Intelligent device battery collaborative charging management system and method based on multi-state prediction

Abstract

The present application relates to the field of cooperative charging technology, in particular to a battery cooperative charging management system for intelligent devices based on multi-state prediction, comprising a multi-state prediction correction module, which dynamically constructs a short-time sliding sequence and trains a dynamic coupling analysis model; the dynamic coupling analysis model outputs a plurality of continuous prediction results under a prediction step m after being trained; it is determined whether the dynamic coupling analysis model is globally optimal; for the dynamic coupling analysis model that is not globally optimal, the prediction results are corrected; a signal output module acquires a safety threshold, receives the prediction results output by the multi-state prediction correction module, and determines whether to output an adjustment signal; a charging current dynamic regulation module receives the adjustment signal and adjusts the charging current; the temperature difference is calculated, the target current is analyzed through the temperature difference and the temperature duration, and then the most suitable real-time charging current in the short-time sliding sequence is selected as the adjustment current through the target current.

Description

Technical Field

[0001] This invention relates to the field of collaborative charging technology, and more specifically, to a smart device battery collaborative charging management system and method based on multi-state prediction. Background Technology

[0002] The collaborative charging management of smart device batteries specifically involves constructing a dynamic control system based on multi-state prediction to address the differentiated chemical condition requirements of various types of smart device batteries, such as drone remote control batteries, dictionary pen built-in batteries, and tablet lithium batteries. During the charging process, due to significant differences in the chemical composition, capacity level, polarization sensitivity, and applicable scenarios of different battery types, device manufacturers implement segmented charging stages, such as constant current and constant voltage stages, to avoid safety risks and battery capacity degradation caused by charging parameter mismatch, including overheating, overcharging, and lithium dendrite precipitation. The constant current stage is used to quickly replenish 70%-80% of the gate layer capacity with a stable large current in the low battery range. The constant voltage stage is used to suppress polarization reaction and control the charging current to gradually decrease when the battery is close to saturation, so as to avoid overcharging and overheating risks while ensuring charging efficiency. Furthermore, to avoid problems such as thermal runaway, lithium dendrite precipitation, and battery capacity degradation caused by excessive temperature and current fluctuations during charging, a high-precision current sensor, surface temperature detector, voltage acquisition module, and internal resistance monitoring unit are used to collect real-time status data such as current, battery temperature, remaining charge (SOC), terminal voltage, and internal resistance changes during charging. Then, by setting a preset safety threshold, the real-time status data is compared with the corresponding threshold one by one to trigger an abnormal intervention mechanism in a timely manner. Specifically, the charging temperature during the charging process is detected by monitoring equipment. If the charging temperature is greater than or equal to the safety threshold, the temperature will be controlled by adjusting the charging current. However, during continuous charging, the charging temperature will gradually increase due to the rise of the SOC gate layer, the accumulation of polarization reactions, fluctuations in ambient temperature, and real-time load changes of the device (such as multitasking on a tablet or standby power consumption of a drone remote controller). The charging temperature will continue to increase. If the charging temperature is greater than or equal to the corresponding safety threshold at a certain moment, due to the lag in battery surface temperature detection, the internal core temperature has already far exceeded the surface monitoring value, and the polarization effect has already formed an accumulated trend, resulting in a strong inertia for temperature rise. At this time, adjusting the charging current can suppress subsequent heat generation and slow down the temperature rise trend to a certain extent. However, if a large adjustment is made at once, it will not only lead to a significant increase in charging time, violating the core requirements of fast charging, but also cause a sudden change in the lithium ion insertion and extraction rate inside the battery, leading to the risk of lithium plating on the negative electrode. At the same time, the sudden change in current will exacerbate the fluctuation of battery internal resistance, resulting in increased energy loss, shortened battery cycle life, and may even cause local overheating due to electrochemical reaction disorder, which will exacerbate safety hazards. In view of this, we propose a smart device battery collaborative charging management system and method based on multi-state prediction. Summary of the Invention

[0003] The purpose of this invention is to solve the problem of how to accurately control the temperature when the charging temperature is greater than or equal to the safety threshold.

[0004] To achieve the above objectives, the present invention provides a smart device battery collaborative charging management system capable of multi-state prediction, comprising: The multi-state prediction and correction module acquires real-time charging temperature, charging current, and remaining power, and dynamically constructs a short-time sliding sequence. The short-time sliding sequence is used to train the dynamic coupling analysis model in real time. After the dynamic coupling analysis model is trained, it outputs multiple consecutive prediction results at a prediction step size m, namely, predicted temperature, predicted current, and predicted power. Determine if the dynamic coupling analysis model is globally optimal; for dynamic coupling analysis models that are not globally optimal, correct the prediction results. The adjustment signal output module obtains the safety threshold, receives the prediction result output by the multi-state prediction correction module, calculates the temperature duration and charging duration by predicting the temperature, predicting the power consumption and the corresponding safety threshold, and determines whether to output an adjustment signal. The dynamic charging current control module receives the adjustment signal and adjusts the charging current: it calculates the temperature difference, analyzes the target current through the temperature difference and temperature duration, and then selects the most suitable real-time charging current in the short-time sliding sequence as the adjustment current, thereby adjusting the charging temperature during the battery charging process.

[0005] As a further improvement to this technical solution, the multi-state prediction correction module includes a dynamically coupled prediction unit, a global optimal determination unit, and a prediction result dynamic correction unit, wherein: The dynamic coupling prediction unit is used to dynamically construct short-time sliding sequences and train dynamic coupling analysis models, which include gating layers and fully connected layers. The global optimal determination unit is used to determine whether the dynamic coupling analysis model is globally optimal. The prediction result dynamic correction unit receives multiple consecutive prediction results output by the dynamic coupling analysis model when it is not in the global optimum, and uses them to correct the corresponding prediction results.

[0006] As a further improvement to this technical solution, when the dynamic coupling prediction unit dynamically constructs the short-time sliding sequence, it takes the initial time as the start and the current time as the end of the short-time sliding sequence, and extracts the charging temperature, charging current, and remaining power within the current time to form the short-time sliding sequence. After obtaining the real-time result of the next time step, the real-time result of the next time step is added to the short-time sliding sequence of the current time step by sliding update, thereby dynamically updating the short-time sliding sequence.

[0007] As a further improvement to this technical solution, the gating layer in the dynamic coupling analysis model specifically includes a forget gate, an input gate, and an output gate: The forget gate receives the real-time result at the current moment and the hidden state at the previous moment, and uses the weight vector between the activation functions; The input gate is used to write key information of the current moment into the cell state. The output gate is used to filter information from the updated cell state and generate the hidden state at the current moment. Fully connected layer: includes temperature prediction branch, current prediction branch, and energy prediction branch, which correspond to predicting temperature, predicting current, and predicting energy, respectively. Each of the three parallel branches uses an independent weight matrix and bias term to perform multi-objective synchronous prediction. Specifically, the final hidden state is input into the fully connected layer, and the prediction step size m is set. The fully connected layer outputs multiple consecutive prediction results under the prediction step size after a short-time sliding sequence, which are multiple consecutive predicted temperatures, predicted currents, and predicted charges under the prediction step size m. It also retrieves the corresponding real-time results from the short-time sliding sequence and calculates the iteration error and mean square error between the predicted results and the real-time results at different times during each iteration. The model parameters of the dynamic coupling analysis model are updated using mean square error backpropagation; It also sets iterative convergence conditions to determine whether the dynamic coupling analysis model has been trained.

[0008] As a further improvement to this technical solution, the dynamic coupling prediction unit determines whether a single training iteration of the dynamic coupling analysis model is complete, specifically as follows: The iterative convergence condition is the maximum number of iterations and the error convergence threshold; If the number of iterations is less than the maximum number of iterations and the training rate of change is less than the error convergence threshold when training the dynamic coupling analysis model in each iteration, then the dynamic coupling analysis model is considered to have completed a single training iteration. After a single training iteration of the dynamic coupling analysis model, the model parameters corresponding to the minimum mean square error after a single training iteration are defined as the optimal parameters, and the minimum iteration error corresponding to the optimal parameters is defined as the single iteration error.

[0009] As a further improvement to this technical solution, the global optimal determination unit is used to determine whether the dynamic coupling analysis model is globally optimal, specifically by: The system receives the single iteration error of the dynamic coupling analysis model after each iteration in the dynamic coupling prediction unit, as well as multiple prediction results output by the dynamic coupling analysis model at prediction step size m after a single iteration. Calculate the training convergence rate between adjacent single-iteration errors, and then use the iterative convergence condition in the dynamic coupling prediction unit to determine whether the dynamic coupling analysis model is globally optimal.

[0010] As a further improvement to this technical solution, the prediction result dynamic correction unit receives multiple consecutive prediction results output by the dynamic coupling analysis model when it is not in the global optimum; a weight factor is set to correct the prediction results corresponding to the non-global optimum, and the prediction results at multiple prediction times under the prediction step size m are corrected by the weight factor. The specific weight factor is as follows: Calculate the training convergence rate between the current training cycle and the previous training cycle; The dynamic adaptation coefficient corresponding to the training convergence rate is set as the weight factor.

[0011] As a further improvement to this technical solution, the adjustment signal output module obtains a preset safety threshold and receives the prediction results output by the dynamic coupling analysis model; Compare the safety thresholds and prediction results in chronological order; If the safety threshold is greater than or equal to the corresponding prediction result, then the time difference between the corresponding prediction time and the current time is defined as the warning duration, where the warning duration is specifically the temperature duration and the charging duration; Compare the temperature duration with the charging duration: if the temperature duration is less than the charging duration, an adjustment signal is output.

[0012] As a further improvement to this technical solution, the charging current dynamic control module includes a target current calculation unit and an adjustment current smoothing adaptation unit, wherein: The target current calculation unit receives the temperature duration from the adjustment signal output module; and calculates the temperature difference between the predicted temperature and the temperature threshold. Adjustment coefficients are set based on temperature differences and temperature duration; The charging current in the short-time sliding sequence at the latest moment is adjusted to the target current by adjusting the coefficients. The current smoothing adaptation unit is used to receive the target current from the target current calculation unit. The charging current at different times within the short-time sliding sequence in the dynamic coupling prediction unit is compared with the target current, and the charging current closest to the target current is selected as the adjustment current.

[0013] The intelligent device battery collaborative charging management method based on multi-state prediction includes the following steps: S1. Dynamically construct short-time sliding sequences to train the dynamic coupling analysis model. After the dynamic coupling analysis model is trained, it outputs multiple consecutive prediction results at a prediction step size m, namely, predicted temperature, predicted current, and predicted charge. Determine whether the dynamic coupling analysis model is globally optimal. For dynamic coupling analysis models that are not globally optimal, correct the prediction results. S2. Obtain the safety threshold, receive the prediction result, and calculate the temperature duration and charging duration by using the predicted temperature, predicted power and corresponding safety threshold, in order to determine whether to output the adjustment signal; S3. Receive adjustment signal and adjust charging current: Calculate temperature difference, analyze target current through temperature difference and temperature duration, and then select the most suitable real-time charging current in the short-time sliding sequence as the adjustment current to adjust the charging temperature during battery charging.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: In this intelligent device battery collaborative charging management system and method based on multi-state prediction, a short-time sliding time window constructed by a dynamically coupled prediction unit enables the prediction model to quickly capture the local dynamic coupling relationship between current, temperature, and charge during the charging process. Relying on real-time sliding updated measured multi-dimensional state data, fine-tuning is performed based on historically optimal parameters instead of retraining from scratch, significantly reducing computational resource consumption and shortening the convergence time of a single training round, dynamically adapting to subtle changes in operating conditions during charging. Furthermore, a global optimality determination unit further calculates the training convergence rate of adjacent single-iteration errors and determines whether the prediction model is globally optimal, dynamically adjusting the prediction results. The correction unit provides accurate global optimum determination criteria and reliable quantitative data for prediction results, clarifying whether the model is currently in a local optimum and the degree of parameter fit. Furthermore, during the correction process of the prediction results by the dynamic correction unit, the negative correlation mapping relationship between the training convergence rate of adjacent training cycles during the training of the prediction unit is constructed and the weight factor is established. This allows for dynamic adjustment of the weight ratio of prediction results in different training cycles, compensating for prediction bias when the model has not reached the global optimum, avoiding the one-sidedness of single-cycle prediction, and improving the fit between the prediction results and the measured multi-dimensional state data. This provides reliable data support for subsequent safety threshold judgment and charging current adjustment. When the dynamic charging current control module generates the target current, it adjusts the charging current in the short-time sliding timing window at the current moment in a stepwise manner by adjusting the temperature difference between the predicted temperature and the safety threshold, and the temperature duration (the duration for which the predicted temperature approaches the threshold) in the signal output module. This avoids the problems caused by a one-time large adjustment of the charging current in existing methods, such as a sharp drop in charging efficiency, a sharp increase in internal battery polarization, lithium-ion insertion / extraction imbalance, and lithium dendrite precipitation. During the adjustment process, the target current calculation unit accurately characterizes the influence of the temperature deviation magnitude and duration on the charging current based on the quantitative adaptation relationship between the temperature difference and the temperature duration, and sets the adjustment coefficient. The system obtains a target current that can suppress temperature rise, avoid thermal runaway risk, and ensure fast charging efficiency. The current adjustment smoothing adaptation unit selects the historical charging current with the smallest difference as the adjustment current by comparing the distance between the historical measured charging current and the target current in a short-time sliding time window. This ensures that the dynamic coupling analysis model maintains a low mean square error and avoids the model's inability to quickly adapt to new coupling relationships due to sudden current changes. It also enables the dynamic coupling analysis model to quickly learn the dynamic coupling law under new current conditions and then gradually adjust the charging temperature to within a safe threshold, ultimately achieving a multiple balance between charging safety, efficiency, and battery cycle life.

[0015] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the figures. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall module of the present invention; Figure 2 This is a flowchart illustrating the working principle of the multi-state prediction and correction module of the present invention. Figure 3 This is a flowchart illustrating the working principle of the dynamic charging current control module of the present invention.

[0017] The meanings of the labels in the diagram are as follows: 100. Multi-state prediction correction module; 110. Dynamic coupling prediction unit; 120. Global optimal determination unit; 130. Prediction result dynamic correction unit; 200. Adjustment signal output module; 300. Charging current dynamic control module; 310. Target current calculation unit; 320. Adjustment current smoothing adaptation unit. Detailed Implementation

[0018] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] refer to Figures 1-3 As shown, the intelligent device battery collaborative charging management system based on multi-state prediction includes: The multi-state prediction correction module 100 acquires real-time charging temperature, charging current, and remaining power, and dynamically constructs a short-time sliding sequence. The short-time sliding sequence is used to train the dynamic coupling analysis model in real time. After the dynamic coupling analysis model is trained, it outputs multiple consecutive prediction results at a prediction step size m, namely, predicted temperature, predicted current, and predicted power. Determine if the dynamic coupling analysis model is globally optimal; for dynamic coupling analysis models that are not globally optimal, correct the prediction results. The adjustment signal output module 200 obtains the safety threshold, receives the prediction result output by the multi-state prediction correction module 100, and calculates the temperature duration and charging duration by predicting the temperature, predicting the power and the corresponding safety threshold, which are used to determine whether to output the adjustment signal. The charging current dynamic control module 300 receives the adjustment signal and adjusts the charging current: it calculates the temperature difference, analyzes the target current through the temperature difference and temperature duration, and then selects the most suitable real-time charging current in the short-time sliding sequence as the adjustment current to adjust the charging temperature during the battery charging process.

[0020] In the implementation of the above embodiments, the multi-state prediction correction module 100 outputs multiple consecutive prediction results (including predicted temperature, predicted current, and predicted charge) at a prediction step size m, capturing the dynamic coupling relationship between current, temperature, and charge during the charging process. This allows for early prediction of potential safety risks such as excessive temperature or overcharging during the subsequent charging process, as well as the specific timing of these risks. This avoids the bias of judging from a single data point and provides accurate timing basis for safety control and current adjustment. The adjustment signal output module 200 obtains a preset safety threshold and compares the safety threshold with the consecutive prediction results output by the multi-state prediction correction module 100 in chronological order. The system defines a corresponding warning duration and determines whether to output a charging current adjustment signal based on the relationship between the temperature duration and the charging duration. After receiving the adjustment signal through the target current calculation unit 310, the charging current dynamic control module 300 sets an adjustment coefficient based on the temperature difference between the predicted temperature and the temperature threshold, and the temperature duration. It then adjusts the charging current in a stepwise manner to obtain the target current. Finally, the adjustment current smoothing adaptation unit 320 selects the adjustment current closest to the target current from the historical charging current set for charging, avoiding the polarization amplification and model error surge caused by current mutations. Ultimately, it achieves a multiple balance between charging safety, efficiency, and battery life.

[0021] The multi-state prediction correction module 100 includes a dynamically coupled prediction unit 110, a global optimal decision unit 120, and a prediction result dynamic correction unit 130, wherein: The dynamic coupling prediction unit 110 acquires multi-dimensional state data, including at least charging temperature, charging current, and remaining power, as well as the corresponding sampling time. The specific multi-dimensional state data is acquired by monitoring equipment. Based on the multi-dimensional state data, a short-time sliding sequence is dynamically constructed, and the dynamic coupling analysis model is trained using the short-time sliding sequence. After training, the dynamic coupling analysis model outputs prediction results for predicted temperature, predicted current, and predicted power. When the dynamic coupling prediction unit 110 dynamically constructs a short-time sliding sequence, it uses the initial time... Starting with the current time The current time is taken as the end point of the short-time sliding sequence. Real-time results such as charging temperature, charging current, and remaining power are used to form a short-time sliding sequence, specifically at the current moment. The short-time sliding sequence expression is: ; in, For a moment The corresponding real-time results include charging temperature, charging current, and remaining battery capacity, and their corresponding expressions are as follows: ; Collect the next moment After obtaining the real-time results, add the next time step using a sliding update method. Real-time results up to the current moment short-time sliding sequence This allows for dynamic updates to the short-time sliding sequence, specifically for the next moment. The corresponding short-time sliding sequence expression is: ; The dynamic coupling analysis model in the dynamic coupling prediction unit 110 includes a gating layer and a fully connected layer for dynamically filtering and memorizing key information in short-time sliding sequences; The gating layer specifically includes a forget gate, an input gate, and an output gate: Forget Gate: Receive the current moment Real-time results Compared to the previous moment The hidden state is used to generate a weight vector between 0 and 1 using the Sigmoid activation function. The specific expression is: ; in, For the current moment The output weight vector of the forget gate , Here are the weight matrix and bias terms for the forget gate. For the previous moment The hidden state, The Sigmoid activation function maps the input to the 0~1 range; Input gate: Used to write key information of the current moment into the cell state, specifically: The updated weight vector is generated using the Sigmoid activation function. This determines the proportion of the current input information that is written. ; Candidate cell states are generated using the Tanh activation function. It includes key features of the current moment (such as the rapid temperature rise caused by high current): ; Multiply the updated weights by the candidate cell states, and then add them to the historical cell states filtered through the forgetting gate to obtain the updated cell states: ; The above This represents the cell state at the previous moment. , , , These are the weight matrices and bias terms for the input gate and candidate cell states, respectively. The hyperbolic tangent activation function maps the input to the interval -1 to 1. Output gate: used to retrieve data from the updated cell state. Filter information to generate the hidden state at the current moment, specifically: Receive real-time results at the current moment Compared to the previous hidden state The output weight vector is generated by the Sigmoid activation function. This determines which information from the cell state needs to be output. ,in , These are the weight matrix and bias term of the output gate, respectively; The cell state is mapped to the -1 to 1 interval using the Tanh activation function, and then multiplied with the output weight vector to obtain the final hidden state. Hidden state Includes historical and current real-time results The fused information serves as the core basis for subsequent predictions of the fully connected layer. Fully connected layer: includes temperature prediction branch, current prediction branch, and energy prediction branch, which correspond to predicting temperature, predicting current, and predicting energy, respectively. Each of the three parallel branches uses an independent weight matrix and bias term to achieve simultaneous prediction of multiple targets. Specifically, the final hidden state Input a fully connected layer and set a prediction step size m (which includes multiple equidistant prediction times). The fully connected layer includes weight matrices and output layer biases corresponding to the predicted temperature, predicted current, and predicted charge at multiple equidistant prediction times under the prediction step size m. The fully connected layer outputs multiple consecutive prediction results (i.e., predicted temperature and predicted charge) under the prediction step size after a short-time sliding sequence through the weight matrices and output layer biases. The specific expressions for the predicted temperature, predicted current, and predicted charge are as follows: The predicted temperature is: ,in (m is the number of prediction steps) , These are the weight matrix and output layer bias of the temperature prediction branch at the i-th prediction time, respectively. The predicted current is: ,in (m is the number of prediction steps) , These are the weight matrix and output layer bias of the current prediction branch at the i-th prediction time, respectively. The predicted power consumption is: ,in , , These are the weight matrix and output layer bias of the power prediction branch at the i-th prediction time, respectively. The training process of the dynamic coupling analysis model using a short-time sliding sequence also includes an iterative mechanism, the specific principle of which is as follows: Real-time results from the short-time sliding sequence used to train the dynamic coupling analysis model are retrieved; the iteration error between the predicted result and the real-time result at different times during each iteration is calculated; and the iteration error is calculated at each time step. Taking the iteration error as an example, the specific expression is as follows: ; in, For a moment Real-time results (specifically referring to the actual temperature) (or actual electricity consumption) ); For a moment The prediction results at that time (specifically referring to the predicted temperature) Or predict electricity consumption ); Calculate the mean squared error: Add the iteration errors corresponding to the prediction results and real-time results in each iteration, and calculate the mean to obtain the mean squared error. ; The model parameters of the dynamic coupling analysis model (specifically the weights and biases of the gating layer and the prediction layer) are updated using mean squared error backpropagation respectively: ,in , These are the model parameters before and after the update, respectively. This is the partial derivative of the iteration error with respect to the parameters of the old model (reflecting the degree of influence of changes in the parameters of the old model on the prediction error). The learning rate; Set the maximum number of iterations With error convergence threshold The iterative convergence condition for iteratively training the dynamic coupling analysis model; Calculate the training rate of change ( , (These are the iteration errors corresponding to the i-th and (i-1)-th backpropagation updates of the dynamic coupling analysis model, respectively). If the number of iterations is used in each iteration to train the dynamic coupling analysis model... Maximum number of iterations And training rate of change <Error convergence threshold If so, it is determined that the dynamic coupling analysis model has completed a single training iteration; Define the minimum mean square error after a single training iteration (i.e., the number of iterations required to satisfy the iteration convergence condition). Maximum number of iterations The minimum mean square error (MSE) corresponds to the optimal model parameters (i.e., the model parameters used by the dynamic coupling analysis model to output the final prediction result at the current moment), and the minimum iteration error corresponding to the optimal parameters is the single iteration error. ,in The minimum function is defined as the sum of all iteration errors that satisfy the condition. In the selection process, the error with the smallest value is chosen, corresponding to the error level of the optimal parameter. The above iterative convergence condition (maximum number of iterations) Error convergence threshold Learning rate Commonly used engineering methods such as empirical value method, k-fold time-series cross-validation, grid search method, and Bayesian optimization method can be used to scientifically set parameters to balance the computational power adaptability of dynamic coupling analysis model training efficiency and prediction accuracy, and to avoid parameter (maximum number of iterations) Error convergence threshold Learning rate Blindly setting parameters can lead to problems such as training oscillations, slow convergence, or excessive prediction bias. Taking the grid search method as an example, the specific working principle is as follows: The maximum number of iterations can be preset. Error convergence threshold Learning rate The corresponding discrete candidate set: The candidate set for the maximum number of iterations is set to {100, 150, 200, 250, 300} (to balance training convergence sufficiency with computational cost), the candidate set for the error convergence threshold is set to {1e-4, 3e-4, 5e-4, 8e-4, 1e-3} (to match the error accuracy standard for multi-state prediction), and the candidate set for the initial learning rate is set to {0.001, 0.003, 0.005, 0.008, 0.01} (to adapt to the training stability of the gated model). A complete parameter grid (5×5×5=125 parameter combinations) is constructed by the Cartesian product of each parameter candidate set. Each parameter combination in the grid is then evaluated one by one. The dynamic coupling analysis model is trained based on the short-time sliding sequence dataset. The maximum number of iterations in the parameter combination is used as the upper limit of the training rounds and the error convergence threshold is used as the termination condition to complete the model training. Then, the mean square error (MSE) of the model is calculated using the validation set covering the full charging conditions (constant current and constant voltage stages) as the performance index of the parameter combination. After all parameter combinations are evaluated, the parameter combination with the smallest MSE in the validation set and the shortest training time is selected as the final setting value of the maximum number of iterations, the error convergence threshold, and the learning rate.

[0022] "Single training completion" specifically refers to training the dynamic coupling analysis model without updating the short-time sliding sequence. Therefore, during each iteration of training the dynamic coupling analysis model, the dynamic coupling prediction unit 110 dynamically constructs a short-time sliding sequence for iterative training and prediction. Specifically, in each iteration, the newly collected real-time results of multiple dimensions such as charging temperature, charging current, and remaining power are added to the short-time sliding sequence of the previous moment in a sliding update manner. Then, the new short-time sliding sequence is used to train the dynamic coupling analysis model. During this training process, the optimal parameters obtained in the previous round of training are used as the basis, and the model parameters are fine-tuned through backpropagation of the corresponding mean square error. This means that during each training of the pre-model, it is not necessary to train the dynamic coupling analysis model from scratch, but only to fine-tune it based on the historical optimal parameters. This reduces the computational resource consumption when training the dynamic coupling analysis model and shortens the convergence time of a single round of training, so as to dynamically adapt to the changes in the coupling relationship of current, temperature, and power during charging until the dynamic coupling analysis model is globally optimal. Furthermore, considering that during the training of the dynamic coupling analysis model by the dynamic coupling prediction unit 110, the iterative convergence condition can be used to determine whether the dynamic coupling analysis model has been trained successfully, and that the short-time sliding sequence will be updated in real time as the charging time progresses and continuously provide dynamic adaptation multi-dimensional data support for model training, in order to avoid the dynamic coupling analysis model from only satisfying single training convergence and getting stuck in local optima, thus failing to cover the dynamic working conditions of the entire charging process, and to ensure the global reliability of the prediction results and the accuracy of subsequent charging control, the global optimum determination unit 120 is used to determine whether the dynamic coupling analysis model is globally optimal. The specific determination principle is as follows: The system receives the single iteration error of the dynamic coupling analysis model after each iteration in the dynamic coupling prediction unit 110, as well as the multiple prediction results of the dynamic coupling analysis model output at prediction step size m after a single iteration. The training convergence rate between adjacent single iteration errors is calculated, and then the iterative convergence condition in the dynamic coupling prediction unit 110 is used again to determine whether the dynamic coupling analysis model is globally optimal. Even when the dynamic coupling analysis model has reached global optimum, subsequent iterative training continues, with the training convergence rate calculated synchronously during the iteration. The dynamic coupling prediction unit 110 then sets corresponding iterative convergence conditions to determine whether the model still maintains global optimum or needs to terminate the iteration. This allows for real-time optimization of the dynamic coupling analysis model under the dual constraints of accuracy and state stability. Furthermore, after iteration termination, the prediction result dynamic correction unit 130 can further correct the prediction results output by the dynamic coupling analysis model. This improves prediction results. With real-time results The fit.

[0023] Furthermore, during the training process of the dynamic coupling analysis model (including single training completion and global optimization), short-time series sequences can be divided into training and validation sets. Through real-time monitoring of training and validation errors and an early stopping mechanism, this ensures that the dynamic coupling analysis model fully learns local dynamic coupling relationships in a single training session and adapts to full charging conditions (constant current / constant voltage stage, SOC 0%-100% range, ambient temperature fluctuations) when globally optimal. Simultaneously, it avoids overfitting and underfitting risks, ensuring prediction accuracy (temperature error ≤ 0.5℃, current error ≤ 0.1A, and power error ≤ 1%). The specific principle is as follows: The partitioning follows the core principle of time-series data: the training set time dimension is completely earlier than the validation set, and it is never randomly split to avoid future data leakage causing the dynamic coupling analysis model to learn false coupling patterns. At the same time, it ensures that the validation set contains at least one complete charging cycle and covers different working conditions to ensure the reliability of the evaluation. In the specific partitioning, a fixed-length rolling window (e.g., each window contains 30 consecutive time-series data points) is set for the short-time series ordered by charging time. The initial training set takes the first 70% of the time-series data (covering early operating conditions such as low SOC and constant current stage), and the validation set takes the last 30% of the data (covering subsequent operating conditions such as high SOC and constant voltage stage). As the model training iterates, the window scrolls forward along the time axis. After each iteration, the training set adds historical data of one window length, while the validation set maintains a fixed length, realizing a rolling validation mode in which the training set gradually expands and the validation set is dynamically updated. During the round-by-round cross-validation process, the model prediction error (mean squared error) is calculated using the validation set after each training round. The changing trends of the training error and validation error are monitored in real time: if the training error continues to decrease while the validation error tends to stabilize, it indicates that the dynamic coupling analysis model has learned sufficiently and has good generalization ability, meeting the conditions for completing a single training session; if the training error continues to decrease but the validation error increases, it is judged as overfitting, and the current training is terminated through the early stopping mechanism and the model complexity is fine-tuned (such as adjusting the number of neurons in the fully connected layer); if both remain high and decrease slowly, it is judged as underfitting, and the model's expressive ability is improved by increasing the number of training rounds or optimizing parameters (such as adjusting the learning rate); for global optimal validation, the validation results of multiple rounds of rolling partitioning need to be accumulated, and the mean and fluctuation amplitude of the validation error under different working conditions are calculated to ensure that the model error remains stable throughout the full charging process.

[0024] Furthermore, during the charging process of smart device batteries, in order to balance charging efficiency and battery safety, the battery capacity is quickly replenished to 70%-80% within the safety threshold. This avoids damage to electrode materials from the initial high current impact, reduces the risk of lithium dendrite precipitation, and suppresses the voltage surge and thermal runaway risks caused by the intensification of polarization reaction, thus extending the battery cycle life. Therefore, constant current charging is adopted. During the constant current charging process of smart device batteries, because the charger maintains a stable output current through current feedback closed-loop control during the constant current stage, the battery voltage gradually increases linearly with the increase of remaining capacity. Moreover, the lithium ion insertion and extraction rate is stable, making the polarization reaction relatively mild. The fluctuation amplitude of the various dimensions of charging current, temperature, and capacity in the short-time sliding sequence is extremely small. The dynamic coupling analysis model does not need to process complex dynamic fluctuation data and can quickly capture the local coupling relationship between the three in the constant current stage, thereby efficiently completing a single round of training and initially adapting to the working condition characteristics of the current charging stage. For the constant current stage, although the dynamic coupling analysis model can be trained quickly, the short-time sliding sequence in the constant current stage can only capture the coupling relationship of the current gate layer, temperature gate layer, and charge gate layer that are currently locally stable. After the dynamic coupling analysis model is trained quickly, it is easy to get stuck in a local optimum rather than a global optimum. This can cause the prediction results output during the training interval between a single training of the dynamic coupling analysis model and the global optimum of the dynamic coupling analysis model to be inaccurate. Therefore, in order to avoid the situation where the prediction results are inaccurate when the dynamic coupling analysis model is not in a global optimum, the prediction result dynamic correction unit 130 receives multiple consecutive prediction results output when the dynamic coupling analysis model is not in a global optimum. A weighting factor is set to correct the prediction result when it is not in the global optimum. This factor is used to correct the prediction results at multiple prediction times under a prediction step size m. The specific principle is as follows: Calculate the training convergence rate between the current training period k and the previous training period k-1: ; in, , These represent the single iteration error when the dynamic coupling analysis model is not in the global optimum at training period k and training period k-1, respectively. Training convergence rate It can accurately characterize the convergence trend and prediction reliability of the dynamic coupling analysis model in the non-global optimal stage. Essentially, it represents the relative change in error between two adjacent iterations. It directly reflects the degree of adaptation and optimization dynamics of the optimal parameters of the dynamic coupling analysis model to the coupling relationship between the current-gated layer, temperature-gated layer, and charge-gated layer during the charging process. Specifically: Training convergence rate The larger the value, the more likely the dynamic coupling analysis model is to be in the stage of dynamic parameter adjustment. At this time, the dynamic coupling analysis model has not yet fully learned the differences in coupling relationships caused by the characteristics of operating conditions such as the evolution of battery polarization effect, ambient temperature fluctuation, and changes in electrolyte ion migration efficiency over a long time scale. Training convergence rate When the value is small, it indicates that the dynamic coupling analysis model has undergone multiple rounds of iterative fine-tuning, and the corresponding optimal parameters have been adapted to the current convergence stage. The prediction parameter update gradually becomes stable. Specifically, the optimal parameters at this time not only fully absorb the effective features in the previous training, but also dynamically follow the subtle operating condition drift during the charging process, and capture the nonlinear coupling relationship between the current gate layer, temperature gate layer, and energy gate layer more accurately. For the reasons mentioned above, the training convergence rate is set. The corresponding dynamic adaptation coefficient is the weighting factor. By establishing the training convergence rate With weighting factors Negative correlation mapping relationship (training convergence rate) Larger weighting factor The smaller the value, the higher the training convergence rate. Smaller weighting factor The larger the value, the more the prediction results are further revised. The specific revised prediction results are as follows: ; in, , These represent the training period k, the prediction step size m in training period k-1, and the prediction time. Uncorrected forecast results; When smart devices are charging batteries together, there are unavoidable interference factors during the charging process, such as instantaneous fluctuations in ambient temperature, random accumulation of battery ohmic polarization and concentration polarization, sensor data acquisition noise, sudden changes in operating conditions during CC-CV stage switching, and activation of hidden defects such as micro-short circuits inside the battery under dynamic operating conditions. These interference factors will cause fluctuations in the newly added real-time results in the short-time sliding sequence, and further increase the training convergence rate of the corresponding training cycle. It is precisely because the training convergence rate increases that, based on its negative correlation with the weight factor, a smaller weight factor will be automatically assigned to the prediction result of this cycle. Through this dynamic adaptive weight adjustment to correct the corresponding prediction result, it is possible to further avoid the interference of the newly added real-time results in the short-time sliding sequence on the overall prediction due to abnormal fluctuations. This eliminates the problem that abnormal data cannot reflect the real temperature, current, and power dynamic coupling relationship of the battery, and ensures that the prediction result is consistent with the actual working state of the battery. In the process of correcting prediction results by training convergence rate, the allocation of weight factors is always based on the convergence stability of model training. When the convergence rate increases (abnormal fluctuation scenario), the impact of abnormal period prediction results is weakened, and when the convergence rate is low (normal learning stage of model), the weight of effective operating condition information is strengthened. At the same time, the stable prediction results of the adjacent previous training cycle are reused as the basic reference to form a smooth transition logic in time sequence. This not only conforms to the time dependence characteristics of short-time sliding sequence, but also avoids large fluctuations in prediction values ​​caused by abnormal data in a single cycle. It also makes the prediction results relatively stable, so that the subsequent formulation of charging current and voltage regulation strategies and the determination of the global optimal state of the dynamic coupling analysis model can be more accurate and reliable.

[0025] The global optimal determination unit 120 first determines whether the dynamic coupling analysis model has reached the global optimal. Then, the prediction result dynamic correction unit 130 makes targeted corrections to the initial prediction results output by the dynamic coupling analysis model based on the global optimal determination result, so that the dynamic coupling analysis model can accurately capture the dynamic coupling relationship of the current gate layer, temperature gate layer, and power gate layer during the charging process and output reliable prediction results that fit the actual working conditions. When the dynamic coupling analysis model outputs the final prediction result, to differentiate between different scenarios during the charging process of a smart device battery—such as efficiently completing charging within a safe threshold versus potential safety risks like excessive temperature (exceeding 45°C above the gate layer can easily lead to electrolyte decomposition, lithium dendrite precipitation, and other thermal runaway risks) and overcharging (leading to loss of active lithium and excessive growth of the SEI gate layer film)—it uses multiple consecutive prediction results from the gate layer at prediction step size m (avoiding the one-sidedness of a single prediction result and improving prediction reliability through multi-step prediction error control) to know in advance the battery's temperature, current, and capacity during the subsequent charging process. Whether the risk approaches or exceeds the safety threshold, and the specific predicted time when the above risks may occur, the timing relationship between the first predicted temperature and the first predicted full charge, thereby shifting safety control from passive response to active prediction, avoiding permanent battery damage or safety accidents caused by sudden risks, and providing accurate timing basis for the charging current adjustment strategy in the subsequent target current calculation unit 310 to balance charging safety and efficiency. Specifically, the adjustment signal output module 200 obtains the preset safety threshold and receives the prediction results (multiple consecutive predicted temperatures or predicted charges under prediction step size m) output by the dynamic coupling analysis model. Compare the safety thresholds and prediction results in chronological order; If the safety threshold (including the temperature safety threshold and the power safety threshold) is greater than or equal to the corresponding prediction result (including the predicted temperature and the predicted power), then the time difference between the corresponding prediction time and the current time is defined as the warning duration, where the warning duration is specifically the temperature duration. and charging time ; Comparison of temperature duration and charging time The system determines whether an adjustment signal for adjusting the charging current during the charging process needs to be output to the target current calculation unit 310. Specifically: If temperature duration ≥ Charging time This indicates that the battery can be fully charged before the temperature exceeds the limit, therefore there is no need to adjust the charging current; If temperature duration <Charging time This indicates that the battery will trigger a risk of exceeding the temperature limit before it is fully charged, so an adjustment signal is output to the target current calculation unit 310.

[0026] To achieve the optimal balance between charging safety and fast charging efficiency, the charging current dynamic control module 300 includes a target current calculation unit 310 and an adjustment current smoothing adaptation unit 320, wherein: The target current calculation unit 310 receives the adjustment signal output by the adjustment signal output module 200 and uses it to analyze the target current. Specifically: Temperature duration in the receive and adjust signal output module 200 ; Calculate the temperature difference between the predicted temperature and the temperature threshold: ,in For the predicted time Predicted temperature; Based on temperature difference With temperature and duration The charging current in the short-time sliding sequence is adjusted in a stepwise manner at the current moment (and the adjusted charging current is synchronously updated to the short-time sliding sequence for subsequent iterative training), enabling the battery temperature to be predicted throughout the charging process. Always controlled by the corresponding temperature threshold Internally, this avoids the risk of thermal runaway caused by excessive temperature during charging, such as lithium dendrite precipitation and electrolyte decomposition, and also utilizes temperature differences. With temperature and duration The quantitative adaptation precisely matches the current adjustment range. The logical relationship between the larger the temperature difference and the temperature duration is as follows: Temperature difference Larger, longer temperature duration The smaller the value, the larger the adjustment range of the charging current each time, in order to quickly control the temperature; The smaller the temperature difference Temperature duration The larger the current, the smaller the adjustment range of the charging current each time, so as to maintain a larger charging current to improve efficiency, and ultimately achieve the optimal balance between charging safety and fast charging efficiency. Based on temperature difference With temperature and duration Set adjustment factor: ; in, The normalized temperature difference The normalized temperature duration , , , These are the maximum and minimum temperature differences and temperature durations (the time difference between the predicted temperature first approaching the safety threshold and the current moment), respectively. Normalization is used to eliminate the dimensional influence between temperature differences and temperature durations and to unify the parameter value range to the [0,1] interval. By adjusting the coefficients Adjust the charging current in the short-time sliding sequence at the latest time to the target current. ,in This refers to the charging current in the short-time sliding sequence at the latest moment; Since the short-time sliding sequence in the dynamic coupling prediction unit 110 is dynamically constructed based on multi-dimensional data such as charging temperature, charging current, and remaining power collected in real time by the monitoring device during the charging process and the corresponding sampling time, in the constant current charging stage of the smart device battery, the power output current is stable and the battery polarization reaction is smooth, resulting in minimal fluctuation of the real-time charging current collected by the monitoring device. This further leads to the dynamic coupling analysis model in the multi-state prediction correction module 100 being able to learn and adapt only to the fixed coupling relationship between charging current, temperature, and power in the above stable state. However, when the signal output module 200 determines that the charging current needs to be adjusted, if the difference between the adjusted charging current and the original stable real-time charging current in the short-time sliding sequence is too large, it will directly change the battery polarization reaction rate, electrolyte ion migration efficiency and heat generation rate, thereby causing dynamic changes in the coupling relationship between the current gate layer, temperature gate layer and charge, forming a new nonlinear coupling law. If the target current is used directly at this time When adjusting the charging current, the original model parameters of the dynamic coupling analysis model are only adapted to the original stable current condition and cannot match the new coupling relationship, which leads to a significant increase in the mean square error during the training process. The dynamic coupling analysis model needs to continuously supplement multi-dimensional data under the new current condition and repeatedly fine-tune the weights and bias terms to learn the new dynamic coupling law, which ultimately makes the process of accurately adjusting the charging current lengthy. Therefore, the current smoothing adaptation unit 320 receives the target current from the target current calculation unit 310. The system calculates the charging current at different times within the short-time sliding sequence in the dynamic coupling prediction unit 110, compares all charging currents with the target current, and selects the charging current closest to the target current as the adjustment current. The battery is then further charged using this adjustment current. This minimizes the difference between the adjustment current and the original charging current adapted to the dynamic coupling analysis model, preventing sudden current changes from causing a sharp increase in ohmic polarization, electrochemical polarization, and concentration polarization within the battery. It also prevents the risk of lithium plating on the negative electrode due to an imbalance in lithium-ion insertion / extraction rates. Furthermore, it avoids the problem of a surge in mean square error caused by the dynamic coupling analysis model's inability to quickly adapt to new coupling relationships. The specific adjustment current is as follows: ; in, This is a collection of charging currents from different historical moments. For a moment The corresponding charging current.

[0027] After each adjustment of the current, if the adjustment signal output module 200 still outputs a control signal, the adjustment coefficient is set again based on the temperature difference and temperature duration corresponding to the output control signal. This coefficient is then used by the target current calculation unit 310 and the adjustment current smoothing adaptation unit 320 to adjust the charging current again until the adjustment signal output module 200 no longer outputs a control signal. In summary: In this implementation, a short-time sliding sequence is dynamically constructed using a dynamic coupling prediction unit 110. Then, a dynamic coupling analysis model is trained in real-time based on this short-time sliding sequence. This allows for rapid capture of the local dynamic coupling relationships of the current-gated layer, temperature-gated layer, and charge-gated layer during the charging process of a smart device battery. Real-time updated multi-dimensional data enhances the timeliness and adaptability of model training, enabling high-precision prediction of the charging state and providing accurate data support for subsequent charging parameter adjustments. Furthermore, during the training of the dynamic coupling analysis model, it is considered that during the charging process of a smart device battery, there are slow accumulations of ohmic polarization, electrochemical polarization, and concentration polarization. The battery internal resistance changes slightly with temperature and the state of the gated layer (SOC), and the electrolyte ion migration efficiency fluctuates slightly. Additionally, the short-time sliding sequence is limited by locally acquired data, resulting in incomplete feature coverage. This may lead to the short-time sliding sequence failing to fully reflect the coupling patterns under all operating conditions. The dynamic coupling analysis model can only learn local stable operating conditions and cannot fully learn the global coupling relationships under complex dynamic operating conditions, easily getting trapped in local optima. The global optimal determination unit 120 determines whether the dynamic coupling analysis model is globally optimal. If it is not globally optimal, the prediction result dynamic correction unit 130 corrects the output result of the dynamic coupling analysis model again, thereby making up for the prediction deviation when the dynamic coupling analysis model does not reach the global optimal, improving the fit between the prediction result and the actual working condition, avoiding the inaccuracy of charging parameter adjustment due to excessive prediction error, and providing a reliable basis for subsequent safety threshold judgment and current adjustment. Furthermore, taking into full account that the charging process is a continuous time sequence, single-moment prediction can easily lead to a lag in the adjustment of charging parameters, making it impossible to respond to potential safety risks in a timely manner. Therefore, the dynamic coupling analysis model can predict the charging state at multiple moments within the prediction step by setting a prediction step size, avoiding the problem of untimely adjustment caused by prediction based on a single moment. This allows sufficient decision time for the dynamic optimization of the charging current, effectively avoiding sudden safety hazards such as temperature exceeding the limit and overcharging. The adjustment signal output module 200 judges the charging process based on the preset safety thresholds (including the temperature threshold gate layer of 45°C, the overcharge voltage threshold, etc. Exceeding the gate layer of 45°C can easily cause electrolyte decomposition, lithium dendrite precipitation and other thermal runaway risks, and overcharging can lead to irreversible loss of active lithium and excessive growth of the SEI gate layer film) and the prediction results output by the multi-state prediction correction module 100 (including the corrected prediction results, based on whether the dynamic coupling analysis model is globally optimal). This clarifies whether the charging process is a normal scenario of efficient charging within the safety threshold or an abnormal scenario with safety risks such as temperature exceeding the limit and overcharging, providing a clear decision direction for subsequent charging current adjustment. The target current calculation unit 310, based on the temperature difference (the difference between the predicted temperature and the temperature threshold) and the adjustment coefficient set by the temperature duration, can quantify the impact of the magnitude and duration of temperature deviation on the charging current. This achieves precise matching between temperature control requirements and charging efficiency, ultimately obtaining a target current that suppresses temperature rise, avoids safety risks, and ensures charging speed. Furthermore, due to the charging characteristics of smart devices (requiring avoidance of polarization amplification and battery damage caused by sudden current changes, and the charging system needing to adapt to dynamically changing operating conditions), in this example, the current smoothing adaptation unit 320 again uses real-time data from the training process... The charging current adjusts the target current in the target current calculation unit 310, thereby minimizing the difference between the adjusted current and the model's adapted operating conditions. This avoids problems such as a sharp increase in internal polarization of the battery and lithium-ion insertion / extraction imbalance caused by sudden current changes, maintains the mean square error of the dynamic coupling analysis model at a low level, ensures the stability of prediction accuracy, and ensures a smooth transition of the charging current. This reduces the impact on the battery plates and the disorder of internal chemical reactions, reduces the capacity decay rate, and extends the battery cycle life. Ultimately, this achieves multiple goals of safe and controllable charging process, high efficiency and energy saving, and extended battery life.

[0028] The foregoing has shown and described 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 to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A smart device battery collaborative charging management system based on multi-state prediction, characterized in that, include: The multi-state prediction and correction module (100) acquires real-time charging temperature, charging current, and remaining power, and dynamically constructs a short-time sliding sequence. A dynamic coupling analysis model is trained in real time using a short-time sliding sequence. After the dynamic coupling analysis model is trained, it outputs multiple consecutive prediction results at a prediction step size m, namely, predicted temperature, predicted current, and predicted charge. Determine if the dynamic coupling analysis model is globally optimal; for dynamic coupling analysis models that are not globally optimal, correct the prediction results. The adjustment signal output module (200) obtains the safety threshold, receives the prediction result output by the multi-state prediction correction module (100), calculates the temperature duration and charging duration by predicting the temperature, predicting the power and the corresponding safety threshold, and determines whether to output the adjustment signal. The charging current dynamic control module (300) receives the adjustment signal and adjusts the charging current: it calculates the temperature difference, analyzes the target current through the temperature difference and temperature duration, and then selects the most suitable real-time charging current in the short-time sliding sequence as the adjustment current, and adjusts the charging temperature during the battery charging process by adjusting the adjustment current.

2. The intelligent device battery collaborative charging management system based on multi-state prediction according to claim 1, characterized in that: The multi-state prediction correction module (100) includes a dynamically coupled prediction unit (110), a global optimal determination unit (120), and a prediction result dynamic correction unit (130), wherein: The dynamic coupling prediction unit (110) is used to dynamically construct short-time sliding sequences and train dynamic coupling analysis models, which include gating layers and fully connected layers respectively. The global optimal determination unit (120) is used to determine whether the dynamic coupling analysis model is globally optimal; The prediction result dynamic correction unit (130) receives multiple consecutive prediction results output when the dynamic coupling analysis model is not in the global optimum, and uses them to correct the corresponding prediction results.

3. The intelligent device battery collaborative charging management system based on multi-state prediction according to claim 2, characterized in that: When the dynamic coupling prediction unit (110) dynamically constructs the short-time sliding sequence, it takes the initial time as the start and the current time as the end of the short-time sliding sequence, and extracts the charging temperature, charging current and remaining power in the current time to form the short-time sliding sequence. After obtaining the real-time result of the next time step, the real-time result of the next time step is added to the short-time sliding sequence of the current time step by sliding update, thereby dynamically updating the short-time sliding sequence.

4. The intelligent device battery collaborative charging management system based on multi-state prediction according to claim 3, characterized in that: The gating layer in the dynamic coupling analysis model specifically includes a forget gate, an input gate, and an output gate: The forget gate receives the real-time result at the current moment and the hidden state at the previous moment, and uses the weight vector between the activation functions; The input gate is used to write key information of the current moment into the cell state. The output gate is used to filter information from the updated cell state and generate the hidden state at the current moment. Fully connected layer: includes temperature prediction branch, current prediction branch, and energy prediction branch, which correspond to predicting temperature, predicting current, and predicting energy, respectively. Each of the three parallel branches uses an independent weight matrix and bias term to perform multi-objective synchronous prediction. Specifically, the final hidden state is input into the fully connected layer, and the prediction step size m is set. The fully connected layer outputs multiple consecutive prediction results under the prediction step size after a short-time sliding sequence, which are multiple consecutive predicted temperatures, predicted currents, and predicted charges under the prediction step size m. It also retrieves the corresponding real-time results from the short-time sliding sequence and calculates the iteration error and mean square error between the predicted results and the real-time results at different times during each iteration. The model parameters of the dynamic coupling analysis model are updated using mean square error backpropagation; It also sets iterative convergence conditions to determine whether the dynamic coupling analysis model has been trained.

5. The intelligent device battery collaborative charging management system based on multi-state prediction according to claim 4, characterized in that: The dynamic coupling prediction unit (110) determines whether a single training iteration of the dynamic coupling analysis model is complete, specifically as follows: The iterative convergence condition is the maximum number of iterations and the error convergence threshold; If the number of iterations is less than the maximum number of iterations and the training rate of change is less than the error convergence threshold when training the dynamic coupling analysis model in each iteration, then the dynamic coupling analysis model is considered to have completed a single training iteration. After a single training iteration of the dynamic coupling analysis model, the model parameters corresponding to the minimum mean square error after a single training iteration are defined as the optimal parameters, and the minimum iteration error corresponding to the optimal parameters is defined as the single iteration error.

6. The intelligent device battery collaborative charging management system based on multi-state prediction according to claim 5, characterized in that: The global optimal determination unit (120) is used to determine whether the dynamic coupling analysis model is globally optimal, specifically as follows: The system receives the single iteration error of the dynamic coupling analysis model after each iteration in the dynamic coupling prediction unit (110), and the multiple prediction results of the dynamic coupling analysis model output at prediction step size m after a single iteration. Calculate the training convergence rate between adjacent single iteration errors, and then use the iterative convergence condition in the dynamic coupling prediction unit (110) to determine whether the dynamic coupling analysis model is globally optimal.

7. The intelligent device battery collaborative charging management system based on multi-state prediction according to claim 6, characterized in that: The prediction result dynamic correction unit (130) receives multiple consecutive prediction results output by the dynamic coupling analysis model when it is not in the global optimum; it sets a weight factor to correct the prediction results corresponding to the non-global optimum, and corrects the prediction results at multiple prediction times under the prediction step size m through the weight factor. The specific weight factor is as follows: Calculate the training convergence rate between the current training cycle and the previous training cycle; The dynamic adaptation coefficient corresponding to the training convergence rate is set as the weight factor.

8. The intelligent device battery collaborative charging management system based on multi-state prediction according to claim 7, characterized in that: The adjustment signal output module (200) acquires a preset safety threshold and receives the prediction results output by the dynamic coupling analysis model; Compare the safety thresholds and prediction results in chronological order; If the safety threshold is greater than or equal to the corresponding prediction result, then the time difference between the corresponding prediction time and the current time is defined as the warning duration, where the warning duration is specifically the temperature duration and the charging duration; Compare the temperature duration with the charging duration: if the temperature duration is less than the charging duration, an adjustment signal is output.

9. The intelligent device battery collaborative charging management system based on multi-state prediction according to claim 8, characterized in that: The charging current dynamic control module (300) includes a target current calculation unit (310) and an adjustment current smoothing adaptation unit (320), wherein: The target current calculation unit (310) receives the temperature duration from the adjustment signal output module (200); and calculates the temperature difference between the predicted temperature and the temperature threshold. Adjustment coefficients are set based on temperature differences and temperature duration; The charging current in the short-time sliding sequence at the latest moment is adjusted to the target current by adjusting the coefficients. The current smoothing adaptation unit (320) is used to receive the target current from the target current calculation unit (310). The charging current at different times within the short-time sliding sequence in the dynamic coupling prediction unit (110) is compared with the target current, and the charging current closest to the target current is selected as the adjustment current.

10. A method for collaborative charging management of batteries in intelligent devices based on multi-state prediction, characterized in that, Includes the following steps: S1. Dynamically construct short-time sliding sequences to train the dynamic coupling analysis model. After the dynamic coupling analysis model is trained, it outputs multiple consecutive prediction results at a prediction step size m, namely, predicted temperature, predicted current, and predicted charge. Determine whether the dynamic coupling analysis model is globally optimal. For dynamic coupling analysis models that are not globally optimal, correct the prediction results. S2. Obtain the safety threshold, receive the prediction result, and calculate the temperature duration and charging duration by using the predicted temperature, predicted power and corresponding safety threshold, in order to determine whether to output the adjustment signal; S3. Receive adjustment signal and adjust charging current: Calculate temperature difference, analyze target current through temperature difference and temperature duration, and then select the most suitable real-time charging current in the short-time sliding sequence as the adjustment current to adjust the charging temperature during battery charging.

Images