SOC detection method and system for battery pack
By establishing the correlation and topology network of battery packs, and combining regression tree and particle swarm optimization algorithms to optimize feature weights, the accuracy and robustness issues of SOC detection in parallel battery systems are solved, enabling optimized management of battery performance and fault early warning.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XIAN YUCHI TENABLE DEFENSE EQUIP RES INST CO LTD
- Filing Date
- 2026-01-20
- Publication Date
- 2026-05-05
AI Technical Summary
Existing parallel battery systems suffer from non-uniform distribution of branch current and circulating current problems due to the difference in equivalent internal resistance between the main and auxiliary battery packs during SOC detection. Furthermore, sampling noise is easily generated during the parallel transient process. Existing algorithms lack robustness and are difficult to track energy flow in real time, resulting in a decrease in SOC identification accuracy.
By establishing the relationship between the main battery pack and the auxiliary battery pack, a two-dimensional association space and topology network are constructed. Combined with the SOC state identification model of the regression tree combination architecture, particle swarm optimization and machine learning models are introduced to optimize feature weights and anomaly thresholds and predict fault events.
It improves the accuracy of SOC detection, enables optimized management of battery performance, timely detection of potential risks and makes control decisions, and enhances the health management capabilities of the battery system.
Smart Images
Figure CN121978531A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of battery pack health management technology, specifically to a method and system for detecting the state of charge (SOC) of a battery pack. Background Technology
[0002] With the rapid development of energy storage systems, energy management technology for lithium battery packs has become the core of industry research, especially for high-voltage, high-capacity energy demands. Parallel systems consisting of multiple battery packs, including a main battery pack and several auxiliary battery packs, have been widely used. However, existing parallel battery systems still have the following drawbacks in SOC detection technology: On the one hand, traditional detection methods usually rely on measuring the total current of the parallel system and assuming that the current is evenly distributed among the branches. However, in actual operating conditions, even if the main and auxiliary battery packs in parallel exhibit the same terminal voltage, the difference in equivalent internal resistance between the main and auxiliary packs will cause impedance deviation, resulting in non-uniform distribution of branch currents. This leads to severe asynchronous drift in the calculation of SOC for each branch, which can cause serious power imbalance under long-term operation. On the other hand, during the transient process of parallel connection, if there is a significant SOC gradient between the main and auxiliary packs, For example, if the main battery pack has a SOC of 80% and the auxiliary battery pack has a SOC of 20%, a huge circulating current will be generated when they are connected in parallel, which can easily produce a lot of sampling noise. Existing SOC identification algorithms often lack effective robust design when dealing with such high-noise, saturated data, which leads to an initial deviation that is difficult to eliminate at the beginning of parallel connection. This is not conducive to achieving in-depth optimization of battery performance and health management throughout the entire life cycle. In addition, existing parallel connection schemes mostly use hard connection. When the SOC of the main and auxiliary packs are extremely inconsistent, it is difficult to track and correct the energy flow during the SOC process in real time and dynamically, resulting in a significant decrease in SOC identification accuracy after the transient connection. Summary of the Invention
[0003] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a method and system for detecting the State of Charge (SOC) of a battery pack. During parallel connection, by determining the electrical signal monitoring results of the main and auxiliary battery packs, the correlation between the main and auxiliary battery packs is established, a two-dimensional correlation space and a battery pack topology network are constructed. Combined with an SOC state identification model based on a regression tree combination architecture, node splitting features are selected according to the magnitude of the anomaly score. A particle swarm optimization algorithm is introduced to optimize feature weights and standard anomaly thresholds, and a predicted SOC label is output, improving the accuracy of SOC detection. Furthermore, a machine learning model is introduced to predict the next warning event, thus solving the problems mentioned in the background technology.
[0004] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, this application provides a method for SOC detection of a battery pack, the method comprising: determining a main battery pack and at least one auxiliary battery pack; wherein the main battery pack performs parallel connection to the auxiliary battery pack and sends a synchronous detection command to acquire the electrical signal monitoring results of each battery pack in real time; Based on the electrical signal monitoring results, the correlation between the main battery pack and the auxiliary battery pack is established. Based on the correlation, the coupling duration and spatial gradient are extracted and input into the SOC state identification model, and the estimated SOC label of each battery pack is output. The SOC state identification model is based on the regression tree combination architecture and introduces the particle swarm algorithm to adjust the feature weights and standard anomaly thresholds according to the control mode and effective coupling period. A sliding window weighted average is applied to the estimated SOC labels to generate an SOC status score. A machine learning model is then introduced to predict the next failure event.
[0005] Furthermore, the electrical signal monitoring results include the shunt current and terminal voltage of each battery pack.
[0006] Furthermore, the relationship between the main battery pack and the auxiliary battery pack is established, including: The main battery pack is mapped as the root topology node, and each sub-battery pack is mapped as a slave topology node. The cell branches inside the slave topology node are identified, and the association between the root topology node and each cell branch is established. The topology edge where each cell branch is located is configured with equivalent resistance and polarization inductance. Simultaneously, by synchronously sampling, the electrical signal monitoring results of each node at the same timestamp are obtained, the first product of equivalent resistance and shunt current and the second product of polarization inductance and current change rate are calculated, and a two-dimensional correlation space is established with the ratio of the first product to the terminal voltage as the horizontal axis and the ratio of the second product to the terminal voltage as the vertical axis.
[0007] Furthermore, the coupling duration and spatial gradient are extracted, including: Monitor the connectivity status of the topology edges; identify the contact detection results of the parallel relays and high-voltage buses corresponding to the auxiliary battery pack to determine the start time of the auxiliary battery pack's connection; identify the disconnection detection results of the parallel relays and high-voltage buses corresponding to the auxiliary battery pack to determine the end time of the auxiliary battery pack's connection; and determine the coupling duration based on the start and end times of the connection. The state evolution trajectory is plotted in the two-dimensional associated space, and the rising and falling gradients of the state evolution trajectory within the coupling duration are identified to determine the spatial gradient.
[0008] Furthermore, determining the coupling duration also includes: All shunt currents of the main battery pack and the auxiliary battery pack within the coupling time are obtained to form corresponding current sequences, including a first current sequence and a second current sequence, wherein the first current sequence corresponds to the main battery pack and the second current sequence corresponds to the auxiliary battery pack. Identify the cross-correlation coefficients of the first and second current sequences, filter and extract the coupling durations where the cross-correlation coefficients are greater than the standard coupling threshold, and obtain the effective coupling time periods.
[0009] Furthermore, the SOC state identification model includes: Extract the association between each topology node and the cell branch, and each branch must carry at least the coupling duration and spatial gradient; Set the state function group of the topology node: including the first function and the second function; obtain the corresponding feature solution matrix by parsing the state function group, and weight and map it with the feature weight of node splitting under the current access mode to generate an anomaly score; at the same time, compare the anomaly score with the standard anomaly threshold. If the anomaly score is greater than the standard anomaly threshold, the spatial gradient is selected as the node splitting feature. If the anomaly score is less than or equal to the standard anomaly threshold, the coupling duration is selected as the node splitting feature. The standard anomaly threshold is a dynamic value.
[0010] Furthermore, the access modes include direct connection and plug-in connection.
[0011] Furthermore, a particle swarm optimization algorithm is introduced to adjust feature weights and standard anomaly thresholds based on the access mode and effective coupling period, including: Construct a high-dimensional target search space; Using feature weights and standard anomaly thresholds as parameters to be optimized, the optimal parameter vector is dynamically searched within the target search space. Based on the current access mode and effective coupling period, evolutionary constraints are provided for population evolution, including: adjusting the position vector by using the access mode as a physical constraint term in the fitness function, and adjusting the velocity vector by using the proportion of effective coupling period as a search constraint term, until the maximum number of iterations is reached; the optimal parameter vector is output with the goal of minimizing the residual, and mapped to the gain of the feature weights of the regression tree node split and the offset of the standard anomaly threshold, respectively.
[0012] Furthermore, the machine learning model has a built-in first branch and a second branch, including: After outputting the predicted SOC label, the reciprocal of the residual is used as the weight coefficient to determine the SOC status score; The first branch extracts the mean, standard deviation, and timestamp from the SOC status score to construct a first feature vector; simultaneously, it retrieves the electrical signal monitoring results at the same timestamp, statistically analyzes the time-series characteristics of voltage range differences, and generates a second feature vector; feature fusion is performed on the first and second feature vectors to generate a fused feature. The fused features are input into the second branch to predict the next failure event.
[0013] Secondly, this application provides a SOC detection system for a battery pack, the system comprising: The target perception module identifies a main battery pack and at least one auxiliary battery pack; wherein the main battery pack performs parallel connection to the auxiliary battery pack and sends a synchronous detection command to obtain the electrical signal monitoring results of each battery pack in real time; The detection and analysis module establishes the correlation between the main battery pack and the auxiliary battery pack based on the electrical signal monitoring results. Based on this correlation, it extracts the coupling duration and spatial gradient, and inputs them into the SOC state identification model to output the estimated SOC of each battery pack. Tags; Among them, the SOC state identification model is based on the regression tree combination architecture, introduces the particle swarm algorithm, and adjusts the feature weights and standard anomaly thresholds according to the control mode and effective coupling period. The fault output module performs a sliding window weighted average on the estimated SOC labels to generate an SOC status score, and introduces a machine learning model to predict the next fault event.
[0014] (III) Beneficial Effects This invention provides a method and system for detecting the state of charge (SOC) of a battery pack, which has the following advantages: 1. This invention designs a main battery pack and a secondary battery pack, and connects the secondary battery pack in parallel through the main battery pack to establish the relationship between the main and secondary battery packs, thus constructing a two-dimensional correlation space. By analyzing the horizontal and vertical slopes in the two-dimensional correlation space, the spatial gradient value is determined through vector synthesis. By analyzing the ascending and descending gradients, the spatial evolution caused by pressure difference can be effectively identified and quantified, providing a basis for the subsequent node splitting of the regression tree model. 2. This invention establishes the relationship between the main battery pack and the auxiliary battery pack, constructs a battery pack topology network, and quantifies the complex relationship of the battery topology network by setting a first function and a second function through mathematical modeling, which facilitates subsequent prediction and analysis. Combined with a SOC state identification model based on a regression tree combination architecture, node splitting features are selected according to the size of the anomaly score, and a particle swarm optimization algorithm is introduced to optimize feature weights and standard anomaly thresholds, thereby improving the robustness of the identification model and enhancing the identification accuracy of SOC. 3. This invention introduces a machine learning model to predict the next fault event. In this process, the prediction accuracy of the model is improved by setting the first branch and the second branch, which plays a role in timely early warning. Attached Figure Description
[0015] Figure 1 This is a schematic flowchart illustrating a SOC detection method according to an exemplary embodiment. Detailed Implementation
[0016] The technical solutions of 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.
[0017] The core of this invention lies in establishing the association between the main and auxiliary battery packs by determining the electrical signal monitoring results of the main and auxiliary battery packs, thereby establishing a corresponding topology network. By monitoring the connectivity of the topology edges and using mathematical modeling to quantify the complex relationships of the topology network, the corresponding node splitting features are obtained for model training. By introducing the particle swarm optimization algorithm, the feature weights and standard anomaly thresholds under different access modes are dynamically iterated to obtain accurate predicted SOC labels. By generating SOC status scores and fault event prediction feedback in real time, potential risks can be identified in a timely manner during actual operation, and corresponding control decisions can be made.
[0018] Example 1: This invention provides a method for detecting the state of charge (SOC) of a battery pack. Figure 1 This is a flowchart illustrating a SOC detection method according to an exemplary embodiment; please refer to [link / reference]. Figure 1 The method includes the following steps: S1: Identify a main battery pack and at least one auxiliary battery pack; wherein the main battery pack connects to the auxiliary battery pack in parallel and sends a synchronization detection command to obtain the electrical signal monitoring results of each battery pack in real time; wherein the electrical signal monitoring results include the shunt current and terminal voltage of each battery pack; In this embodiment, after the power-on signal initialization is triggered, the bus nodes are scanned to read the hardware ID and BMS master / slave flag of each battery pack, and preliminary screening of primary and secondary battery pack candidates is performed. For example: Hardware ID: the primary pack ID starts with M, and the secondary pack ID starts with S; BMS master / slave flag: Master=1 represents the primary pack, and Slave=0 represents the secondary pack; System-level commands are sent to the primary and secondary battery pack candidates, such as: Sleep command: enters battery sleep mode, the one that responds is the primary battery pack, and the one that does not respond is the secondary battery pack; The above is an example of determining the primary and secondary battery packs. The specific determination process is set according to the actual situation; By sending a synchronization detection command, the primary and secondary packs are monitored in parallel. The primary battery pack is monitored through the communication bus. For example, the CAN bus sends a synchronous detection command, which connects the main battery pack to the auxiliary battery pack in parallel. Sampling is performed on each battery pack, and the electrical signal monitoring results are monitored using Hall current sensors and high-voltage sampling probes. Specifically, Hall current sensors are used to monitor the branch current. For instance, the sampling frequency is set to 100Hz to 1000Hz, and a hardware filter removes high-voltage pulse interference. The high-voltage sampling probe monitors the terminal voltage, directly obtaining the total potential difference across the battery cell assembly. It should be noted that the sensors used for the above monitoring are not shown in the diagram. Adaptive installation is possible; for example, Hall current sensors are connected in series to the positive output terminal of each battery pack, and high-voltage sampling probes are connected to the positive and negative busbars of the battery cell assembly. Further details are omitted here.
[0019] S2: Based on the electrical signal monitoring results, establish the correlation between the main battery pack and the auxiliary battery pack. Based on the correlation, extract the coupling duration and spatial gradient, and input them into the SOC state identification model to output the estimated SOC label of each battery pack. The SOC state identification model is based on the regression tree combination architecture and introduces the particle swarm algorithm to adjust the feature weights and standard anomaly thresholds according to the control mode and effective coupling period. Establish the relationship between the main battery pack and the auxiliary battery pack, including: The main battery pack is mapped as the root topology node, and each sub-battery pack is mapped as a slave topology node, establishing a battery pack topology network. Cell branches within each slave topology node are identified, and a hierarchy of each slave topology node and all its internal cell branches is established using topology path coding. The relationship between the root topology node and each cell branch is also established, with each cell branch's topology edge configured with equivalent resistance and polarization inductance. The equivalent resistance characterizes the DC ohmic internal resistance of the cell branch and is obtained using a hybrid pulse power characteristic test. For example, at 25°C, a 50A pulse current is applied to the branch for 10 seconds, and the voltage change ΔU is recorded and calculated as ΔU / 50. The polarization inductance characterizes the inductive component of the cell branch as the current changes over time and is measured using the step current method. For example, apply a current step of 1A / s, record the induced voltage, calculate the ratio of the induced voltage generated by the cell branch during the dynamic change of current to the rate of change of the branch current; at the same time, obtain the electrical signal monitoring results of each node at the same time stamp through synchronous sampling, calculate the first product of the equivalent resistance and the branch current, the first product is used to characterize the instantaneous voltage drop of the cell branch due to the ohmic resistance, calculate the second product of the polarization inductance and the rate of change of current, the second product is used to characterize the dynamic induced voltage generated by the cell branch during current fluctuations, such as the instant of parallel connection, by normalizing the first product and the second product to eliminate the influence of dimensions, and establish a two-dimensional correlation space with the ratio of the first product to the terminal voltage as the horizontal axis and the ratio of the second product to the terminal voltage as the vertical axis; Extracting coupling duration and spatial gradient includes: monitoring the connectivity of topological edges, characterizing the electrical continuity from the root topological edge to the topological edge from the topological edge or from the topological node to the branch topological edge, including but not limited to disconnection and connection, reflecting the actual working state of the parallel relay; Based on the first closure detection result, the start time of the auxiliary battery pack connection is determined. The first closure detection result is the contact detection result of the parallel relay and high-voltage bus corresponding to the auxiliary battery pack. This contact detection result is for the closed state detection of the parallel relay and high-voltage bus of the auxiliary battery pack, used to confirm whether the relay has gone from open to closed and whether the contacts have truly made contact. Based on the first disconnection detection result, the end time of the auxiliary battery pack connection is determined. The first disconnection detection result is the disconnection detection result of the parallel relay and high-voltage bus, used to confirm whether the relay has gone from closed to open and whether the contacts have truly disconnected. The coupling duration is generated based on the time difference between the start and end times of connection. Determining the coupling duration also includes: obtaining all shunt currents of the main battery pack and the auxiliary battery pack within the coupling duration, forming corresponding current sequences, including a first current sequence and a second current sequence, wherein the first current sequence corresponds to the main battery pack and the second current sequence corresponds to the auxiliary battery pack; Spatiotemporal calibration of the current sequences is performed to ensure that the two current sequences have the same time base, thus ensuring the accuracy of subsequent correlation analysis. Due to the presence of high-frequency switching noise in the parallel circuit, a sliding window is used to perform local mean denoising on the two current sequences. The cross-correlation coefficient between the first and second current sequences after denoising is calculated. Coupling durations with cross-correlation coefficients greater than the standard coupling threshold are selected and extracted to obtain the effective coupling period. It should be noted that the standard coupling threshold is 0.9, and the cross-correlation coefficient ranges from -1 to 1. If the cross-correlation coefficient approaches 1, it is determined that the load distribution between the main and auxiliary battery packs is more balanced and the energy coupling relationship is more stable. If the cross-correlation coefficient approaches 0, it is determined that the energy coupling relationship between the main and auxiliary battery packs is less stable. If the cross-correlation coefficient approaches -1, it is determined that the main and auxiliary battery packs have severe self-balancing circulating currents, which need to be corrected. The state evolution trajectory is determined based on the two-dimensional association space. The horizontal and vertical slopes within the coupling duration are extracted based on the state evolution trajectory. The square root algorithm is used to synthesize the horizontal and vertical slopes at the same moment into the instantaneous spatial gradient value at that moment. That is, the square of the horizontal slope at the same moment is calculated, the square of the vertical slope at the same moment is calculated, the two calculation results are added together and the square root is taken to obtain the instantaneous spatial gradient value at that moment. According to the positive and negative signs of the slopes, all instantaneous spatial gradients within the coupling duration are divided into ascending gradients and descending gradients to determine the spatial gradient. It should be noted that during normal charging and discharging, the main and auxiliary packages are in a stable parallel state, and the currents of the main and auxiliary packages change synchronously. Since the horizontal axis is the ratio of the first product to the terminal voltage and the vertical axis is the ratio of the second product to the terminal voltage, they will slowly move along the direction of the ray emitted from the origin in the two-dimensional associated space. Therefore, the trajectory of state evolution is represented as a relatively flat and smooth curve, and the slopes of the horizontal and vertical axes change little. When there is a voltage difference between the main and auxiliary packages in parallel, the branch current will experience a nonlinear surge, which will force the generation of circulating current. Therefore, the trajectory of state change is represented as a radial curve, and the slopes of the horizontal and vertical axes change significantly.
[0020] By analyzing the horizontal and vertical slopes in the two-dimensional correlation space, the spatial gradient value is determined through vector synthesis. By analyzing the ascending and descending gradients, the spatial evolution caused by pressure difference can be effectively identified and quantified, providing a basis for the node splitting of the subsequent regression tree model.
[0021] The SOC state identification model includes: extracting the association between each topology node and the cell branch; wherein, the topology node is the main battery pack and the auxiliary battery pack participating in parallel, and the cell branch is the physical path connecting each topology node to the actual high-voltage bus and the sampling channel, and each branch carries at least the coupling duration and spatial gradient; setting the state function group of the topology node: including the first function and the second function; The construction steps of the first function are as follows: Based on the battery pack topology network, the equivalent impedance of the main battery pack node is determined as the reference impedance. The set of all sub-battery pack nodes participating in parallel is obtained. For any sub-battery pack node, its real-time equivalent impedance is subtracted from the reference impedance to obtain the impedance deviation. A mapping equation is established using the sum of the impedance deviation and the reference value. By solving the equations of all nodes, the above mapping equations are combined to form the first function. The construction steps of the second function are as follows: Based on the battery pack topology network, the polarization inductance on all cell branches is determined. In the parallel circuit, the terminal voltage of the main battery pack is used as the reference voltage. The first product consisting of the branch current and the equivalent resistance, and the second product consisting of the polarization inductance and the rate of change of current are subtracted from the reference voltage to establish the correlation equation. By solving the correlation equations of all cell branches participating in parallel at the same time, the second function is formed. The corresponding feature solution matrix is obtained by analyzing the state function set. To eliminate the influence of dimensions, the feature solution matrix is normalized and weighted and mapped with the feature weights of node splitting under the current access mode, which includes direct connection and plug-in access. Initial feature weights are obtained by calling the database. The feature solution matrix is combined with the corresponding feature weights using a matrix weighting algorithm. An anomaly score is obtained by multiplying a component of the feature solution matrix with the corresponding feature weight and accumulating the results. It should be noted that the larger the anomaly score, the greater the degree of deviation of the system from steady-state parallel operation and the stronger the nonlinear interference of the data. When the regression tree model performs SOC identification, the feature of the anomaly score is adaptively switched, including: comparing the anomaly score with the standard anomaly threshold. If the anomaly score is greater than the standard anomaly threshold, it indicates that the system is in a pressure shock or contact fault state. At this time, the current integration logic fails, and the spatial gradient is selected as the node splitting feature, and the two-dimensional correlation space is called to correct and predict the SOC label. If the anomaly score is less than or equal to the standard anomaly threshold, it indicates that the parallel system is operating smoothly and the data confidence is high. In this case, the coupling duration is selected as the node splitting feature. The standard anomaly threshold is a dynamic value. The process involves adjusting feature weights and standard anomaly thresholds based on the control mode and effective coupling period, including: constructing a high-dimensional target search space; using feature weights and standard anomaly thresholds as parameters to be optimized, mapping them to position vectors within the target search space; dynamically searching for the optimal parameter vector within the target search space by randomly generating an initial population of m particles, each containing a unique position and velocity vector; employing a binary bidding strategy to select superior individuals from the population for mating pools, performing crossover and mutation operations on the particles in the mating pools to generate subpopulations; and simultaneously providing evolutionary constraints for population evolution based on the current access mode and effective coupling period until the maximum number of iterations is reached, where m is a positive integer; for example: number of particles m=30, maximum number of iterations=100, crossover probability=0.7, mutation probability=0.1. By using the access mode as a physical constraint term in the fitness function to adjust the position vector, the following steps are taken: First, identify the current access mode. If the access mode is direct connection, add a penalty term to the fitness function for solutions with low spatial gradient weights to guide the position vector towards a direction that enhances dynamic feature recognition. For example, the lower the spatial gradient feature weight, the less it meets the dynamic feature requirements of direct connection. If the access mode is plug-in access, add a penalty term to the fitness function for solutions with low standard anomaly thresholds to guide the position vector towards a region with high impedance tolerance. For example, the lower the standard anomaly threshold, the lower the tolerance for anomaly scoring, and the less it meets the fluctuation characteristics of plug-in access. Specific parameters are set according to the actual situation and are not detailed here. Further explanation is needed; it should be noted that the fitness function is the sum of the core term and the physical constraint term. The core term is the calculation of minimizing the residual, and the physical constraint term is the adaptation access mode to avoid poor accuracy. The effective coupling period ratio is then used as a search constraint term to adjust the velocity vector, including: real-time calculation of the ratio of the effective coupling period to the coupling duration to obtain the effective coupling period ratio; when the effective coupling period ratio increases, the inertia weight of the velocity vector is reduced by a preset ratio, so that the search behavior switches to a local search mode centered on the current optimal solution; when the effective coupling period ratio decreases, the magnitude of the velocity vector is increased by a preset ratio, so that the search behavior switches to a global search mode with a large step size to skip the local extreme value region formed by noise interference. Simultaneously, the optimal parameter vector is output with the goal of minimizing the residual. Minimizing the residual involves summing the squares of the instantaneous residuals at each sampling moment during iterative searching of the optimal parameter vector within the search space to obtain a loss function. The position and velocity vectors are continuously corrected using a particle swarm optimization algorithm to minimize the loss function, ultimately outputting the optimal parameter vector. The optimal parameter vector is then used to configure the parameters of the regression tree model, mapping each component of the optimal parameter vector to the gain of the feature weights for the regression tree node split and the offset of the standard anomaly threshold. For example, from the position vector of the (n+1)-dimensional optimal parameter vector, scalar values of each dimension are extracted sequentially according to the index order. The first n-dimensional components are defined as the gain of the feature weights, and the (n+1)-th component is defined as the offset of the standard anomaly threshold. By adding the offset to the initial standard anomaly threshold of the system, a dynamic standard anomaly threshold adapted to the current access mode and coupling duration is generated; where n is a positive integer. By establishing the relationship between the main battery pack and the auxiliary battery pack, a battery pack topology network is constructed. Through mathematical modeling, a first function and a second function are set to quantify the complex relationships of the battery topology network, facilitating subsequent predictive analysis. Combined with a SOC state identification model based on a regression tree combination architecture, node splitting features are selected according to the size of the anomaly score. Particle swarm optimization is introduced to optimize feature weights and standard anomaly thresholds. Parameters are dynamically adjusted through physical constraints and search constraints, overcoming the scenario limitations of fixed parameter models, improving the robustness of the model, and enhancing the identification accuracy of SOC.
[0022] S3: Perform a sliding window weighted average on the estimated SOC labels to generate an SOC status score, and introduce a machine learning model to predict the next failure event; The predicted SOC labels output over multiple consecutive time periods are obtained. The fitness of the best particle output by the particle swarm optimization algorithm is extracted as the weight coefficient. That is, the influence of the constraint term is ignored, and the reciprocal of the minimized residual is directly used as the weight coefficient. The weighted average calculation is performed. The SOC state score is generated by multiplying the weight coefficients of multiple time periods and the predicted SOC labels and accumulating them. Example: Assume there are three consecutive time periods: Time period 1: estimated SOC label is 50%, residual is 0.5, fitness is 2; Time period 2: estimated SOC label is 51%, residual is 0.1, fitness is 10; Time period 3: estimated SOC label is 51.5%, residual is 0.05, fitness is 20. The total weight base is obtained by calculating 2 + 10 + 20 = 32. Each period is divided by this total weight base to obtain the final weight percentage. Rounded to two decimal places, 2 / 32 = 6%, 10 / 32 = 31%, 20 / 32 = 63%. Therefore, the SOC status score is 50% × 6% + 51% × 31% + 51.5% × 63% = 51.25%. The first branch comprises an input layer, a first hidden layer, a second hidden layer, and an output layer. The mean, standard deviation, and timestamp extracted from the SOC (State of Occurrence) score are input into the input layer, which has 3 neurons. The first hidden layer maps the features to a high-dimensional space, capturing the non-linear relationship of SOC features, including the combined characteristics of decreasing mean and increasing standard deviation; this first hidden layer has 16 neurons. The second hidden layer reduces the dimensionality and filters core features to avoid overfitting; this second hidden layer has 8 neurons. The output layer outputs a first feature vector, which has 4 neurons. Simultaneously, the electrical signal monitoring results at the same timestamp are retrieved, and the temporal features of voltage extreme differences are statistically analyzed to generate a second feature vector. Feature fusion is performed on the first and second feature vectors, and the fusion weights are adjusted by effectively coupling time periods to generate fused features. The second branch comprises an input layer, a first hidden layer, a second hidden layer, and an output layer. The fused features are input into the input layer (8 neurons). The first hidden layer maps the fused features to a high-dimensional space, capturing the non-linear relationship between the fused features and the fault. This first hidden layer has 32 neurons. The second hidden layer filters the core fault features (16 neurons). The output layer outputs four fault types (4 neurons). These fault types include no fault, battery imbalance, sudden increase in SOC, and sharp decrease in SOC. The type of the next fault event is predicted based on the highest probability value among the fault types. During training, historical data is divided into training, validation, and test sets in a 7:2:1 ratio. The Adam optimizer is used with a learning rate of 0.001, resulting in fast and stable convergence, aiding in accurate model prediction.
[0023] By introducing a machine learning model to predict the next fault event, the model's prediction accuracy is improved by setting the first and second branches. It can successfully distinguish abnormal fluctuations caused by faults such as battery imbalance, sudden increase in SOC, and sharp decrease in SOC, thus playing a role in timely early warning and making corresponding control decisions, achieving a dual improvement in battery performance optimization and safety management.
[0024] Example 2: This invention provides a battery pack SOC detection system, which includes a target sensing module, a detection and analysis module, and a fault output module, wherein the target sensing module, the detection and analysis module, and the fault output module are communicatively connected. The target perception module identifies a main battery pack and at least one auxiliary battery pack; wherein the main battery pack performs parallel connection to the auxiliary battery pack and sends a synchronous detection command to obtain the electrical signal monitoring results of each battery pack in real time; The detection and analysis module establishes the correlation between the main battery pack and the auxiliary battery pack based on the electrical signal monitoring results. Based on the correlation, it extracts the coupling duration and spatial gradient and inputs them into the SOC state identification model, outputting the estimated SOC label of each battery pack. The SOC state identification model is based on the regression tree combination architecture and introduces the particle swarm algorithm to adjust the feature weights and standard anomaly thresholds according to the control mode and effective coupling period. The fault output module performs a sliding window weighted average on the estimated SOC labels to generate an SOC status score, and introduces a machine learning model to predict the next fault event.
[0025] In the application, the various formulas mentioned are all calculated by removing dimensions and taking their numerical values. The formulas are derived from the most recent real-world situation by collecting a large amount of data and conducting software simulations. The formulas are set by those skilled in the art according to the actual situation.
[0026] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.
[0027] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.
[0028] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.
Claims
1. A method for detecting the state of charge (SOC) of a battery pack, characterized in that, The method includes: identifying a main battery pack and at least one auxiliary battery pack; wherein the main battery pack performs parallel connection to the auxiliary battery pack and sends a synchronization detection command to acquire the electrical signal monitoring results of each battery pack in real time; Based on the electrical signal monitoring results, the correlation between the main battery pack and the auxiliary battery pack is established. Based on the correlation, the coupling duration and spatial gradient are extracted and input into the SOC state identification model, and the estimated SOC label of each battery pack is output. The SOC state identification model is based on the regression tree combination architecture and introduces the particle swarm algorithm to adjust the feature weights and standard anomaly thresholds according to the control mode and effective coupling period. A sliding window weighted average is applied to the estimated SOC labels to generate an SOC status score. A machine learning model is then introduced to predict the next failure event.
2. The SOC detection method for a battery pack according to claim 1, characterized in that, The electrical signal monitoring results include the shunt current and terminal voltage of each battery pack.
3. The SOC detection method for a battery pack according to claim 2, characterized in that, Establish the relationship between the main battery pack and the auxiliary battery pack, including: The main battery pack is mapped as the root topology node, and each sub-battery pack is mapped as a slave topology node. The cell branches inside the slave topology node are identified, and the association between the root topology node and each cell branch is established. The topology edge where each cell branch is located is configured with equivalent resistance and polarization inductance. Simultaneously, by synchronously sampling, the electrical signal monitoring results of each node at the same timestamp are obtained, the first product of equivalent resistance and shunt current and the second product of polarization inductance and current change rate are calculated, and a two-dimensional correlation space is established with the ratio of the first product to the terminal voltage as the horizontal axis and the ratio of the second product to the terminal voltage as the vertical axis.
4. The SOC detection method for a battery pack according to claim 3, characterized in that, Extracting coupling duration and spatial gradient, including: Monitor the connectivity status of the topology edges; identify the contact detection results of the parallel relays and high-voltage buses corresponding to the auxiliary battery pack to determine the start time of the auxiliary battery pack's connection; identify the disconnection detection results of the parallel relays and high-voltage buses corresponding to the auxiliary battery pack to determine the end time of the auxiliary battery pack's connection; and determine the coupling duration based on the start and end times of the connection. The state evolution trajectory is plotted in the two-dimensional associated space, and the rising and falling gradients of the state evolution trajectory within the coupling duration are identified to determine the spatial gradient.
5. The SOC detection method for a battery pack according to claim 4, characterized in that, Determining the coupling duration also includes: All shunt currents of the main battery pack and the auxiliary battery pack within the coupling time are obtained to form corresponding current sequences, including a first current sequence and a second current sequence, wherein the first current sequence corresponds to the main battery pack and the second current sequence corresponds to the auxiliary battery pack. Identify the cross-correlation coefficients of the first and second current sequences, filter and extract the coupling durations where the cross-correlation coefficients are greater than the standard coupling threshold, and obtain the effective coupling time periods.
6. The SOC detection method for a battery pack according to claim 5, characterized in that, The SOC state identification model includes: Extract the association between each topology node and the cell branch, and each branch must carry at least the coupling duration and spatial gradient; Set the state function group of the topology node: including the first function and the second function; obtain the corresponding feature solution matrix by parsing the state function group, and weight and map it with the feature weight of node splitting under the current access mode to generate an anomaly score; at the same time, compare the anomaly score with the standard anomaly threshold. If the anomaly score is greater than the standard anomaly threshold, the spatial gradient is selected as the node splitting feature. If the anomaly score is less than or equal to the standard anomaly threshold, the coupling duration is selected as the node splitting feature. The standard anomaly threshold is a dynamic value.
7. The SOC detection method for a battery pack according to claim 6, characterized in that, Access modes include direct connection and plug-in connection.
8. The SOC detection method for a battery pack according to claim 7, characterized in that, A particle swarm optimization algorithm is introduced to adjust feature weights and standard anomaly thresholds based on access patterns and effective coupling periods, including: Construct a high-dimensional target search space; Using feature weights and standard anomaly thresholds as parameters to be optimized, the optimal parameter vector is dynamically searched within the target search space. Based on the current access mode and effective coupling period, evolutionary constraints are provided for population evolution, including: adjusting the position vector by using the access mode as a physical constraint term in the fitness function, and adjusting the velocity vector by using the proportion of effective coupling period as a search constraint term, until the maximum number of iterations is reached; the optimal parameter vector is output with the goal of minimizing the residual, and mapped to the gain of the feature weights of the regression tree node split and the offset of the standard anomaly threshold, respectively.
9. The SOC detection method for a battery pack according to claim 8, characterized in that, Machine learning models have built-in first and second branches, including: After outputting the predicted SOC label, the reciprocal of the residual is used as the weight coefficient to determine the SOC status score; The first branch extracts the mean, standard deviation, and timestamp from the SOC status score to construct a first feature vector; simultaneously, it retrieves the electrical signal monitoring results at the same timestamp, statistically analyzes the time-series characteristics of voltage range differences, and generates a second feature vector; feature fusion is performed on the first and second feature vectors to generate a fused feature. The fused features are input into the second branch to predict the next failure event.
10. A battery pack SOC detection system, characterized in that, The system includes: a target sensing module, which identifies a main battery pack and at least one auxiliary battery pack; wherein the main battery pack performs parallel connection to the auxiliary battery pack and sends a synchronous detection command to acquire the electrical signal monitoring results of each battery pack in real time; The detection and analysis module establishes the correlation between the main battery pack and the auxiliary battery pack based on the electrical signal monitoring results. Based on the correlation, it extracts the coupling duration and spatial gradient and inputs them into the SOC state identification model, outputting the estimated SOC label of each battery pack. The SOC state identification model is based on the regression tree combination architecture and introduces the particle swarm algorithm to adjust the feature weights and standard anomaly thresholds according to the control mode and effective coupling period. The fault output module performs a sliding window weighted average on the estimated SOC labels to generate an SOC status score, and introduces a machine learning model to predict the next fault event.