Method for predicting service life of battery pack

Abstract

The invention relates to a battery pack life prediction method, which comprises the following steps of: respectively presetting scores corresponding to CPU / GPU utilization rate, battery pack inconsistency characteristic and data quality of a system, acquiring corresponding scores according to the CPU / GPU utilization rate, battery pack inconsistency characteristic and data quality of the current system, and respectively recording the scores; performing calculation; and executing the following prediction mode according to the value of lambda so as to accurately predict the service life of the battery pack, thereby improving the safety and reliability of battery pack management.

Description

Technical Field

[0001] This invention relates to the field of battery information, and more particularly to a method for predicting battery pack lifespan. Background Technology

[0002] A battery pack is an integrated energy module composed of multiple individual cells connected in series, parallel, or a hybrid configuration, and equipped with a battery management system (BMS), electrical protection structure, and thermal management system. Series connection increases the output voltage, while parallel connection expands the total capacity, enabling it to meet the high-voltage demands of equipment and extend power supply time. The BMS monitors voltage, temperature, and current in real time, achieving intelligent balancing and overload protection to ensure stable and safe operation. It is widely used in new energy vehicles, energy storage power stations, and consumer electronics, and is a core component of modern portable and clean energy technologies.

[0003] However, in existing technologies, battery pack monitoring systems typically only collect the voltage, current, and surface temperature of individual cells, lacking synchronous perception of multi-dimensional information such as the internal thermal field distribution of the module and vibration of the battery pack's operating environment. This makes it impossible to accurately locate hot spots or assess the impact of environmental stress on the aging process. Furthermore, most battery pack life predictions rely on purely data-driven models, which can lead to significant inaccuracies in predictions when there is high data noise, sensor failure, or encountering unprecedented extreme conditions. Consequently, the predictions of battery pack life are not accurate. Summary of the Invention

[0004] Based on the aforementioned problems in the existing technology, the purpose of this invention is to achieve accurate prediction of battery pack lifespan, thereby improving the safety and reliability of battery pack management.

[0005] To achieve the above objectives, this invention proposes a battery pack lifespan prediction method, comprising: pre-setting scores corresponding to the system's CPU / GPU utilization, battery pack inconsistency characteristics, and data quality; obtaining corresponding scores based on the current system's CPU / GPU utilization, battery pack inconsistency characteristics, and data quality, and recording them as follows: , ;calculate , For bias terms, , and The weights are used; based on the value of λ, the following prediction mode is executed: when λ>0.7, the first prediction mode is used to predict the remaining lifespan of the battery pack. When λ < 0.3, the second prediction mode is used to predict the remaining lifespan of the battery pack. When 0.3 ≤ λ ≤ 0.7, a fusion of the first and second prediction modes is used to predict the remaining lifespan of the battery pack. .

[0006] To achieve the above objectives, this invention proposes a battery pack life prediction method, comprising: when the generalized health state... When the battery level falls below a preset failure threshold, the iteration stops, and the number of virtual cycles elapsed from the current moment is used as the predicted remaining lifespan of the battery pack. Otherwise, repeat the following steps: obtain the estimated SOH health value for each monomer at time t. Calculate the estimated state of health (SOH) of each cell during the next charging cycle. : , Let be the change in the health estimate of the i-th individual between time t and t+1.

[0007] Calculate the generalized state of health of the battery pack : .

[0008] This represents the current capacity degradation value. The maximum capacity degradation threshold, This is the value of the internal resistance divergence. The maximum internal resistance divergence threshold, and These are the weighting coefficients.

[0009] .

[0010] This represents the minimum SOH value of all cells within the battery pack. This represents the average SOH of all monomers within the group. Let S be the standard deviation of the state of harmonics (SOH) of all cells within the battery pack. Starting from the current time, using a trained GNN model, with the combined adjacency matrix of the battery pack at the current time and the historical data from the relative iteration start time as input, predict the state of harmonics for each cycle in the future. The value forms a predicted trajectory, and based on the lifespan end threshold, the predicted trajectory is checked backward from the current moment to find the first... The difference between the cycle number at the time point that is less than or equal to the end-of-life threshold and the current time point is used as the predicted remaining lifespan of the battery pack. Final predicted remaining lifespan of the battery pack λ is the preset weight.

[0011] Furthermore, prior to achieving the aforementioned objectives, this invention proposes a battery pack health status monitoring method, comprising: constructing a multi-level sensing network to collect multi-dimensional operating parameters at three spatial levels: individual cells, modules, and battery packs. The multi-level sensing network is divided into individual cell-level sensing units, module-level sensing units, and battery pack-level sensing units. The individual cell-level sensing units are configured to monitor voltage, current, and surface temperature. The module-level sensing units are configured to use infrared thermal imaging or distributed fiber optic sensors to capture the thermal field distribution inside the module in real time. The battery pack-level sensing units are configured to integrate environmental temperature and humidity, vibration, and air pressure sensors to achieve comprehensive perception of the operating environment. Based on different battery pack operating states, adaptive data acquisition is performed, including: enabling high-frequency acquisition during charging to monitor voltage increments and temperature rise rates; and employing... A combination of mid- and high-frequency strategies is employed to balance power output stability and thermal runaway early warning. During the resting phase, a low-frequency inspection mode is switched to continuously track self-discharge behavior and voltage relaxation processes, avoiding resource consumption. The collected operating data undergoes data preprocessing, and battery health status-related features are extracted based on the processed operating parameters. These features include capacity inconsistency coefficient, internal resistance divergence, maximum temperature gradient, and voltage distribution entropy. The capacity inconsistency coefficient is obtained by calculating the ratio of the standard deviation to the average value of the available capacity among individual cells in the battery pack. The internal resistance divergence is the degree of difference in AC internal resistance among individual cells, obtained by fitting pulse response data during the charging and discharging process. The voltage distribution entropy is obtained by quantifying the dispersion of voltage curves under static or dynamic conditions based on information entropy theory, discretizing the voltage values ​​of all cells in the pack into multiple intervals, and then calculating the information entropy value.

[0012] The beneficial effects of this invention are: This application improves the accuracy and robustness of battery pack remaining lifespan, enabling stable remaining battery pack lifespan even under conditions of data scarcity or extreme operating conditions. Furthermore, it constructs a multi-level sensing network and adaptive sampling strategy across "cell-module-battery pack," achieving high-quality data acquisition across the entire battery pack lifecycle and multiple dimensions, monitoring battery pack health status, and providing data support for predicting battery pack remaining lifespan. Attached Figure Description

[0013] Figure 1 This is a flowchart of a battery pack life prediction method according to the present invention.

[0014] Figure 2 This is a schematic diagram of the network-side server provided according to the second embodiment of the present invention. Detailed Implementation

[0015] Regarding the following issues mentioned in the background technology: (1) The existing technology lacks synchronous perception of multi-dimensional information such as the internal thermal field distribution of the module and the vibration of the battery pack operating environment, making it impossible to accurately locate hot spots or assess the impact of environmental stress on the aging process. (2) The existing technology uses a pure data-driven model for prediction, which may result in serious inaccuracies when there is high data noise, sensor failure, or encountering extreme working conditions that have not been experienced before.

[0016] This disclosure proposes a battery life prediction method. Based on considerations of system CPU / GPU utilization, battery pack inconsistency characteristics, and data quality, it determines a battery pack life prediction pattern to improve the accuracy and reliability of battery pack life prediction. Furthermore, this disclosure also proposes a battery pack health status monitoring method.

[0017] The technical solution of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, 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.

[0018] An implementation method for monitoring the health status of a battery pack is proposed. By constructing a multi-level sensing network of "cell-module-battery pack" and an adaptive sampling strategy, high-quality data collection of the entire life cycle and multiple dimensions of the battery pack is achieved, and data support is provided for predicting the remaining lifespan of the battery pack.

[0019] In the battery pack health status monitoring method, this application defines the health status-related characteristics of the battery pack and defines a generalized battery pack health status index.

[0020] The multi-level sensing network is used to collect multi-dimensional data at three spatial levels: individual cells, modules, and battery packs. Specifically, the multi-level sensing network is divided into individual cell-level sensing units, module-level sensing units, and battery pack-level sensing units. The individual cell-level sensing units use high-precision sensors to detect three core parameters: voltage, current, and surface temperature. The module-level sensing units introduce infrared thermal imaging modules or distributed fiber optic sensors to capture the thermal field distribution inside the module in real time. The battery pack-level sensing units integrate environmental temperature and humidity, vibration, and air pressure sensors to achieve comprehensive perception of the operating environment.

[0021] As an example, in the single-unit hierarchical sensing unit, the voltage sensor has an accuracy of ±1mV to capture the slight voltage deviation between units; the current sensor has an accuracy of ±1% FS (full scale) to accurately calculate the charge and discharge capacity; and the temperature sensor uses an NTC thermistor with an accuracy of ±0.5℃ to monitor the temperature rise of the unit.

[0022] The module-level sensing unit detects and monitors the two-dimensional temperature field, accurately locates hot spots, and calculates the maximum temperature gradient within the module. This provides a basis for precise intervention in the thermal management system.

[0023] The battery pack-level sensing unit collects the total voltage and total current on the bus; it is paired with an ambient temperature and humidity sensor (accuracy ±5% RH) to record the battery pack's operating environment, which is used to correct the impact of environmental factors on battery performance and aging rate.

[0024] The adaptive acquisition strategy selects the corresponding frequency acquisition strategy based on different battery pack operating states. Specifically, it dynamically adjusts the sampling frequency of sensors at each level according to whether the battery is charging, discharging, or in a static state; during charging, high-frequency acquisition is enabled to monitor voltage increment and temperature rise rate, preventing overcharging risks; during discharging, a combination of medium and high frequency strategies is adopted to balance power output stability and thermal runaway warning; during the static stage, it switches to a low-frequency inspection mode to continuously track self-discharge behavior and voltage relaxation process, avoiding resource occupation.

[0025] As an example, high-frequency acquisition (10-100Hz) is suitable for high-dynamic and high-rate operating conditions such as fast charging, rapid acceleration, and rapid deceleration. Under such conditions, parameters such as the battery's polarization voltage and internal resistance will undergo rapid transient changes, and high-frequency acquisition can fully capture the transient change characteristics.

[0026] Intermediate frequency acquisition (1-10Hz): Adaptable to relatively stable operating conditions such as constant current charging / discharging and constant voltage charging; while ensuring effective monitoring of battery status change trends, it reduces the total amount of data, achieving a balance between data volume and timeliness.

[0027] Low-frequency triggered acquisition (0.01-0.1Hz): The acquisition action is triggered by specific events, such as when the battery reaches the preset state of charge threshold, when the temperature fluctuates abnormally, or when the system self-test command is issued. It only acquires key data when necessary, which significantly reduces system power consumption and communication load, extends the service life of the monitoring system, and is suitable for unattended operation mode in long-term energy storage scenarios.

[0028] Next, the operating parameters collected from the battery pack are preprocessed. Specifically, the 3σ criterion is used to remove outliers from voltage and current data, filtering out transient interference and sensor noise. For missing data caused by communication interruptions or temporary sensor failures, linear interpolation is used to complete the data, ensuring the continuity and integrity of the data sequence. Subsequently, Z-score normalization is performed on voltage, current, temperature, and environmental parameters to unify the dimensions and eliminate differences in numerical scale between parameters. Then, the multi-source sensor data is time-series aligned based on timestamps to control the time deviation between signal channels within 10ms, ensuring the spatiotemporal consistency of subsequent feature extraction and state estimation.

[0029] Then, health status-related features of the battery pack are extracted based on the processed operating parameters. Furthermore, early warning prompts can be issued based on the extracted health status-related features of the battery pack.

[0030] The health status characteristics of the battery pack include capacity inconsistency coefficient, internal resistance divergence, maximum temperature gradient, and voltage distribution entropy. Specifically, the capacity inconsistency coefficient is calculated by the ratio of the standard deviation to the average value of the available capacity among the individual cells in the battery pack, reflecting the inconsistency in aging; the internal resistance divergence characterizes the degree of difference in AC internal resistance among individual cells, and is obtained by fitting pulse response data during the charging and discharging process. When the internal resistance divergence is greater than 20%, under high current conditions, cells with high internal resistance will generate much more heat than their counterparts, causing local overheating and posing a safety hazard; the maximum temperature gradient captures the non-uniformity of the internal thermal field distribution of the module and identifies potential thermal runaway risk points; the voltage distribution entropy quantifies the dispersion of the voltage curve under static or dynamic conditions based on the information entropy theory, discretizing the voltage value distribution of all cells in the pack into multiple intervals and calculating its information entropy value. The more uniform and orderly the voltage distribution, the lower the entropy value; the more dispersed and disordered the distribution, the higher the entropy value. The entropy value reflects the overall consistency evolution trend of the battery pack.

[0031] The generalized health status ( The indicator is a composite indicator that considers both the shortcomings and average levels of individual performance while incorporating penalties for inconsistencies, based on the generalized health status. The calculation formula for ) is as follows.

[0032]

[0033] This represents the minimum SOH value of all cells within the battery pack. This represents the average SOH of all monomers within the group. This represents the standard deviation of SOH for all cells within the battery pack. , , For the weights, where The penalty weights are α=0.5, β=0.3, and γ=0.2 in one implementation.

[0034] An enhancement correction is introduced based on the basic generalized state of health (G-SOH), and the formula for enhancing the generalized state of health is as follows.

[0035]

[0036] in, and These are the weighting coefficients. =0.3, =0.2; A normalized indicator of capacity degradation; This represents the current capacity degradation value. The maximum capacity degradation threshold, It is a normalized index for the divergence of internal resistance; This is the value of the internal resistance divergence. The threshold for maximum internal resistance divergence is used. When inconsistency exceeds the corresponding threshold, the enhancement of the generalized health status is proportionally attenuated and corrected. The decline accelerates, allowing for earlier risk warnings.

[0037] The change in health estimates within the i-th time interval is calculated using the following formula. :

[0038] In the formula, Let be the duration of the i-th time interval. The change in the health estimate within the i-th time interval is the change in the health status of the i-th individual between time t and t+1.

[0039] The aging rate of each individual cell in the battery pack can be given by the Arrhenius model (the Arrhenius model is used to describe the relationship between chemical reaction rate and temperature): , The preceding factor was obtained by fitting a large amount of accelerated aging experimental data. This indicates the sensitivity of the aging reaction to temperature. Activation energy is the minimum energy required for a reaction to occur. The gas constant is Temperature; the exponential accelerating effect of temperature on aging rate is quantified using an aging rate modeling formula.

[0040] It is an inconsistency evolution function used to measure inconsistencies in battery packs. , Let represent the magnitude of the load current of the i-th individual unit, with an exponent of 0.5, indicating that the impact of operating conditions on the inconsistency evolution is sublinear; The maximum temperature gradient in the battery pack reflects the quality of thermal management within the pack. The greater the temperature difference, the faster the performance differences between individual battery cells will diverge, exacerbating inconsistency. An index of 0.3 indicates that the impact of temperature non-uniformity on the evolution of inconsistency is a relatively weak positive correlation. Initial capacity inconsistency represents the difference in the initial capacity of the battery pack, with an index of 0.2, indicating that the initial inconsistency has a very weak impact on the subsequent evolution of inconsistency.

[0041] In one embodiment, the battery health status monitoring method further includes: calculating the Euclidean distance of the aging trajectory using the SOH (health status) decay sequence of each cell during its life cycle.

[0042] Specifically, the Euclidean distance of the aging trajectory is used to calculate the Euclidean distance between any two monomer vectors. The Euclidean distance is calculated as follows.

[0043]

[0044] in, This represents the distance between the health status trajectories of monomer i and monomer j at T monitoring times. and The values ​​are the health estimates of the i-th and j-th individuals at time t, respectively. The smaller the distance, the more similar their aging paths are; the larger the distance, the more significant the difference in their degenerative behavior.

[0045] Based on the Euclidean distance of the aging trajectory of each individual monomer, the differences in state evolution paths among monomers are analyzed. As an example, when When the value is >0.15, this pair of monomers is marked as a "highly different monomer pair", indicating that the aging pathways between the two monomers have been significantly separated and require close attention.

[0046] The Euclidean distance of the aging trajectory and the health status-related characteristics of the battery pack are used as inconsistency features. The average or maximum value of each inconsistency feature is calculated as an inconsistency index, which can be used for time-series trend analysis.

[0047] The time-series trend analysis: For time-series data of inconsistent indicators, a sliding window method (the window size is usually set to 10 charging cycles) is used to perform linear fitting to capture the pattern of indicator changes over time, quantify the evolution trend of performance differences among individual cells within the battery pack, and thus identify abnormal changes in the health status of the battery pack in advance.

[0048] By calculating the Euclidean distance of the aging trajectory and performing time-series trend analysis, the degree of aging difference between individual cells in the battery pack is dynamically evaluated. By combining preset thresholds with real-time data comparison, abnormal cells that deviate from the normal degradation path can be identified, enabling early identification and warning.

[0049] The data collected in the above-mentioned battery health status monitoring methods can also provide support for predicting the remaining service life of the battery pack, as detailed below.

[0050] A method for predicting battery pack lifespan, see [link to relevant documentation]. Figure 1 This includes the following steps.

[0051] S1, obtain the battery inconsistency comprehensive score, data quality score, and computing resource score respectively.

[0052] The scoring can be preset, with scores corresponding to the system's CPU / GPU utilization, battery pack inconsistency characteristics, and data quality respectively.

[0053] The battery inconsistency comprehensive score is obtained by normalizing each inconsistency indicator to a range of 0-1 and then taking a weighted average. The higher the inconsistency, the lower the score.

[0054] The data quality is assessed based on the data collected from the battery pack, considering factors such as data missing rate, data noise level, and sensor health status. The data quality score is calculated by taking the currently collected operating parameter data during battery pack health status monitoring, obtaining the data missing rate, data noise level, and sensor health / fault status, normalizing each item to a value between 0 and 1, and then calculating a weighted average. Higher data quality results in a higher score.

[0055] Within a fixed time window (such as a complete charging cycle), the missing rate of each parameter (voltage, current, temperature, etc.) is calculated. The specific formula for quantitatively evaluating the missing data rate is as follows.

[0056]

[0057] To assess the reliability and cleanliness of the data, select a stable data segment and calculate the ratio of its signal mean to standard deviation. The quantitative assessment formula for data noise level is as follows.

[0058]

[0059] in, The mean of the data. The standard deviation of the data. The higher the value, the lower the noise level.

[0060] Using federated Kalman filtering or consensus algorithms, a "fusion truth value" for key parameters (such as voltage) is generated. The absolute deviation of each sensor reading from this truth value is calculated. The quantitative evaluation formula for sensor health status is as follows.

[0061]

[0062] The smaller the deviation, the better the health of the sensor.

[0063] The acquisition of computing resource scores involves considering the current system's CPU / GPU utilization. A higher utilization rate results in a lower score, indicating a heavy system load and limited resources available for additional computing. The current system's memory usage also plays a role; a higher memory utilization rate results in a lower score, indicating a shortage of available memory resources.

[0064] The weights involved in the weighted average in this step can be preset.

[0065] S2. Construct an adaptive gating fusion mechanism: Obtain the gating coefficient based on the acquired battery inconsistency comprehensive score, data quality score, and computing resource score.

[0066] Specifically, a gating coefficient λ is constructed, and the formula for calculating the gating coefficient is as follows:

[0067] in, , and As weight, To calculate resource scores; A comprehensive score for battery pack inconsistency; The data quality is scored, with all three scoring ranges between 0 and 1. Used to assess the overall status of current computing resources, inconsistencies, and data quality; This is a bias term used to correct the model's predictions and adjust the overall offset of the gating coefficients. In one implementation, =0.3, =0.4, =0.2, b=0.1.

[0068] As an example: Train a linear regression model with inputs of computing resources, battery pack inconsistency, and data quality, and output λ. Select a model (GNN-dominated, balanced fusion, or physical model-dominated) based on the value of λ. Construct labels for the three modes: GNN with a label of 1, balanced fusion with a label of 0.5, and physical model-dominated with a label of 0. Use the linear regression model to train the mapping from the three features to the labels, and obtain the optimal weights and bias parameters for the gating coefficient λ.

[0069] S3. Execute the corresponding prediction mode based on the value of λ: When λ > 0.7 (sufficient resources, high consistency, and good data): the first prediction mode is used for prediction, with a weighting of 70%, to obtain the predicted remaining lifespan of the battery pack. .

[0070] When 0.3 ≤ λ ≤ 0.7 (moderate conditions): A fusion of the first and second prediction modes is used to predict the remaining lifespan of the battery pack. .

[0071] In some implementations, when 0.3 ≤ λ ≤ 0.7, That is, the weights of the first and second prediction models are each 50%.

[0072] When λ < 0.3 (resource-constrained or poor data): Only the second prediction mode is enabled to ensure basic prediction capability and obtain the predicted value of the remaining lifespan of the battery pack. .

[0073] In one embodiment, the first prediction mode is a GNN-based prediction mode, which quantifies the coupling relationship between individual entities by constructing a graph structure model that quantifies the coupling relationship between multiple physics fields.

[0074] Specifically, the battery pack is abstracted as a graph structure. Here, the node set V represents each battery cell, the edge set E represents the interactions between cells, and the adjacency matrix A quantifies the strength of the interactions. Each node The feature vector is defined as [ (Health status or health assessment) (Ohmic internal resistance). (Polarization resistance). (temperature), (State of charge)], each node It encompasses multiple dimensions of attributes, including health status, internal resistance, temperature, and state of charge. While using individual cell health estimates for battery pack life prediction, it can also collaboratively predict other characteristics to obtain the overall battery pack status. Based on this high-precision and robust collaborative prediction of battery pack status and remaining lifespan, it achieves accurate prediction of battery pack lifespan and improves the safety and reliability of battery pack management.

[0075] As an example, the edge set E contains the following three coupling relationships.

[0076] Electrically connected edges: Based on the series-parallel topology, the weight formula for electrically connected edges is as follows.

[0077]

[0078] The value of directly reflects the tightness of the current coupling between current node i and current node j. When there is a difference in current, the weight will decrease. The greater the difference in current, the lower the weight of the electrical connection, thus quantifying the coupling strength of the current distribution.

[0079] Thermal coupling edge: The battery cells exchange heat through thermal conduction and thermal convection. The closer the distance, the greater the temperature difference and the stronger the heat exchange. The weight formula for thermal coupling edge is as follows.

[0080]

[0081] in, This represents the temperature difference between cell i and cell j in the battery pack; the larger the thermal coupling weight, the stronger the thermal interaction.

[0082] Aging-related edges: Euclidean distance of aging trajectories calculated based on S15 The formula for the weights of aging-related edges is as follows.

[0083]

[0084] Among them, distance The smaller the value, the more similar the aging paths of the two monomers are. The closer to 1, the greater the connection weight, and the more likely it is to identify and propagate similar aging trends.

[0085] The above three edge relationships are combined into a comprehensive adjacency matrix, and the formula for the comprehensive adjacency matrix is ​​as follows.

[0086]

[0087] Among them, electrical connection, as the direct pathway for energy exchange in the battery pack, is assigned the highest weight of 0.4; thermal management and aging consistency are assigned the same and second-highest weight of 0.3. The multi-dimensional interaction relationships are uniformly quantified through a comprehensive adjacency matrix formula. The comprehensive adjacency matrix formula shows the mutual influence of the battery pack in terms of electricity, heat, and aging.

[0088] A graph neural network is constructed and trained to learn and predict the state of the battery pack over a period of time based on the graph structure and the comprehensive adjacency matrix of the battery pack.

[0089] Specifically, a graph neural network (GNN) is introduced. Through message passing, the adjacency matrix is ​​used as the core input of the graph neural network (GNN) to modulate the intensity of information transmission between nodes, allowing nodes in the graph to exchange information with their neighboring nodes, thereby learning the topology of the entire system. By utilizing the message passing mechanism of the GNN, each node can aggregate the state information of its neighbors. The GNN predicts future states (such as the evolution of voltage, temperature, and SOH) based on the current state, thereby explicitly modeling the graph structure relationships defined in S21.

[0090] The message passing function formula for a Graph Neural Network (GNN) is as follows:

[0091] in, This refers to some intermediate calculated value from the j-th node to the i-th node in the l-th layer; It is a non-linear activation function; Here is the weight matrix for the l-th layer; Let be the feature vector of the j-th node in the l-th layer; Let l be the bias vector of the l-th layer; The weights are derived from the composite adjacency matrix.

[0092] The message passing function computes the information sent from node j to node i. The content of the information depends on the state of the neighbors, and the strength is determined by the coupling weights between them. modulation.

[0093] After the message is passed, the node's state is updated by aggregating messages from its neighbors. The specific node update function is as follows.

[0094]

[0095] in, This is the feature vector of the i-th node in the (l+1)-th layer; To modify the activation function of the linear unit; Here is the weight matrix for the l-th layer; Let be the feature vector of the i-th node in the l-th layer; Let be the weight matrix of the l-th layer. To satisfy conditional Perform summation.

[0096] As an example, a multi-head graph attention mechanism (setting 4 attention heads) is introduced, and LeakyReLU (slope=0.2) is used as the activation function, enabling the model to adaptively focus on important neighbors and improve its ability to capture complex coupling relationships. The network training uses the Adam optimizer, with an initial learning rate of 1e-4, 200 training epochs, and a batch size of 32 to ensure stable convergence.

[0097] The first prediction mode, specifically, includes the following steps: starting from the current time, using a trained GNN model, with the combined adjacency matrix of the battery pack at the current time and the historical time relative to the start of the iteration as input, predicting the probability of each cycle in the future period. The value forms a predicted trajectory, and based on the lifespan end threshold, the predicted trajectory is checked backward from the current moment to find the first... The difference between the cycle number at the time point that is less than or equal to the end-of-life threshold and the current time point is used as the predicted remaining lifespan of the battery pack. .

[0098] Because graph neural networks (GNNs) can be unstable under data scarcity or extreme conditions, data cannot be verified if only GNNs are used for prediction. Therefore, it is necessary to establish an enhanced physical model based on the first principles of electrochemistry and thermodynamics to provide reliable baseline predictions even when the GNN fails, thus making the data verifiable. Furthermore, even when sensors malfunction, it can determine whether battery life has failed, making the prediction reliable. This application uses physical model prediction as a second prediction mode.

[0099] The physical model models the terminal voltage of individual cells, analyzes each component, and identifies their role in the battery pack. The formula for the terminal voltage of an individual cell is as follows:

[0100] in, The open-circuit voltage is a function of state of charge (SOC) and temperature. For ohmic voltage drop; This is the polarization overpotential; To correct for thermal effects, the impact of temperature fluctuations on voltage is quantified; This is a cross-current interference term; The coupling coefficient; This is the random error term.

[0101] The single-cell terminal voltage data and information carrier, based on the data components in the formula, displays the current internal state of the battery. As an observation value for inconsistency assessment, it reflects the degree of voltage dispersion between cells. Abnormal voltage of cells (such as too high, too low, or sudden changes) usually indicates direct problems such as overcharging, over-discharging, or internal short circuits. It is a signal that triggers short-term safety warnings and serves as a basis for judging battery life failure when the single-cell voltage deviates significantly.

[0102] The second prediction model, specifically, includes the following steps: When broad health status When the battery level falls below a preset failure threshold, the iteration stops, and the number of virtual cycles elapsed from the current moment is used as the predicted remaining lifespan of the battery pack. Otherwise, repeat the following steps: Obtain the estimated SOH health value for each monomer at time t. Calculate the estimated state of health (SOH) of each cell during the next charging cycle. : , Let be the change in the health estimate of the i-th individual between time t and t+1; Calculate using the formula for calculating the generalized state of health of the battery pack mentioned above. .

[0103] In the fusion mode of the first and second prediction modes, based on the calculation method of the gating coefficient λ, the fusion mode essentially dynamically adjusts the weights of physical drive and data drive according to data quality: when the data quality is high ( →1), Small, The system fully trusts the first prediction pattern. However, when the data quality is low ( →0, such as when there is high noise, many missing values, or sensor failure. big, Small, the system trusts the second prediction pattern.

[0104] Furthermore, in some embodiments, warnings are triggered at different time scales based on the predicted remaining lifespan of the battery pack. Specifically, based on the predicted remaining lifespan of the battery pack, warnings are divided into three categories: short-term warnings, medium-term warnings, and long-term warnings. <20 cycles, triggering a short-term warning, outputting a real-time alert, when <50 cycles, triggering a mid-term warning, outputting a maintenance signal, when, when <100 cycles, output replacement signal.

[0105] It should be noted that the above steps do not have to be performed in order. Instead, they can be performed in reverse order or simultaneously. Furthermore, one or more additional steps can be added to the sequence, and one or more steps can be removed from the above sequence.

[0106] An implementation method for predicting battery pack lifespan quantifies the complex electro-thermal-aging coupling relationships and inconsistent evolution paths between individual cells by extracting deep health features and constructing a physical information graph neural network hybrid model. This achieves high-precision and robust collaborative prediction of battery pack status and remaining lifespan. Furthermore, based on the fused prediction results, a tiered early warning system is triggered, enabling accurate prediction of battery pack lifespan and improving the safety and reliability of battery pack management. Specifically, when the generalized health state... When the battery level falls below a preset failure threshold, the iteration stops, and the number of virtual cycles elapsed from the current moment is used as the predicted remaining lifespan of the battery pack. Otherwise, repeat the following steps: obtain the estimated SOH health value for each monomer at time t. Calculate the estimated state of health (SOH) of each cell during the next charging cycle. : , Let i be the change in the health estimate of the i-th cell between time t and t+1; calculate the generalized health state of the battery pack. .

[0107]

[0108] This represents the current capacity degradation value. The maximum capacity degradation threshold, This is the value of the internal resistance divergence. The maximum internal resistance divergence threshold, and These are the weighting coefficients.

[0109]

[0110] This represents the minimum SOH value of all cells within the battery pack. This represents the average SOH of all monomers within the group. Let S be the standard deviation of the state of harmonics (SOH) of all cells within the battery pack. Starting from the current time, using a trained GNN model, with the combined adjacency matrix of the battery pack at the current time and the historical data from the relative iteration start time as input, predict the state of harmonics for each cycle in the future. The value forms a predicted trajectory, and based on the lifespan end threshold, the predicted trajectory is checked backward from the current moment to find the first... The difference between the cycle number at the time point that is less than or equal to the end-of-life threshold and the current time point is used as the predicted remaining lifespan of the battery pack. ; Final predicted remaining lifespan of the battery pack λ is the preset weight The second embodiment of the present invention relates to a network-side server, such as... Figure 2 As shown, it includes at least one processor 302; and a memory 301 communicatively connected to at least one processor 302; wherein the memory 301 stores instructions executable by at least one processor 302, the instructions being executed by at least one processor 302 to enable at least one processor 302 to perform the above-described data processing method.

[0111] The memory 301 and processor 302 are connected via a bus, which may include any number of interconnecting buses and bridges. The bus connects various circuits of one or more processors 302 and memory 301 together. The bus can also connect various other circuits, such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. A bus interface provides an interface between the bus and the transceiver. The transceiver can be a single element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. Data processed by processor 302 is transmitted over a wireless medium via an antenna, which further receives data and transmits it to processor 302.

[0112] Processor 302 is responsible for managing the bus and general processing, and can also provide various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. Memory 301 can be used to store data used by processor 302 during operation.

[0113] The fourth embodiment of the present invention relates to a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the battery pack life prediction method of the first embodiment.

[0114] The steps of the various methods described above are only for clarity. In practice, they can be combined into one step or some steps can be split into multiple steps. As long as they include the same logical relationship, they are all within the scope of protection of this patent. Adding insignificant modifications or introducing insignificant designs to the algorithm or process, but without changing the core design of the algorithm and process, are also within the scope of protection of this patent.

[0115] The above descriptions are merely embodiments of the present invention. Commonly known structures and characteristics are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

[0116] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A battery pack life prediction method characterized by, Comprising: Respectively preset scores corresponding to CPU / GPU utilization of the system, battery pack inconsistency characteristics and data quality, and obtain the corresponding scores according to the CPU / GPU utilization of the current system, battery pack inconsistency characteristics and data quality, respectively denoted as , and ; Computes , is a bias term, , and is a weight; According to the value of λ, the following prediction modes are performed: when λ > 0.7, a first prediction mode is used to predict the remaining useful life of the battery pack ; when λ < 0.3, a second prediction mode is used to predict the remaining useful life of the battery pack ; When 0.3≤λ≤0.7, the remaining service life prediction value of the battery pack is obtained by using a fusion mode of the first prediction mode and the second prediction mode for prediction .

2. The method of claim 1, wherein, The second prediction mode comprises the following steps: When the generalized health state If the generalized health state is lower than the preset failure threshold, the iteration is stopped, and the experienced number of virtual cycles from the current time is taken as the remaining service life prediction value of the battery pack ; otherwise, repeat the following steps: SOH health estimate value of each monomer at time t , calculate the SOH health estimate value of each monomer when the next charging cycle is performed : , is the health estimate value change amount of the i th monomer between time t and t+1 Computing a generalized state of health of a battery pack : is a current capacity degradation value, is a maximum capacity degradation threshold value, is an internal resistance divergence value, is a maximum internal resistance divergence threshold value, and is a weight coefficient; ； the minimum value of the SOH of all monoblocs within the battery pack; the average value of the SOH of all monoblocs within the battery pack, the standard deviation of the SOH of all monoblocs within the battery pack.

3. The method of claim 2, wherein, Health estimate change amount The following formula is used: In the formula, is the length of the i-th time interval; is the aging rate of each single cell in the battery pack, is a pre-factor, obtained by fitting a large amount of accelerated aging experimental data, denotes the sensitivity of the aging reaction to temperature, is the activation energy, which denotes the minimum energy required for the reaction to occur, is the gas constant, is the temperature;​ for measuring inconsistency in a battery pack, is a magnitude of the i-th cell load current, is a maximum temperature gradient in the battery pack, is an initial capacity inconsistency.​ 4. The method of claim 1, wherein, The first prediction mode comprises the following steps: From the current moment, using the trained GNN model, taking the graph structure and the integrated adjacency matrix of the battery pack at the current moment and the history of the relative iteration start moment as input, predicting the value of each cycle in the future period , forming a prediction trajectory, and according to the life termination threshold, checking the prediction trajectory backward from the current moment, finding the first time point less than or equal to the life termination threshold, taking the cycle number difference between the time point and the current moment as the remaining service life prediction value of the battery pack ; The graph structure is a graph constructed by taking the battery monomer as the node and the interaction between the battery monomers as the edge, and the node attribute includes the SOH health estimation value of the monomer; The integrated adjacency matrix is a matrix on the interaction strength between battery cells, the matrix element : , and is the current size of the current-coupled cell i and cell j, represents the temperature difference between cell i and cell j in the battery pack.

5. The method of claim 2, wherein, The method further comprises triggering a warning of different time scales based on the remaining life prediction result of the battery pack.

6. The method of claim 2, wherein, The method further comprises preprocessing the collected data in the battery pack, comprising: Adopting 3σ criterion to remove outliers of voltage and current data, filtering out transient interference and sensor noise; For missing data caused by communication interruption or temporary sensor failure, linear interpolation method is used for completion to ensure the continuity and integrity of the data sequence; Carrying out Z-score normalization processing on voltage, current, temperature and environmental parameters, unifying dimension, and eliminating the numerical scale difference between parameters; Based on the time stamp, the time sequence of the multi-source sensor data is aligned, the time deviation between the signal channels is controlled within a preset range, and the spatio-temporal consistency of the subsequent feature extraction and state estimation is ensured.

7. A battery pack life prediction method characterized by, Comprising: When the generalized health state If the generalized health state is lower than a preset failure threshold, the iteration is stopped, and the experienced number of virtual cycles from the current time is taken as the remaining useful life prediction value of the battery pack Otherwise, the following steps are repeated: SOH health estimate value of each monomer at time t , calculate the SOH health estimate value of each monomer when the next charging cycle is performed : , is the health estimate value change amount of the i th monomer between time t and t+1 Computing a generalized state of health of a battery pack : is a current capacity degradation value, is a maximum capacity degradation threshold value, is an internal resistance divergence value, is a maximum internal resistance divergence threshold value, and is a weight coefficient; ； the minimum value of the SOH of all monoblocs within the battery pack; the average value of the SOH of all monoblocs within the battery pack, the standard deviation of the SOH of all monoblocs within the battery pack; From the current moment, using the trained GNN model, taking the integrated adjacency matrix of the battery pack at the current moment and the history of the relative iteration start moment as input, predicting each cycle of the battery pack in the future period Value, form a prediction trajectory, and according to the life termination threshold, check the prediction trajectory backward from the current moment, find the first Time point less than or equal to the life termination threshold, and the cycle number difference between the time point and the current moment as the remaining service life prediction value of the battery pack ; Final battery pack remaining useful life prediction value where λ is a predetermined weight.

8. A battery pack state of health monitoring method characterized by, Comprising: A multi-level perception network is constructed to collect multi-dimensional operating parameters of three spatial levels of monomers, modules and battery packs, and the multi-level perception network is divided into monomer level sensing unit, module level sensing unit and battery pack level sensing unit; wherein the monomer level sensing unit is configured to monitor voltage, current and surface temperature; the module level sensing unit is configured to use infrared thermal imaging or distributed optical fiber sensor to capture the thermal field distribution inside the module in real time; the battery pack level sensing unit is configured to integrate environmental temperature and humidity, vibration and air pressure sensors to realize comprehensive perception of the operating environment; Based on different battery pack operating states, adaptive collection of operating data is carried out, including: enabling high-frequency collection during charging to monitor voltage increment and temperature rise rate; during discharging, a high-frequency and medium-frequency combination strategy is adopted to balance power output stability and thermal runaway warning; in the static stage, a low-frequency inspection mode is switched to to continuously track self-discharge behavior and voltage relaxation process, avoiding resource occupation; The collected operating data is preprocessed, and battery health state related features are extracted based on the processed operating parameters, including capacity inconsistency coefficient, internal resistance divergence, maximum temperature gradient and voltage distribution entropy; the capacity inconsistency coefficient is obtained by calculating the ratio of the standard deviation to the average value of the available capacity of the monomers in the battery pack; the internal resistance divergence is the difference degree of the AC internal resistance of each monomer, which is obtained by fitting the pulse response data in the charging and discharging process; the voltage distribution entropy is obtained by quantifying the dispersion degree of the voltage curve under static or dynamic working conditions based on information entropy theory, and the voltage value distribution of all monomers in the group is discretized into multiple intervals to calculate the information entropy value.

9. The method of claim 8, wherein, The method comprises: Using the SOH (State of Health) degradation sequence of each monobloc over its lifetime, the aging trajectory Euclidean distance is calculated: , and are the health estimations of the i, j-th monobloc at time t, T is the total number of monitoring times. Based on the Euclidean distance of the aging trajectory, the average value and the maximum value are calculated, and the health state related features of the battery pack are taken as inconsistency features.

10. The method of claim 8, wherein, The method also includes calculating a generalized state of health of the battery pack : is a current capacity degradation value, is a maximum capacity degradation threshold value, is an internal resistance divergence value, is a maximum internal resistance divergence threshold value, and is a weight coefficient; ； the minimum value of the SOH of all monoblocs within the battery pack; the average value of the SOH of all monoblocs within the battery pack, the standard deviation of the SOH of all monoblocs within the battery pack.

Images