as the health factor input sequence of the health state estimation model.

[0055] In some embodiments, the health model module is specifically configured to pass the comprehensive feature representation output by the bidirectional long short-term memory neural network to the fully connected layer that follows as an input vector; the fully connected layer performs linear transformation on the input vector through a weight matrix and a bias vector, realizes linear weighted summation of all input features, and generates a preliminary linear prediction value; The preliminary linear prediction value output by the fully connected layer is input to the output layer; the output layer is configured with a Sigmoid activation function, which non-linearly compresses and maps the preliminary linear prediction value to the interval of 0 to 1 through the Sigmoid activation function, forming the native output value of the model; The native output value is de-normalized to convert and restore it from the theoretical output range of 0 to 1 to the actual estimated value of the current health state in percentage form relative to the rated capacity of the battery, obtaining a scalar value.

[0056] In some embodiments, the dynamic balancing module is specifically configured to calculate the standard deviation of the current health state values of all battery cells in the battery pack as a quantitative indicator of the health state difference representing the degree of dispersion of the health states among the battery cells; The health state difference is compared with a first threshold value, and when the health state difference is greater than the first threshold value, it is determined that balancing needs to be started, and a balancing trigger signal is generated; After the balancing trigger, the battery cell with the highest and lowest current health state value in the battery pack is identified and marked as the energy source cell and the energy receiving cell, respectively; The balancing strength coefficient is calculated by a preset proportional-integral algorithm based on a relative size of the state-of-health difference and the first threshold value, wherein the larger the state-of-health difference is, the higher the balancing strength coefficient calculated is; According to the balancing strength coefficient, the identifiers of the energy source battery cell and the energy receiving battery cell, a dynamic balancing instruction containing a balancing current size, a balancing duration and a balancing path is generated.

[0057] In some embodiments, the dynamic balancing module is specifically configured to subtract the first threshold value from the state-of-health difference at the current time, and a result obtained is defined as a current error value; The integral value at the last time is added to a product of the current error value and a sampling time interval, and a result obtained is used as the integral value at the current time; A proportional term obtained by multiplying the current error value by a preset proportional gain coefficient is added to an integral term obtained by multiplying the integral value at the current time by an integral gain coefficient, and a sum value obtained is used as the balancing strength coefficient at the current time, wherein the proportional gain coefficient and the integral gain coefficient are both positive numbers.

[0058] In some embodiments, the dynamic balancing module is specifically configured to multiply the balancing strength coefficient by a preset maximum allowed balancing current to obtain an actual balancing current value of the current balancing operation, wherein the actual balancing current value does not exceed a preset upper limit of a safety current; Based on the state-of-health difference between the battery cells and the actual balancing current value, a balancing duration is determined by querying a preset balancing time mapping table, wherein the larger the state-of-health difference is, the longer the balancing duration required is; According to the identifiers of the energy source battery cell and the energy receiving battery cell, a corresponding switch control signal sequence is generated to configure a conduction state of a power switching device in the active balancing circuit, and an energy transfer path from the energy source battery cell to the energy receiving battery cell is established; The actual balancing current value, the balancing duration and the balancing path configuration information are encapsulated into a complete dynamic balancing instruction, and the complete dynamic balancing instruction is sent to a balancing execution circuit.

[0059] Embodiments of the present application also disclose a control device.

[0060] Specifically, the control device includes a memory and a processor, and the memory stores a computer program capable of being loaded and executed by the processor to implement the above-mentioned method for predicting a state-of-health of a charging main battery cell based on multi-sensor fusion.

[0061] Embodiments of the present application also disclose a computer readable storage medium.

[0062] Specifically, the computer readable storage medium, which stores the computer program capable of being loaded and executed by the processor, such as the above-mentioned charging main electric core dynamic balancing and health prediction method based on multi-sensor fusion, for example, includes: U disk, mobile hard disk, read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), various storage program codes such as magnetic disk or optical disk.

[0063] The above are preferred embodiments of the present application, not limited to the protection scope of the present application, therefore: all equivalent changes made according to the structure, shape, principle of the present application should be covered within the protection scope of the present application.

Claims

1. A method for dynamic balancing and health prediction of charging main cells based on multi-sensor fusion, characterized in that, include: During the charging process of the battery pack, the voltage, current and temperature data of each cell are collected simultaneously to form multi-source sensor time-series data; Based on the multi-source sensor time-series data, a set of health factors that can characterize the aging state of the battery cell are calculated and extracted. The health factors include parameters related to charging time, voltage curve changes, temperature changes, and internal resistance. The health factors are input into a pre-trained health status estimation model to estimate and output the current health status value of each cell in real time. The current health status values ​​of each cell are compared to calculate the health status difference between cells. When the health status difference exceeds a preset first threshold, a dynamic balancing command is generated. The balancing strength of the dynamic balancing command is positively correlated with the magnitude of the health status difference. According to the dynamic balancing command, the active balancing circuit is controlled to redistribute energy to the corresponding cells.

2. The method for dynamic balancing and health prediction of charging main cells based on multi-sensor fusion according to claim 1, characterized in that, Based on the multi-source sensor time-series data, a set of health factors that can characterize the aging state of the battery cell are calculated and extracted, including: The time-series data from the multi-source sensors are preprocessed to eliminate outliers and missing values. Based on the current data during the charging process, constant current charging time and constant voltage charging time are extracted as the first type of health factor. Based on the voltage data during the charging process, the voltage difference within the same time interval and the area enclosed by the charging voltage curve and the time axis are calculated as the second type of health factor. Based on temperature data, the highest temperature, average temperature and temperature rise rate during the charging process are extracted as the third type of health factor. Based on voltage and current data, the internal resistance of the battery cell is calculated or obtained through electrochemical impedance spectroscopy analysis, and is used as a fourth type of health factor.

3. The method for dynamic balancing and health prediction of charging main cells based on multi-sensor fusion according to claim 1, characterized in that, Before inputting the health factors into the pre-trained health status estimation model to estimate and output the current health status value of each cell in real time, the method further includes: Acquire historical battery aging data, which includes health factor sequences collected within multiple charge-discharge cycles and the measured actual battery capacity corresponding to each cycle; calculate the true health status value for each cycle based on the actual battery capacity, and generate a labeled training dataset; A bidirectional long short-term memory neural network is selected as the basic architecture of the health status estimation model. The basic architecture includes a forward propagation layer and a backpropagation layer. The hyperparameters of the basic architecture are initialized, including the number of hidden layer nodes, the maximum number of training iterations, the initial learning rate, and the L2 regularization parameter. An improved whale optimization algorithm is used to globally optimize the initial hyperparameters of the infrastructure and search for the optimal combination of hyperparameters: an initial whale population is generated using a quasi-backward learning strategy to enhance population diversity; based on the fitness value, the position of individual whales is updated by introducing a nonlinear convergence factor and an adaptive weight strategy to balance the algorithm's global exploration and local exploitation capabilities. Perform random differential mutation operation on individual whales to prevent the algorithm from getting trapped in local optima too early; repeat position update and differential mutation operation until the termination condition is met to obtain the optimal hyperparameter combination that minimizes the model's prediction error; The optimal hyperparameter combination is used to configure and train the bidirectional long short-term memory neural network, and the trained model is solidified into the health status estimation model.

4. The method for dynamic balancing and health prediction of charging main cells based on multi-sensor fusion according to claim 3, characterized in that, The step of inputting the health factors into a pre-trained health status estimation model to estimate and output the current health status value of each cell in real time includes: The latest health factors collected and extracted from the current battery cell in real time will be combined with recent historical data to form a fixed-length health factor input sequence arranged in chronological order. The health factor input sequence is input into the trained health status estimation model, and forward computation is performed through the bidirectional long short-term memory neural network within the model. The forward computation includes: processing the health factor input sequence in chronological order through the model's forward propagation layer to capture the historical-to-current dependencies; simultaneously, processing the health factor input sequence in reverse order through the back propagation layer to capture the future-to-current contextual dependencies; and concatenating and fusing the hidden states of the forward and back propagation layers at each time step to form a comprehensive feature representation containing complete sequence contextual information. The comprehensive feature representation is passed to the fully connected layer and output layer at the end of the model. Through linear weighting and nonlinear Sigmoid activation function, the high-dimensional features are mapped to a scalar value between 0 and 1. The scalar value is the estimated current health status value of the battery cell.

5. The method for dynamic balancing and health prediction of charging main cells based on multi-sensor fusion according to claim 4, characterized in that, The latest health factors, collected and extracted from the current battery cell in real time, are combined with recent historical data to form a fixed-length, chronologically ordered health factor input sequence, including: In the memory of the battery management system, a first-in-first-out data buffer is maintained for each cell to cyclically store all health factors corresponding to the cell at the most recent M sampling times in chronological order, where M is a positive integer greater than a fixed length N; At each new sampling moment, the latest set of health factors collected and extracted in real time at the current moment is written into the data buffer; From the data buffer, extract the health factor data corresponding to the most recent N time steps in chronological order from morning to evening; The extracted N groups of health factors are arranged in chronological order to form an N×K two-dimensional matrix, where N represents the time step of the sequence and K represents the number of types of health factors extracted at each time step. The data in the two-dimensional matrix are standardized so that the mean of the data in each health factor dimension is 0 and the variance is 1. The standardized sequence is then used as the health factor input sequence for the health status estimation model.

6. The method for dynamic balancing and health prediction of charging main cells based on multi-sensor fusion according to claim 6, characterized in that, The process of passing the comprehensive feature representation to the fully connected layer and output layer at the end of the model, and mapping the high-dimensional features to a scalar value between 0 and 1 through linear weighting and a non-linear sigmoid activation function, includes: The comprehensive feature representation output by the bidirectional long short-term memory neural network is used as an input vector and passed to the following fully connected layer. The fully connected layer performs a linear transformation on the input vector through a weight matrix and a bias vector to achieve a linear weighted sum of all input features and generate a preliminary linear prediction value. The initial linear prediction value output by the fully connected layer is input to the output layer; the output layer is configured with a Sigmoid activation function, which non-linearly compresses and maps the initial linear prediction value to the interval between 0 and 1, forming the model's native output value; The original output value is denormalized to convert and restore its theoretical output range of 0 to 1 to an actual estimate of the current health status as a percentage relative to the battery's rated capacity, thus obtaining the scalar value.

7. The method for dynamic balancing and health prediction of charging main cells based on multi-sensor fusion according to claim 1, characterized in that, The current health status values ​​of each battery cell are compared to calculate the health status difference between the cells; when the health status difference exceeds a preset first threshold, a dynamic balancing instruction is generated, including: Calculate the standard deviation of the current health status values ​​of all cells in the battery pack as a quantitative indicator of the health status difference, representing the degree of dispersion of health status among cells; The difference in health status is compared with a preset first threshold. When the difference in health status is greater than the first threshold, it is determined that equalization needs to be initiated, and an equalization trigger signal is generated. After equalization is triggered, the cells with the highest and lowest current health status values ​​in the battery pack are identified and marked as energy source cells and energy receiving cells, respectively. Based on the relative magnitude of the health status difference and the first threshold, the equilibrium intensity coefficient is calculated using a preset proportional-integral algorithm, wherein the greater the health status difference, the higher the calculated equilibrium intensity coefficient. Based on the equalization intensity coefficient, the identifiers of the energy source cell and the energy receiver cell, a dynamic equalization command is generated, which includes the equalization current magnitude, equalization duration, and equalization path.

8. The method for dynamic balancing and health prediction of charging main cells based on multi-sensor fusion according to claim 7, characterized in that, The step of calculating the equilibrium intensity coefficient based on the relative magnitude of the health status difference and the first threshold using a preset proportional-integral algorithm includes: Subtracting the first threshold from the current health status difference is defined as the current error value. Add the product of the previous integral value and the current error value multiplied by the sampling time interval, and use the result as the current integral value. The proportional term obtained by multiplying the current error value by the preset proportional gain coefficient is added to the integral term obtained by multiplying the current integral value by the integral gain coefficient. The sum is used as the equilibrium intensity coefficient at the current moment. Both the proportional gain coefficient and the integral gain coefficient are positive numbers.

9. The method for dynamic balancing and health prediction of charging main cells based on multi-sensor fusion according to claim 8, characterized in that, The step of generating the dynamic balancing command, which includes the balancing current magnitude, balancing duration, and balancing path, based on the balancing intensity coefficient, the identifiers of the energy source cell, and the energy receiver cell, includes: Multiply the equalization intensity coefficient by the preset maximum allowable equalization current to obtain the actual equalization current value of this equalization operation, wherein the actual equalization current value does not exceed the preset safe current limit. Based on the differences in health status between cells and the actual balancing current value, the balancing duration is determined by querying a preset balancing time mapping table. The greater the difference in health status, the longer the required balancing duration. Based on the identifiers of the energy source cell and the energy receiving cell, a corresponding switch control signal sequence is generated, the conduction state of the power switching devices in the active equalization circuit is configured, and an energy transfer path from the energy source cell to the energy receiving cell is established. The actual equalization current value, equalization duration, and equalization path configuration information are encapsulated into a complete dynamic equalization command and sent to the equalization execution circuit.

10. A device for dynamic balancing and health prediction of charging main cells based on multi-sensor fusion, characterized in that, The device includes: The data acquisition module is used to simultaneously collect voltage, current and temperature data of each cell during the charging process of the battery pack, forming multi-source sensor time-series data; The health characterization module is used to calculate and extract a set of health factors that can characterize the aging state of the battery cell based on the time-series data of the multi-source sensors. The health factors include parameters related to charging time, voltage curve changes, temperature changes and internal resistance. The health model module is used to input the health factors into a pre-trained health status estimation model, and to estimate and output the current health status value of each cell in real time. The dynamic balancing module is used to compare the current health status values ​​of each cell and calculate the health status difference between cells. When the health status difference exceeds a preset first threshold, a dynamic balancing command is generated. The balancing intensity of the dynamic balancing command is positively correlated with the magnitude of the health status difference. According to the dynamic balancing command, the active balancing circuit is controlled to redistribute energy to the corresponding cells.

Images