Intelligent battery management and safety control system and electric flame cartridge stove

Abstract

The application relates to the technical field of battery management systems, and discloses an intelligent battery management and safety control system and an electric flame cartridge stove. The system comprises a perception engine for collecting battery multi-physical field data through a heterogeneous sensor array; a calculation engine for extracting time-space features of the data to generate dynamic safety boundary parameters; an analysis engine for analyzing external power requests and fusing a power allocation constraint set generated by the safety boundary; a strategy generation engine for synchronously solving an optimal power output trajectory and a thermal management control instruction in the constraint set; and an execution engine for executing power driving and heat dissipation adjustment. The system realizes accurate perception of the internal state of the battery and early warning of safety risks through the dynamic safety boundary, and ensures efficient and stable power output of the battery within the safety boundary through collaborative control of power and thermal management, thereby improving the safety and energy efficiency of the overall system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of battery management system, in particular to an intelligent battery management and safety control system and electric flame cartridge stove. BACKGROUND

[0002] The current battery management system generally adopts a protection strategy based on fixed threshold. This kind of system monitors the basic parameters such as voltage, current and surface temperature of the battery through discrete sensors, and sets a single safety upper limit for it. When any parameter exceeds the pre-set fixed threshold, the system will take protective action such as cutting off the circuit. The limitation of this method lies in its static and isolated judgment logic, which cannot perceive the complex electrochemical state changes and spatial and temporal dynamic characteristics of uneven temperature distribution inside the battery.

[0003] Static threshold management cannot cope with the comprehensive effects of battery aging, working condition fluctuations and internal inconsistency. The health status and safety margin of the battery are dynamically changed with its internal electrochemical reaction, temperature distribution and historical load, and the fixed boundary parameter is either too conservative leading to performance waste or cannot timely identify potential risks such as internal short circuit and lithium precipitation under boundary conditions. The existing technology lags behind in judging risks, and often responds passively after the abnormality has occurred, lacking the ability of early warning and self-adaptive adjustment.

[0004] At the system control level, power output management and thermal management are usually two independent control loops. The power distribution module decides the output power according to the external demand instruction, while the thermal management system is passively started to dissipate heat after the temperature rises caused by the power output. This series-type and responsive control architecture leads to delayed and fragmented control. The system is difficult to respond to high power demand while prospectively coordinating heat dissipation resources, so it is easy to fall into a vicious cycle of "performance limitation- overheating-reduction of performance" under extreme conditions, affecting user experience and accelerating battery aging. SUMMARY

[0005] The purpose of the present application is to provide an intelligent battery management and safety control system and electric flame cartridge stove to solve the problems raised in the background art.

[0006] To achieve the above purpose, the present application provides an intelligent battery management and safety control system, which comprises:

[0007] A battery state perception engine for collecting multi-physical field data through a heterogeneous sensor array embedded inside the battery pack, the heterogeneous sensor array comprising a voltage sampling unit, a current monitoring unit, a temperature distribution detection unit and an electrochemical impedance spectrum analysis unit;

[0008] A dynamic safety boundary calculation engine is configured to perform spatiotemporal feature extraction on multi-physical field data, construct a high-dimensional feature representation of the battery state, and generate dynamic safety boundary parameters.

[0009] A power demand analysis engine is configured to receive a power request signal from the range control core, decode the power mode code and priority label embedded in the signal, and generate a power allocation constraint set in combination with the dynamic safety boundary parameters.

[0010] An adaptive control strategy generation engine is configured to solve an optimal power output trajectory within the power allocation constraint set and synchronously generate a thermal management control instruction.

[0011] A power execution engine is configured to convert the optimal power output trajectory into a driving signal for the power semiconductor device and adjust the working state of the heat dissipation execution mechanism according to the thermal management control instruction.

[0012] Preferably, the battery state perception engine comprises:

[0013] The end voltage waveform data, charge and discharge current ripple characteristics, surface temperature field distribution, and electrochemical impedance phase angle of each cell in the battery pack are synchronously collected by distributed sensing nodes.

[0014] A battery state model based on a knowledge graph is constructed, and the collected raw data is mapped to a state node in a unified spatiotemporal coordinate system.

[0015] A time series convolution network is used to extract features from the state node, generating a fusion feature vector containing voltage stability indicators, current load capacity evaluation, temperature gradient change rate, and internal resistance evolution trend.

[0016] The fusion feature vector is normalized and bound with a battery identifier to generate a battery state feature matrix.

[0017] Preferably, the dynamic safety boundary calculation engine comprises:

[0018] A multi-scale graph neural network based on an attention mechanism is established, in which node features represent individual cell states and edge weights represent cell coupling strengths.

[0019] Neighbor node information is aggregated through a graph attention layer to generate a collaborative state feature of a cell cluster.

[0020] A long short-term memory network is used to perform time series prediction on the collaborative state feature, outputting a state evolution path of the battery system within a future time window.

[0021] The state evolution path is compared with a preset safety threshold to generate dynamic safety boundary parameters including the maximum allowed discharge current, instantaneous power upper limit, and temperature safety boundary.

[0022] Preferably, the power demand analysis engine comprises:

[0023] analyzing the base power demand, power variation pattern and emergency level identification in the power request signal;

[0024] calling the battery state of health evaluation model to calculate the fitness score of the current battery pack for different power demands;

[0025] converting the dynamic safety boundary parameters into power domain constraints, including power ramp rate limit, sustained output time threshold and power fluctuation tolerance;

[0026] using the constraint satisfaction algorithm to integrate the power domain constraints, fitness score and power demand characteristics to generate an optimal power allocation constraint set containing priority weights.

[0027] Preferably, the adaptive control strategy generation engine comprises:

[0028] building a reinforcement learning framework based on double-delay deep deterministic policy gradient, in which the policy network outputs power control actions and the value network evaluates the long-term value of state-action pairs;

[0029] designing a multi-objective reward function that takes into account energy conversion efficiency, temperature control effect, battery life loss and power tracking accuracy;

[0030] training the reinforcement learning framework through the experience replay mechanism, so that it can learn online and generate an optimal power output trajectory that meets the power allocation constraint set;

[0031] Based on the power output trajectory, predict the battery heat generation rate, and generate a fan speed control curve and phase change material activation strategy based on the characteristics of the heat dissipation system.

[0032] Preferably, the power execution engine comprises:

[0033] Discretize the optimal power output trajectory into a power control instruction sequence, and generate the switching timing of the power semiconductor device through the space vector modulation algorithm;

[0034] Design a heat dissipation control algorithm based on fuzzy logic to dynamically adjust the working parameters of the heat dissipation execution mechanism according to the thermal management control instructions;

[0035] Establish a cooperative control mechanism for power output and heat dissipation execution, and start active cooling measures in advance when high-power output is detected;

[0036] Real-time monitoring of the junction temperature change of the power semiconductor device, automatically derating when the safety threshold is exceeded.

[0037] Preferably, the battery state perception engine further comprises:

[0038] performing real-time anomaly detection processing on the multi-physical field data to generate a data integrity index and a signal credibility score;

[0039] performing data repair compensation on the high-dimensional feature representation based on the data integrity index to generate a calibrated battery state feature matrix;

[0040] performing time series smoothing processing on the calibrated battery state feature matrix using a sliding window mechanism to eliminate transient noise interference.

[0041] Preferably, the dynamic safety boundary calculation engine further comprises:

[0042] integrating environmental parameters collected by external environmental sensing devices, the environmental parameters including environmental temperature distribution, humidity gradient, and mechanical vibration spectrum;

[0043] performing multi-physical field coupling analysis on the environmental parameters and the state evolution path to generate environment-adaptive safety boundary parameters;

[0044] dynamically adjusting the calculation weights of the maximum allowed discharge current and the instantaneous power upper limit according to the environment-adaptive safety boundary parameters.

[0045] Preferably, the adaptive control strategy generation engine further comprises:

[0046] establishing a power output deviation monitoring mechanism to compare the difference value between the optimal power output trajectory and the actual power output sequence in real time;

[0047] when the difference value exceeds a preset tolerance threshold, activating a control strategy recalculation process to regenerate a power output trajectory based on the current battery state features;

[0048] using an incremental learning algorithm to integrate historical deviation data into the training process of the reinforcement learning framework to update the action selection policy of the policy network online.

[0049] Preferably, the present application further comprises an intelligent battery management and safety control electric flame cartridge stove, the electric flame cartridge stove comprising:

[0050] a battery state perception engine for collecting multi-physical field data through a heterogeneous sensor array embedded inside a battery pack, the heterogeneous sensor array including a voltage sampling unit, a current monitoring unit, a temperature distribution detection unit, and an electrochemical impedance spectroscopy analysis unit;

[0051] a dynamic safety boundary calculation engine for performing spatiotemporal feature extraction on the multi-physical field data, constructing a high-dimensional feature representation of the battery state, and generating dynamic safety boundary parameters;

[0052] a power demand analysis engine configured to receive a power request signal from the stove control core, decode the power mode code and priority tag embedded in the signal, and generate a power allocation constraint set in combination with the dynamic safety boundary parameter;

[0053] an adaptive control strategy generation engine configured to solve an optimal power output trajectory within the power allocation constraint set and synchronously generate a thermal management control instruction;

[0054] a power execution engine configured to convert the optimal power output trajectory into a driving signal of the power semiconductor device to control the electric flame generation unit, and adjust the working state of the heat dissipation execution mechanism according to the thermal management control instruction.

[0055] Compared with the prior art, the present application has the following beneficial effects:

[0056] The multi-physical field data including electrochemical impedance spectroscopy is collected by the heterogeneous sensor array, and the spatial and temporal characteristics are extracted to construct a high-dimensional feature representation reflecting the real state of the battery. Based on this, the dynamic safety boundary parameter is generated, so that the safety boundary is no longer a fixed value, but a dynamic function that evolves with the battery health, working stress and internal environment in real time. This mechanism can more accurately depict the instantaneous safety margin of the battery and realize early detection of internal micro-failure signs. The system can dynamically tighten the boundary parameter based on the boundary parameter, and take preventive measures before the physical parameters exceed the standard, improving the safety guarantee from passive protection to active early warning and prevention, and enhancing the robustness of the system in dealing with complex working conditions and potential risks.

[0057] By fusing the power request signal from the outside with the dynamic safety boundary parameter, a power allocation constraint set is generated. The adaptive control strategy engine synchronously solves and generates the power output trajectory and the thermal management instruction within the constraint set. This method changes the traditional series and lag response of power output and heat dissipation control to parallel and forward-looking optimization. The system calculates the heat dissipation strategy required to maintain thermal balance at the same time of deciding the output power, realizing the closed-loop cooperation of energy flow and thermal management. This cross-domain real-time optimization enables the battery system to maximize the instantaneous performance within the safety boundary, while avoiding the performance drop caused by thermal management lag, ensuring the smooth and stable output power and the overall optimization of system efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0058] Figure 1 a working principle diagram of the intelligent battery management and safety control system described in the present application;

[0059] Figure 2 a flowchart of data processing of the battery state perception engine;

[0060] Figure 3A flow chart for dynamic security boundary computation engine state prediction and boundary generation;

[0061] Figure 4 A system diagram for power demand intelligent analysis and dynamic allocation optimization;

[0062] Figure 5 A system diagram for adaptive power control and thermal management collaborative optimization. DETAILED DESCRIPTION

[0063] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work are within the scope of protection of the present application.

[0064] Please refer to Figure 1 The present application provides an intelligent battery management and safety control system, which comprises: a battery state perception engine that collects multi-physical field data through a heterogeneous sensor array embedded in a battery pack, the array comprising a voltage sampling unit, a current monitoring unit, a temperature distribution detection unit and an electrochemical impedance spectrum analysis unit, these units being distributedly arranged at key positions of the battery pack, a dynamic security boundary computation engine that extracts spatio-temporal features from the collected multi-physical field data, constructs a high-dimensional feature representation of the battery state, and generates dynamic security boundary parameters based on the representation, a power demand analysis engine that receives a power request signal from a stove control core, decodes the power mode code and priority label embedded in the signal, and generates a power allocation constraint set in combination with the dynamic security boundary parameters, an adaptive control strategy generation engine that solves an optimal power output trajectory within the power allocation constraint set and synchronously generates a thermal management control instruction, and a power execution engine that converts the optimal power output trajectory into a driving signal of a power semiconductor device and adjusts the working state of a heat dissipation execution mechanism according to the thermal management control instruction, the whole system being modularly designed, the engines communicating through a high-speed data bus to realize real-time data processing and closed-loop control.

[0065] Embodiment 1: Please refer to Figure 2In a specific implementation, the implementation of the battery state perception engine relies on the precise layout of distributed sensing nodes and the synchronous acquisition mechanism. The distributed sensing nodes are connected to a central acquisition coordinator in a star topology. The central acquisition coordinator periodically broadcasts a synchronization signal to ensure that all distributed sensing nodes can collect the end voltage waveform data, charge and discharge current ripple characteristics, surface temperature field distribution, and electrochemical impedance phase angle of each cell in the battery pack at the same time. The end voltage waveform data is collected through a high-precision analog-to-digital converter with a sampling rate set to no less than 100 kSPS to capture the instantaneous fluctuations of the voltage. The charge and discharge current ripple characteristics are measured using a closed-loop Hall effect sensor with a bandwidth covering the DC to 100 kHz frequency range, which can accurately reflect the high-frequency ripple component of the current. The surface temperature field distribution is obtained by a micro infrared thermocouple array arranged on the main surface of each cell. The thermocouple array is distributed in a grid pattern with a spatial resolution of millimeters. The analysis of the electrochemical impedance phase angle is completed by a dedicated impedance spectrum analysis unit, which injects a small amplitude AC excitation signal with a frequency range of 10 mHz to 10 kHz into the cell and simultaneously measures the voltage and current response to calculate the phase angle information. All raw data is preliminarily formatted at the distributed sensing nodes and transmitted to the main processing unit of the battery state perception engine through an isolated CAN bus network.

[0066] In a specific implementation, constructing a battery state model based on a knowledge graph is the core step of mapping the collected raw data to state nodes in a unified space-time coordinate system. The knowledge graph uses the Resource Description Framework as the data model, where each state node corresponds to a uniform resource identifier, and the node attributes include the timestamp, cell physical location coordinates, and measurement values. The mapping process first aligns the data from different distributed sensing nodes in time, using the timestamp information from the central acquisition coordinator to eliminate errors caused by network transmission delays. Then, spatial registration is performed to associate each measurement value to a specific cell and location in the three-dimensional model of the battery pack. In the knowledge graph, state nodes are connected by edges representing physical adjacency, electrical connection, or thermal coupling. The edge attributes can include distance, connection resistance, or thermal conductivity. The constructed knowledge graph is stored in a graph database, supporting efficient adjacency queries and path traversal operations.

[0067] In a specific implementation, a time series convolutional network is used to extract features of the state nodes, generating a fusion feature vector containing voltage stability indicators, current load capacity assessment, temperature gradient change rate, and internal resistance evolution trend. The time series convolutional network architecture includes an input layer, multiple dilated convolutional layers, batch normalization layers, activation function layers, and a fully connected output layer. The input layer receives the state node sequence from the knowledge graph, with a fixed sequence length of 64 most recent sampling time data; the dilated convolutional layer uses dilated convolution kernels with increasing expansion factors to exponentially expand the receptive field, effectively capturing long-term dependencies in time series without significantly increasing the number of parameters; each dilated convolutional layer is followed by a batch normalization layer to speed up the training process, and a ReLU activation function is used to introduce nonlinearity; the final fully connected output layer maps the learned time features to a fixed-dimensional fusion feature vector. Voltage stability indicators are calculated by analyzing the fluctuation variance and decay trend of the voltage sequence; current load capacity assessment is based on the amplitude and frequency components of current ripple characteristics; temperature gradient change rate is obtained by differential operation of temperature field data in space and time dimensions; internal resistance evolution trend is extracted from the curve of electrochemical impedance phase angle change with frequency.

[0068] In a specific implementation, the fusion feature vector is normalized and bound with a battery identity identifier to generate a battery state feature matrix. The normalization process uses the min-max scaling method to linearly transform the numerical values in each dimension of the fusion feature vector to the [0, 1] interval, and the scaling parameters are pre-calculated based on historical operation data of the battery type. The normalized fusion feature vector is bound with a globally unique battery identity identifier, which is written into the read-only memory at the time of battery pack production, containing information such as battery chemistry, rated capacity, production batch, etc. The binding process generates a two-dimensional matrix data structure, with rows corresponding to different feature dimensions and columns corresponding to different sampling times. The battery identity identifier is stored as metadata along with the matrix itself. The battery state feature matrix is finally packaged into a specific format of data packet and delivered to the dynamic safety boundary calculation engine through shared memory.

[0069] In a specific implementation, the battery state perception engine also performs real-time anomaly detection processing on the multi-physical field data to generate a data integrity index and a signal credibility score. The anomaly detection processing uses an unsupervised learning algorithm based on an isolation forest. The algorithm independently constructs an isolation forest model for each sensor data stream. The isolation forest model constructs multiple isolation trees by randomly selecting features and randomly selecting split points. In the training phase, historical normal data is used. In the inference phase, the path length of each new data point is calculated. Data points with significantly short path lengths are determined as abnormal points. The data integrity index is calculated as the ratio of the number of abnormal points to the total number of data points for a single sensor within a recent time window. The signal credibility score is calculated by weighting the data integrity index, the sensor's own health status, and the signal's physical rationality.

[0070] In a specific implementation, based on the data integrity index, data repair compensation is performed on the high-dimensional feature representation to generate a calibrated battery state feature matrix. The data repair compensation logic triggers different compensation strategies according to the threshold of the data integrity index. When the data integrity index is higher than the preset threshold, the data quality is considered good, and only individual abnormal data points are compensated by linear interpolation. When the data integrity index is in a moderate range, a smoothing prediction algorithm based on Kalman filtering is started to optimally estimate missing or abnormal data using a state space model. When the data integrity index is too low, redundant sensor data or predicted data based on the cell model is switched to. The calibrated battery state feature matrix marks the repaired data points and attaches a repair confidence.

[0071] In a specific implementation, a sliding window mechanism is used to perform time series smoothing processing on the calibrated battery state feature matrix to eliminate transient noise interference. The sliding window mechanism defines a fixed-length time window that slides with the arrival of new data. The smoothing processing is performed within the window. For each feature sequence in the battery state feature matrix, a Savitzky-Golay filter is applied for convolution smoothing. The Savitzky-Golay filter is a filter based on local polynomial least squares fitting, which can effectively smooth noise while preserving the true characteristics of the signal (such as peak value and width). The length of the sliding window and the polynomial order, window length, and other parameters of the Savitzky-Golay filter can be adaptively adjusted according to the current sampling frequency and noise characteristics to balance the smoothing effect and signal response speed. The processed battery state feature matrix has its transient glitches suppressed, and the data curve is smoother, which facilitates stable and reliable state evaluation and decision-making by subsequent engines.

[0072] Embodiment 2: see Figure 3In a specific implementation, the establishment of the dynamic security boundary calculation engine based on the multi-scale graph neural network with attention mechanism is a core link. The node features of the graph neural network are directly derived from the battery state feature matrix generated by the battery state perception engine. Each node feature vector contains the voltage, current, temperature, and impedance features of a single cell at the current time. The initialization of the edge weight is based on the physical design drawing of the battery pack. The Euclidean distance between cells is calculated and substituted into the pre-defined thermal coupling coefficient formula for calculation. The coefficients of the formula are determined based on the chemical properties of the cells and the thermal conductivity of the packaging materials. The graph neural network includes multiple graph attention layers. Each graph attention layer performs aggregation and update of node information. The graph attention layer uses a multi-head attention mechanism with a default configuration of 8 attention heads. Each attention head independently calculates the attention weight of the adjacent nodes to the center node. The calculation of the attention weight uses the LeakyReLU activation function. The neighbor node information after weighted summation is concatenated with the original feature of the center node, and a learnable linear transformation matrix is used for dimension reduction to generate the updated node feature representation. This process is progressive from capturing the interaction of local adjacent cells to finally perceiving the global state of the entire battery pack.

[0073] In a specific implementation, the graph attention layer is used to aggregate neighbor node information to generate the cooperative state feature of the cell cluster. The cooperative state feature is a high-dimensional vector with the same dimension as the individual node feature, but it contains the mutual influence information between the cells in the battery module. The generation process of the cooperative state feature is to perform graph pooling operation on all node features processed by multiple layers of graph attention layers. The graph pooling operation uses the global average pooling method, which calculates the average value of the feature vectors of all nodes by element to obtain a fixed-size comprehensive feature vector that can represent the current cooperative state of the entire battery pack. This cooperative state feature vector not only contains the individual health status of each cell, but more importantly, it encodes the cluster effect information such as stress and thermal runaway propagation risk generated by the inconsistency between cells, providing rich input for subsequent time series prediction.

[0074] In a specific implementation, the long short-term memory network is used to predict the state evolution path of the battery system in the future time window based on the collaborative state features. The input of the long short-term memory network is the collaborative state feature vector of a past time sequence, and the time step is usually set to 60 historical sampling points. The network structure includes an input layer, two hidden layers, and an output layer. The forget gate, input gate, and output gate of the long short-term memory network cooperate together to control the retention and forgetting of information in the cell state, and learn the dynamic pattern of the collaborative state features in the time dimension. After training, the network can predict the collaborative state features at multiple future time steps in a recursive manner based on the past state sequence, thereby forming a complete battery system state evolution path. This path predicts the future trends of key parameters such as voltage, temperature, and internal pressure.

[0075] In a specific implementation, the state evolution path is compared with the preset safety threshold to generate dynamic safety boundary parameters including the maximum allowed discharge current, the upper limit of instantaneous power, and the temperature safety boundary. The preset safety threshold is stored in a configurable database, and the threshold is dynamically set according to the chemical system, model specification, and historical operation data of the battery. The comparison process is a time-step-by-time-step scanning process. The algorithm compares each future state point on the predicted path with the safety threshold at the corresponding time, focusing on monitoring whether the voltage is lower than the minimum discharge voltage, whether the temperature exceeds the material tolerance limit, and whether the predicted temperature rise rate is abnormal. Based on the most stringent constraint condition, the algorithm reversely calculates the maximum discharge current value allowed at the current time, the peak upper limit of instantaneous power, and the temperature safety boundary that must be maintained to ensure that the battery does not violate any safety threshold within the entire prediction time window. These parameters together constitute the dynamic safety boundary parameter set.

[0076] In a specific implementation, the dynamic safety boundary calculation engine also integrates environmental parameters collected by external environmental sensing devices, including environmental temperature distribution, humidity gradient, and mechanical vibration spectrum. The environmental temperature distribution is measured by digital temperature sensors arranged at multiple positions on the battery pack shell, the humidity gradient is obtained by a capacitive humidity sensor array, and the mechanical vibration spectrum is collected by a three-axis MEMS accelerometer and then subjected to fast Fourier transform to obtain the frequency spectrum information. These environmental parameters are read in real time into the memory of the dynamic safety boundary calculation engine through the I2C digital bus, and are time-synchronized with the internal state data of the battery.

[0077] In a specific implementation, the environmental parameters are coupled with the state evolution path for multi-physics analysis to generate environment-adaptive safety boundary parameters. The coupling analysis employs a numerical simulation module based on the finite element method, which establishes a simplified battery pack electro-thermal-mechanical multi-physics model. The model takes the internal state evolution path prediction of the battery as the internal heat source, the environmental temperature distribution as the boundary condition, and the mechanical vibration spectrum as the structural load input. By solving the coupled partial differential equations, the specific effects of external environmental factors on the internal electrochemical state, thermal distribution, and mechanical stress of the battery are simulated. The simulation results are used to correct the previously internal state prediction-based state evolution path, for example, higher environmental temperature can cause the battery to have difficulty dissipating heat, thereby requiring the predicted temperature curve to be adjusted downward, ultimately generating a set of more realistic operating condition-adaptive safety boundary parameters.

[0078] The calculation weight of the maximum allowable discharge current and the instantaneous power upper limit is dynamically adjusted according to the environment-adaptive safety boundary parameters. This adjustment process is realized through a fuzzy reasoning system. The input of the fuzzy reasoning system is the severity evaluation value of the environmental parameters, such as the deviation of the environmental temperature from the ideal working temperature, the vibration intensity level, etc. These input values are fuzzified into language variables such as "low", "medium", "high", etc. The fuzzy reasoning system has a set of fuzzy rules defined by expert knowledge, such as "if the environmental temperature is high and the vibration intensity is high, then the weight factor of the maximum allowable discharge current should be significantly reduced". Through fuzzy reasoning and defuzzification calculation, the system outputs specific adjustment coefficients for the calculation weight of the maximum allowable discharge current and the instantaneous power upper limit, so that in harsh environments, the calculation of safety boundary parameters will be more conservative, prioritizing battery safety, while in ideal environments, more battery performance potential can be released.

[0079] In a specific implementation, the implementation of the power demand analysis engine begins with the reception and analysis of the power request signal from the stove control core. The power request signal is transmitted in the form of a data frame through a high-speed serial peripheral interface, and the data frame is encapsulated using a communication protocol. The analysis process first performs a cyclic redundancy check to ensure data integrity, and after passing the check, the payload field in the data frame is extracted. The payload field contains encoded information of the basic power demand, power change mode, and emergency level identification. The basic power demand is represented as a floating-point number indicating the requested power value. The power change mode is represented by a predefined enumeration code, for example, constant power mode corresponds to code 1, pulse power mode corresponds to code 2, and ramp-up power mode corresponds to code 3. The emergency level identification is an integer value ranging from 1 to 5, with a larger value representing a higher priority request. The decoding algorithm converts the binary encoding into the standardized data structure required for internal processing according to the protocol specification.

[0080] In a specific implementation, the battery health state assessment model is called to calculate the fitness score of the current battery pack for different power demands, which is a pre-trained support vector machine classifier, and the model file is stored in the non-volatile memory. The input feature vector of the support vector machine classifier comes from the latest timestamp data of the battery state feature matrix generated by the battery state perception engine, and the feature vector contains the normalized values of voltage stability indicators, current load capacity assessment, temperature gradient change rate and internal resistance evolution trend. The support vector machine classifier calculates a fitness score between 0 and 1 based on these features, and the closer the score is to 1, the more the current health state of the battery can adapt to high power demand, and the closer the score is to 0, the more the battery has limiting factors that need to be handled carefully.

[0081] In a specific implementation, converting dynamic safety boundary parameters into power domain constraints is a key mapping process, which includes maximum allowed discharge current, instantaneous power upper limit and temperature safety boundary. The conversion process uses a set of linear transformation formulas, such as multiplying the maximum allowed discharge current by the rated voltage of the battery pack to get the reference value of the power rise rate limit, directly using the instantaneous power upper limit as the instantaneous power upper limit in the power domain constraint, and getting the allowed continuous output time threshold value by backstepping the temperature safety boundary through the battery thermal model. The power fluctuation tolerance is calculated by analyzing the current and voltage change rate limit in the dynamic safety boundary parameters. The converted power domain constraints form a constraint set, including power rise rate limit, continuous output time threshold and power fluctuation tolerance, which define the limits on the power operating space in the form of numerical ranges.

[0082] In a specific implementation, a constraint satisfaction algorithm is used to integrate the power domain constraints, fitness score and power demand characteristics to generate an optimal power allocation constraint set containing priority weights. The constraint satisfaction algorithm models the power allocation problem as a constraint satisfaction problem, where the variables are the power parameters to be allocated, the constraints include the converted power domain constraints, the soft constraints based on the fitness score, and the boundary conditions defined by the power demand characteristics. The algorithm uses a backtracking search strategy to find a solution that satisfies all hard constraints and maximizes the fitness score and priority matching degree in the solution space. During the search process, the urgency identifier directly affects the allocation of priority weights, and requests with high urgency will be given higher weights, so that the final solution is more inclined to meet their demands. The output of the optimal power allocation constraint set is a structure, which contains the allowed power range, change rate limit, duration and priority weight of each constraint.

[0083] In some embodiments, the parsing of the power request signal can involve more complex signal features, for example, the power request signal can contain a power demand curve in the form of a time series instead of a single numerical value, in which case the parsing engine needs to extract key feature parameters of the curve such as peak power, average power, duration pattern, etc. The encoding of the power variation pattern can also be extended to a more fine-grained classification, for example, to distinguish between different duty cycle pulse patterns or different slope ramp patterns. The parsing of the urgency level identifier can be linked to a system pre-defined policy template, different levels of identifier trigger different degrees of constraint relaxation. The implementation of the battery health state assessment model can have different choices, in addition to the support vector machine classifier, a deep neural network or a random forest machine learning model can also be used. The calculation of the fitness score can not only be based on the current state, but also introduce a prediction of the historical performance decay trend, making the score forward-looking. The selection of the model input features can be optimized according to the battery type, for example, for lithium iron phosphate batteries, more attention can be paid to voltage platform stability, while for ternary lithium batteries, more attention can be paid to temperature sensitivity.

[0084] In some embodiments, the specific implementation of the constraint satisfaction algorithm can use different search strategies, when the solution space is large, heuristic search algorithms such as genetic algorithms can be used to speed up the solution process. For scenarios with extremely high real-time requirements, the algorithm can be simplified to a rule-based speed priority matching, sacrificing part of the optimality to ensure response speed. The calculation method of the priority weight can be designed as a nonlinear function, so that the weight of the urgent request presents an exponential growth, thereby ensuring that high-priority requests are fully met. Optionally, the generation process of the power allocation constraint set can introduce a safety buffer coefficient, which is dynamically adjusted according to the confidence of the battery health state assessment model. When the model has low confidence in calculating the fitness score, the safety buffer coefficient is automatically increased, making the generated power allocation constraint set more conservative and enhancing the robustness of the system. The adjustment of the buffer coefficient follows a pre-defined lookup table, which defines the mapping relationship between the confidence interval and the coefficient value.

[0085] It can be understood that each processing step of the power demand parsing engine needs to be completed within a strict time window, usually requiring the whole process from signal reception to constraint set generation to be completed within a control cycle (e.g. 10 milliseconds). It can be understood that the mapping relationship between the power domain constraint conditions and the physical boundary parameters needs to be calibrated according to the specific battery management system hardware characteristics, for example, the switching characteristics of power semiconductor devices, the efficiency of the heat dissipation system and other factors will affect the coefficient setting in the mapping formula. The calibration process is usually completed in the system initialization stage by running a series of diagnostic tests.

[0086] The formula representing the power mapping in the power demand parsing engine is:

[0087]

[0088] wherein: represents the calculated power domain limit value, represents the maximum allowed discharge current in the dynamic safety margin parameter, represents the rated voltage of the battery pack, represents the power mapping efficiency coefficient, which takes into account both line losses and conversion efficiency. The power mapping efficiency coefficient ranges from 0 to 1, and its specific value is determined through preliminary system calibration experiments and stored in the configuration file for real-time query. This formula ensures the physical consistency of the conversion from the current boundary to the power boundary.

[0089] Referring to Figure 4 , the complete processing flow of the system for the power request signal from the stove control core is presented, including the reception and analysis process of the power request signal. Through three main curves, the actual value change of the basic power demand, the optimized power value calculated by the system according to the battery health state evaluation model, and the power domain safety margin converted from the dynamic safety margin parameter are displayed. The power request curve in the chart reflects the characteristics of different power change modes, including constant power mode, pulse power mode and ramp-up power mode, and smooth transitions of various working states. The allocated power curve reflects the optimization effect of the constraint satisfaction algorithm, which maximizes the satisfaction of power demand under the premise of ensuring system safety. The safety margin curve shows the safety boundary maintaining ability of the system under different working conditions, providing an important constraint condition for power allocation decision. The entire chart vividly presents the complete technical process of how the power demand analysis engine receives data frames through the high-speed serial peripheral interface, performs cyclic redundancy check to ensure data integrity, extracts the encoded information in the payload field, and finally generates an optimal power allocation constraint set containing priority weights.

[0090] In a specific implementation, the implementation of the adaptive control strategy generation engine is based on a reinforcement learning framework of double-delay deep deterministic policy gradient, which includes two core components, a policy network and a value network. The policy network is responsible for generating power control actions according to the current system state, and the value network is used to evaluate the long-term value of state-action pairs to guide policy optimization. Both the policy network and the value network use a deep neural network structure, the input layer receives the optimal power distribution constraint set from the power demand analysis engine and the battery state feature matrix from the battery state perception engine, the hidden layer is composed of multiple fully connected layers, and the ReLU activation function is used to introduce nonlinearity, and the output layer generates a continuous action vector and a scalar value estimate respectively. The double-delay deep deterministic policy gradient algorithm stabilizes the training process by introducing two target value networks and a delayed policy update mechanism. The target value network is a slow update copy of the value network, used to calculate the target value to reduce estimation bias, and the update frequency of the policy network is lower than that of the value network to reduce volatility risk. The reinforcement learning framework runs on an embedded computing platform, using GPU to accelerate the forward inference and back propagation calculation of the neural network.

[0091] In a specific implementation, designing a multi-objective reward function is a key step in the reinforcement learning framework. The multi-objective reward function needs to consider multiple performance indicators such as energy conversion efficiency, temperature control effect, battery life loss, and power tracking accuracy. The specific form of the reward function is a weighted sum, each objective corresponds to a reward item and a weight coefficient. The energy conversion efficiency reward item is calculated based on the ratio of actual output power to input power, encouraging efficient energy utilization. The temperature control effect reward item is negatively related to the degree of deviation of the battery temperature from the ideal working interval, promoting thermal management. The battery life loss reward item estimates the life decay according to the current stress and temperature stress model, punishing accelerated aging operations. The power tracking accuracy reward item measures the error between the actual power output and the requested power, ensuring response accuracy. The weight coefficient is initialized through domain knowledge and can be fine-tuned during training to balance the priority of different objectives. See Table 1 for the weight coefficients of the multi-objective reward function.

[0092] Table 1: Reward function weight configuration table

[0093]

[0094] In specific implementations, the reinforcement learning framework is trained by an experience replay mechanism to enable online learning and generate optimal power output trajectories that satisfy the set of power allocation constraints. The experience replay mechanism uses a fixed-size circular buffer to store experience tuples of agent-environment interactions, each containing the current state, the action taken, the immediate reward, the next state, and a terminal flag. During training, a small batch of tuples is randomly sampled from the buffer to update the network parameters. This random sampling breaks the correlation between data, improving the stability and efficiency of learning. The generation of optimal power output trajectories is the process of the policy network outputting a sequence of actions given the current state. The sequence of actions needs to satisfy the boundary conditions defined by the set of power allocation constraints. Constraint handling is achieved through projection, which projects the original action output by the network into the feasible region.

[0095] In specific implementations, the battery heat generation rate is predicted based on the power output trajectory and combined with the characteristics of the heat dissipation system to generate a fan speed control curve and a phase change material activation strategy. The prediction of the heat generation rate uses a simplified thermal model, which takes the current and voltage data in the power output trajectory as input and outputs the heat generation power of the battery pack. The thermal model can be represented as:

[0096]

[0097] where: represents the heat generation rate, represents the operating current, represents the internal resistance of the battery, represents the polarization voltage. The fan speed control curve is calculated by PID control based on the predicted heat generation rate and current temperature feedback to generate a speed setpoint. The phase change material activation strategy is based on temperature threshold judgment. When the predicted temperature exceeds the phase change point of the phase change material, the phase change material layer is triggered to absorb excess heat.

[0098] In some embodiments, the specific architecture of the policy network and the value network can be adjusted according to the limitations of computing resources. For resource-constrained embedded systems, the network structure can be simplified to fewer layers and fewer neurons. For example, the policy network only contains one hidden layer with 64 neurons, and the value network contains two hidden layers with 32 and 16 neurons, respectively. The training algorithm can also use a distributed version, with multiple agent instances collecting experience data in parallel and a centralized parameter server aggregating gradient updates to the model to speed up training convergence. The management strategy of the experience replay buffer can be diversified, such as using a priority experience replay mechanism that assigns a sampling priority to each experience tuple based on the size of the time difference error. Tuples with larger errors have a higher probability of being sampled, thereby improving learning efficiency. The size of the buffer capacity needs to balance the diversity of memory and the proportion of old and new experiences. Typically, it is set to cover tens of thousands of experience data at time steps.

[0099] The power output deviation monitoring mechanism compares the difference value between the optimal power output trajectory and the actual power output sequence in real time. The deviation monitoring mechanism collects the actual output data of the power execution engine through a high-precision power sensor to form an actual power output sequence. The sequence is aligned with the optimal power output trajectory generated by the strategy network on the same time axis. The difference value is calculated using the root mean square error method. The deviation of each sampling point is squared and averaged, and then the square root is taken to obtain a comprehensive deviation index. The deviation monitoring runs at a fixed control period, for example, the difference value in the current time window is calculated once every 100 milliseconds. When the difference value exceeds the preset tolerance threshold, the control strategy recalculation process is activated to regenerate the power output trajectory based on the current battery state characteristics. The preset tolerance threshold is dynamically set according to the urgency of the power demand and the system safety requirements. For example, for high-priority requests, the tolerance threshold is smaller, requiring more stringent tracking accuracy. The control strategy recalculation process interrupts the current power output and immediately calls the reinforcement learning framework for re-inference. The re-inference process uses the latest battery state feature matrix and power allocation constraint set as input. The strategy network quickly calculates the new power output trajectory through forward calculation. The recalculation process requires to be completed within a very short time, for example, no more than 10 milliseconds, to ensure the real-time response of the system.

[0100] The incremental learning algorithm is used to integrate historical deviation data into the training process of the reinforcement learning framework to update the action selection policy of the strategy network online. The incremental learning algorithm uses the online gradient descent method. After receiving new deviation data each time, the strategy network and the value network are updated by one step of gradient. The update step is dynamically adjusted according to the learning rate scheduler. Historical deviation data is converted into additional reward signals or constraint conditions to be integrated into training. For example, larger deviations will produce negative rewards, encouraging the strategy to avoid similar actions. Incremental learning allows the reinforcement learning framework to continuously adapt to system dynamics, such as battery aging or load characteristic changes. It can be understood that the recalculation trigger condition of the power output deviation monitoring mechanism needs to avoid too frequent switching to avoid causing control oscillation, so the setting of the tolerance threshold usually has a certain hysteresis characteristic, for example, only when the deviation exceeds the threshold for a period of time, the recalculation is triggered. It can be understood that the learning rate of the incremental learning algorithm needs to be carefully adjusted. Too high learning rate may lead to unstable strategy, while too low learning rate will make the adaptation process too slow. The learning rate usually adopts a decay strategy, gradually decreasing with the increase of training steps.

[0101] Referring to Figure 5, which shows the operation effect of the adaptive control strategy generation engine based on the double-delay deep deterministic policy gradient reinforcement learning framework. The chart adopts a double-Y-axis design, with the main Y-axis showing the power control-related parameter changes and the secondary Y-axis showing the thermal management-related parameter dynamics, fully presenting the control effect of multi-objective collaborative optimization. The optimal power output trajectory curve in the figure reflects the optimization decision-making ability of the reinforcement learning framework under the premise of meeting the power allocation constraint set, with smooth trajectory and rapid response to system state changes. The actual power output curve shows the tracking performance of the power execution engine on the control command, and the slight difference between the two reflects the control accuracy of the system. The fan speed control curve shows the active thermal management strategy of the system according to the heat generation rate predicted by the power output trajectory, embodying the idea of forward-looking control. The heat generation rate curve reflects the heat generation characteristics of the battery under different power output conditions. The overall chart shows how the adaptive control strategy generation engine considers multiple performance indicators such as energy conversion efficiency, temperature control effect, battery life loss, and power tracking accuracy, and how it learns and generates optimal power output trajectories that meet complex constraint conditions online through the experience replay mechanism, while simultaneously generating thermal management control instructions based on the heat model prediction.

[0102] In a specific implementation, the core task of the power execution engine is to convert the optimal instructions issued by the adaptive control strategy generation engine into precise control of power semiconductor devices and heat dissipation execution mechanisms. The optimal power output trajectory is a continuous function with time as the independent variable. The power execution engine first needs to discretize this continuous trajectory into a power control instruction sequence. The discretization process is based on a fixed control period for sampling, and each sampling point corresponds to a power set value, thereby generating a digitized instruction sequence. This power control instruction sequence is then sent to the space vector modulation algorithm module, which is specialized for the control of three-phase inverters. The algorithm calculates the required voltage vector based on the current instruction power value, determines the conduction and shutdown state of the six power semiconductor devices in the next switching period by querying the pre-stored switching table, and generates precise pulse width modulation waveforms. The calculation of switching timing aims to minimize harmonic distortion and improve direct current voltage utilization.

[0103] In specific implementations, a heat dissipation control algorithm based on fuzzy logic dynamically adjusts the working parameters of the heat dissipation execution mechanism according to the thermal management control instructions, which include the fan speed control curve and the phase change material activation strategy predicted by the adaptive control strategy generation engine. The input variables of the fuzzy logic heat dissipation control algorithm include "temperature deviation" and "temperature change rate", and the output variables are "fan speed adjustment" and "pump speed adjustment". The fuzzification process converts precise input values into fuzzy language variables such as "negative large", "negative small", "zero", "positive small", "positive large", etc., conducts reasoning based on a series of fuzzy rules, and finally converts the fuzzy output into precise control quantities through defuzzification methods, thereby dynamically adjusting the speed of the fan or liquid cooling pump.

[0104] In specific implementations, the establishment of a cooperative control mechanism for power output and heat dissipation execution is the key to achieving efficient thermal management. The cooperative control mechanism is essentially a feedforward control strategy. When the power execution engine analyzes the optimal power output trajectory and detects an upcoming high-power output segment, the cooperative control mechanism sends a pre-start instruction to the heat dissipation execution mechanism in advance. The content of the instruction is determined based on historical data or model prediction, such as increasing the fan speed to a higher base level 200 milliseconds in advance, rather than waiting for the temperature sensor to detect temperature rise before acting. This proactive strategy of starting active cooling measures in advance can effectively counteract the thermal inertia caused by high power output, keeping the battery temperature within the safety boundary.

[0105] In specific implementations, real-time monitoring of the junction temperature change of power semiconductor devices and automatic derating when it exceeds the safety threshold are important steps to ensure hardware safety. Junction temperature monitoring is done through thermistors embedded in IGBT or MOSFET modules or indirectly estimated using the known relationship between on-state voltage drop and temperature. The monitoring circuit samples the temperature signal at a high frequency, and the junction temperature safety threshold is set according to the specifications of the power semiconductor device, usually 150 degrees Celsius. When the real-time monitored junction temperature approaches or exceeds the safety threshold, the derating logic is triggered. Derating usually uses linear or segmented linear strategies to proportionally reduce the power set value in the power control instruction sequence, thereby reducing the current through the power semiconductor device and reducing heat generation, ensuring that the junction temperature returns to the safety interval. The power derating ratio can be calculated by the following formula:

[0106]

[0107] Where: represents the power derating ratio, represents the real-time monitored junction temperature, represents the junction temperature safety threshold, represents the derating coefficient (a positive number less than 1, for example 0.02 / degree Celsius). When Time, No derating.

[0108] In some embodiments, the implementation of the space vector modulation algorithm can consider different optimization objectives, for example, to reduce switching loss, a discontinuous space vector modulation strategy can be adopted, which forces certain phase leg switching devices to maintain a constant state within each sampling period, thereby reducing one-third of the switching times. Another optimization direction is the overmodulation strategy, when the instruction voltage vector amplitude exceeds the maximum linear output capability of the inverter, by modifying the trajectory of the reference voltage vector, it can tolerate higher harmonic content while ensuring the amplitude of the fundamental component, to expand the output voltage range. The design of the fuzzy logic cooling control algorithm can be more refined, for example, more input variables such as "ambient temperature", "current fan speed" or "battery health status" can be introduced, so that the control rules can better adapt to complex working conditions. The shape and range of the membership function can also be adjusted and optimized according to the characteristics of the specific cooling system to obtain a smoother control response.

[0109] Optionally, the feedforward action in the cooperative control mechanism can be based on a simple linear model, for example, the pre-start cooling power is proportional to the predicted power increment, and the proportional coefficient is obtained through system identification experiments. More complex nonlinear mapping relationships can also be used, such as lookup tables or neural networks, to more accurately predict the required pre-cooling intensity under different power steps. In addition to using thermistors, the junction temperature monitoring of power semiconductor devices can also use a physical model based on heat conduction principles for soft measurement, which takes power loss (on-state loss and switching loss) as input, considers the thermal resistance and thermal capacity of the heat sink, and estimates the junction temperature in real time. This method can not rely on hardware sensors and provides a redundant monitoring method.

[0110] It can be understood that each sub-module of the power execution engine needs to complete the calculation and execution in a very short time, usually requiring all processing to be completed within one switching cycle, so the algorithm needs to be highly optimized to meet the real-time requirements. It can be understood that the effectiveness of the synergistic control mechanism of power output and heat dissipation execution is heavily dependent on the accuracy of the optimal power output trajectory prediction. If the predicted trajectory deviates too much from the actual power demand, it may lead to insufficient or excessive heat dissipation in advance, so it needs to work closely with the upstream adaptive control strategy generation engine. In specific implementation, the hardware platform of the power execution engine usually contains a digital signal processor or a high-performance microcontroller. The digital signal processor is responsible for executing tasks that require high-speed mathematical operations such as space vector modulation, while the microcontroller can handle logical control tasks such as fuzzy logic control, synergistic strategy, and junction temperature monitoring. The drive circuit of the power semiconductor device provides the necessary level conversion, isolation, and amplification functions to ensure that the control signal can safely and reliably drive high-power devices. The heat dissipation execution mechanism such as a brushless DC fan or a liquid cooling pump is controlled by a dedicated drive chip, which receives pulse width modulation signals or analog voltage signals from the microcontroller to achieve smooth speed adjustment.

[0111] It should be noted that the relational terms herein such as first and second and the like are used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any such actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0112] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. An intelligent battery management and safety control system, characterized in that, The system includes: A battery state sensing engine is used to collect multi-physics data through a heterogeneous sensor array embedded inside the battery pack. The heterogeneous sensor array includes a voltage sampling unit, a current monitoring unit, a temperature distribution detection unit, and an electrochemical impedance spectroscopy analysis unit. The dynamic safety boundary calculation engine is used to extract spatiotemporal features from multiphysics data, construct a high-dimensional feature representation of battery state, and generate dynamic safety boundary parameters. The power demand parsing engine is used to receive power request signals from the stove control core, decode the power mode code and priority mark embedded in the signal, and generate a set of power allocation constraints by combining dynamic safety boundary parameters. An adaptive control strategy generation engine is used to solve for the optimal power output trajectory within the power allocation constraint set and simultaneously generate thermal management control commands. The power execution engine is used to convert the optimal power output trajectory into drive signals for power semiconductor devices and adjust the working state of the heat dissipation actuator according to thermal management control commands.

2. The intelligent battery management and safety control system according to claim 1, characterized in that, The battery status sensing engine includes: The terminal voltage waveform data, charge and discharge current ripple characteristics, surface temperature field distribution and electrochemical impedance phase angle of each cell in the battery pack are collected synchronously through distributed sensing nodes. A knowledge graph-based battery state model is constructed, mapping the collected raw data into state nodes in a unified spatiotemporal coordinate system; A temporal convolutional network is used to extract features from the state nodes, generating a fused feature vector that includes voltage stability indicators, current load capacity assessment, temperature gradient change rate, and internal resistance evolution trend. The fused feature vector is normalized and bound to the battery identifier to generate a battery state feature matrix.

3. The intelligent battery management and safety control system according to claim 1, characterized in that, The dynamic security boundary calculation engine includes: A multi-scale graph neural network based on an attention mechanism is established, where node features represent the state of a single battery cell and edge weights represent the coupling strength between battery cells. By aggregating neighbor node information through the graph attention layer, the collaborative state characteristics of the battery cell cluster are generated. A long short-term memory network is used to predict the time series of cooperative state features and output the state evolution path of the battery system within a future time window; The state evolution path is compared with the preset safety threshold to generate dynamic safety boundary parameters that include the maximum allowable discharge current, the instantaneous power limit, and the temperature safety boundary.

4. The intelligent battery management and safety control system according to claim 1, characterized in that, The power demand analysis engine includes: Analyze the basic power demand, power change pattern, and urgency level indicator in the power request signal; Call the battery health status assessment model to calculate the current battery pack's adaptability score to different power demands; The dynamic safety boundary parameters are converted into power domain constraints, including power rise rate limits, continuous output time thresholds, and power fluctuation tolerance. The constraint satisfaction algorithm integrates power domain constraints, fitness scores, and power demand characteristics to generate an optimal power allocation constraint set that includes priority weights.

5. The intelligent battery management and safety control system according to claim 1, characterized in that, The adaptive control strategy generation engine includes: A reinforcement learning framework based on dual-delay deep deterministic policy gradient is constructed, in which the policy network outputs power to control actions, and the value network evaluates the long-term value of state-action pairs. Design a multi-objective reward function that comprehensively considers energy conversion efficiency, temperature control effect, battery life loss and power tracking accuracy; The reinforcement learning framework is trained through an experience replay mechanism, enabling it to learn online and generate the optimal power output trajectory that satisfies the power allocation constraint set. Based on the power output trajectory, the battery heat generation rate is predicted, and the fan speed control curve and phase change material activation strategy are generated by combining the characteristics of the heat dissipation system.

6. The intelligent battery management and safety control system according to claim 1, characterized in that, The power execution engine includes: The optimal power output trajectory is discretized into a power control command sequence, and the switching timing of the power semiconductor device is generated by a space vector modulation algorithm. Design a heat dissipation control algorithm based on fuzzy logic to dynamically adjust the working parameters of the heat dissipation actuator according to the thermal management control command; Establish a coordinated control mechanism for power output and heat dissipation, and activate active heat dissipation measures in advance when high power output is detected; It monitors the junction temperature changes of power semiconductor devices in real time and automatically derates when the temperature exceeds a safe threshold.

7. The intelligent battery management and safety control system according to claim 1, characterized in that, The battery status sensing engine also includes: Real-time anomaly detection processing is performed on the multiphysics data to generate data integrity indicators and signal reliability scores; Based on the data integrity index, data repair and compensation are performed on the high-dimensional feature representation to generate a calibrated battery state feature matrix. A sliding window mechanism is used to perform time-series smoothing on the calibrated battery state feature matrix to eliminate transient noise interference.

8. The intelligent battery management and safety control system according to claim 3, characterized in that, The dynamic security boundary calculation engine also includes: The system integrates environmental parameters collected by external environmental sensors, including ambient temperature distribution, humidity gradient, and mechanical vibration spectrum. The environmental parameters and the state evolution path are coupled through multiphysics analysis to generate environmentally adaptive safety boundary parameters. The calculation weights of the maximum allowable discharge current and the instantaneous power limit are dynamically adjusted based on the environmentally adaptive safety boundary parameters.

9. The intelligent battery management and safety control system according to claim 5, characterized in that, The adaptive control strategy generation engine also includes: Establish a power output deviation monitoring mechanism to compare the difference between the optimal power output trajectory and the actual power output sequence in real time. When the difference value exceeds the preset tolerance threshold, the control strategy recalculation process is activated to regenerate the power output trajectory based on the current battery state characteristics. An incremental learning algorithm is used to incorporate historical bias data into the training process of the reinforcement learning framework, and the action selection strategy of the policy network is updated online.

10. A smart battery management and safety control electric flame portable gas stove, characterized in that, The electric flame portable gas stove includes: A battery state sensing engine is used to collect multi-physics data through a heterogeneous sensor array embedded inside the battery pack. The heterogeneous sensor array includes a voltage sampling unit, a current monitoring unit, a temperature distribution detection unit, and an electrochemical impedance spectroscopy analysis unit. The dynamic safety boundary calculation engine is used to extract spatiotemporal features from multiphysics data, construct a high-dimensional feature representation of battery state, and generate dynamic safety boundary parameters. The power demand parsing engine is used to receive power request signals from the stove control core, decode the power mode code and priority mark embedded in the signal, and generate a set of power allocation constraints by combining dynamic safety boundary parameters. An adaptive control strategy generation engine is used to solve for the optimal power output trajectory within the power allocation constraint set and simultaneously generate thermal management control commands. The power execution engine is used to convert the optimal power output trajectory into drive signals for power semiconductor devices to control the electric flame generation unit, and to adjust the working state of the heat dissipation actuator according to the thermal management control command.

Images