Storage battery intelligent early warning and service life prediction system and method fusing physical constraints

Abstract

The invention discloses a storage battery intelligent early warning and service life prediction system and method fusing physical constraints. Belongs to the technical field of battery monitoring. According to the system, voltage, current, temperature and SOC data of a storage battery are acquired in real time through an intelligent acquisition unit, high-precision prediction of the health state is realized by fusing an artificial intelligence unit with a physical constraint correction layer based on a CNN-BiLSTM-SelfAttention model, and the residual service life is predicted by adopting an AE-CNN-BiLSTM model in combination with a time attention mechanism; and meanwhile, the batteries are dynamically grouped through an impedance spectrum clustering algorithm, and early fault early warning is realized by utilizing five-dimensional feature anomaly detection. The method overcomes the hysteresis defect of traditional single-parameter threshold alarm, remarkably improves the accuracy and operation and maintenance efficiency of battery life prediction, and is suitable for full-life-cycle management of the storage battery under complex working conditions.

Description

Technical Field

[0001] This invention relates to the field of battery monitoring technology, and more specifically to a smart early warning and life prediction system for batteries that integrates physical constraints. Background Technology

[0002] In the field of existing battery monitoring technology, traditional monitoring solutions basically adopt a single-parameter threshold alarm mechanism, which can only realize fault detection but cannot predict faults. For example, by collecting parameters such as battery voltage, current, and temperature, it is only possible to estimate the remaining charge (SOC) or state of health (SOH); the threshold judgment of temperature parameters can only issue an alarm when thermal management fails, but cannot provide early warning and can only be dealt with after the fact; the monitoring of a single parameter cannot detect common internal aging problems such as plate sulfation, making it difficult to predict battery failures and remaining battery life, resulting in untimely maintenance and incidental losses. Summary of the Invention

[0003] To address the aforementioned shortcomings in the existing technology, this invention provides a battery intelligent early warning and life prediction system that integrates physical constraints.

[0004] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: A battery intelligent early warning and life prediction system integrating physical constraints is characterized by comprising: an intelligent acquisition unit for acquiring battery parameters; an artificial intelligence unit connected to the intelligent acquisition unit for data fusion analysis and early warning; a communication unit connected to the artificial intelligence unit for data interaction; and a human-computer interaction platform connected to the communication unit for visualized operation and maintenance management.

[0005] Furthermore, the intelligent acquisition unit is configured to acquire the voltage, current, temperature, and remaining power (SOC) of a single battery, and is connected to the artificial intelligence unit using a parallel bus architecture.

[0006] Furthermore, the artificial intelligence unit includes a data processing module and a grouping early warning module. The data processing module calculates the state of health (SOH) using the CNN-BiLSTM-SelfAttention algorithm, while the grouping early warning module performs battery grouping, anomaly detection, and remaining useful life (RUL) calculation.

[0007] Furthermore, the data processing module includes a Convolutional Neural Network (CNN) layer, a Bidirectional Long Short-Term Memory (BiLSTM) layer, a self-attention mechanism layer, a fully connected layer, and a physical constraint correction layer.

[0008] Furthermore, the grouping early warning module adopts a dynamic grouping equalization strategy based on impedance spectrum clustering, and divides the battery into a healthy group, a degradation group and a fault early warning group through the XGBoost algorithm.

[0009] Furthermore, the communication unit supports both wired and wireless communication methods, enabling local data interaction and remote platform data uploading.

[0010] A method for intelligent early warning and life prediction of batteries that integrates physical constraints includes the following steps: S1. Install and deploy system hardware and complete network configuration, specifically including installing intelligent acquisition unit, artificial intelligence unit and communication unit, and configuring chain cascade connection and network parameters; S2. Load the artificial intelligence algorithm model and set the parameters, specifically including loading the CNN-BiLSTM-SelfAttention algorithm for calculating the state of health (SOH), the dynamic grouping and equalization strategy based on impedance spectrum clustering for battery grouping, and the AE-CNN-BiLSTM model for calculating the remaining useful life (RUL). S3. Conduct system testing and parameter optimization, specifically including verifying the accuracy of data acquisition and adjusting algorithm parameters to ensure output accuracy; S4. Achieve visualized operation and maintenance management through a human-computer interaction platform, specifically including deploying digital twin models, real-time monitoring, and alarm linkage verification.

[0011] Furthermore, step S2, which uses the CNN-BiLSTM-SelfAttention model to predict SOH, includes the following sub-steps: S21. Perform outlier removal and standardization on the collected voltage, current, temperature and time series data to generate a sample set; S22. Use convolution kernels to extract local spatial features and reduce dimensionality, and reduce input redundancy through convolution operations; S23. Capture the long-term dependencies of the charging and discharging process through input gate, forget gate and output gate mechanisms; S24. Calculate the correlation between each part of the sequence and the SOH state, and dynamically allocate attention weights to focus on key features; S25. After fusing the deep features, output the preliminary SOH prediction value; S26. Calculate the residuals based on the physical equation of battery capacity decay, and construct a comprehensive loss function to correct the predicted values.

[0012] Furthermore, the anomaly detection and fault warning in step S3 includes the following sub-steps: S31. Collect battery AC impedance spectrum data, extract impedance amplitude and phase angle features, use principal component analysis algorithm for dimensionality reduction, and then use extreme gradient boosting algorithm for clustering to divide the battery into healthy group, degraded group and fault warning group. S32. A five-dimensional feature space is constructed based on voltage, internal resistance, temperature, SOC, and SOH to comprehensively characterize the static and dynamic operating characteristics of the battery; S33. Using a single-class support vector machine algorithm, a radial basis kernel function is introduced to construct the minimum hypersphere that encloses normal data in five-dimensional space; S34. When the real-time sample features exceed the boundary of the hypersphere, it is judged as an anomaly and an early warning is triggered.

[0013] Furthermore, step S2, which uses the AE-CNN-BiLSTM model to predict RUL, includes the following sub-steps: S41. Local features of time series are extracted through convolutional neural network layers, bidirectional long short-term memory network layers are used to capture forward and backward time dependencies, and autoencoders are used for feature compression and noise reduction. S42. A time attention mechanism is introduced to assign weights to different time steps, highlighting key moments of influence. Then, a bidirectional long short-term memory network layer is used to mine long-term trends, and a fully connected layer outputs the RUL prediction value. S43. Using Gaussian process regression as a surrogate model, the optimal parameters are searched globally through the acquisition function to adaptively optimize the model performance.

[0014] The present invention has the following beneficial effects: (1) Statistical regression methods based on a single or few features, such as linear regression and support vector regression, cannot make full use of multi-source time series data and have limited prediction accuracy; (2) Time series prediction methods based on single neural network structures such as LSTM and GRU have the problem of weak feature extraction ability and difficulty in capturing local spatiotemporal features; (3) Some literature uses CNN or RNN, but most of them do not achieve multi-structure fusion and are insufficient for modeling charging and discharging data under complex working conditions; (4) Most data-driven models rely solely on historical sample fitting and lack physical interpretability, which can easily lead to physical inconsistencies and poor generalization ability across operating conditions.

[0015] This algorithm achieves innovation through structural optimization, physical fusion, and parameter adaptation, specifically addressing the following issues: traditional CNNs can only extract spatial features, BiLSTM lacks spatial dimensionality reduction capabilities, and single SelfAttention is computationally redundant; data-driven models deviate from the physical laws of battery capacity decay, leading to large prediction biases under extreme conditions (high and low temperatures, high current); and the problem of noisy collected data and weak feature correlation, resulting in insufficient model generalization ability. Attached Figure Description

[0016] Figure 1 The present invention provides a system hardware topology diagram for an artificial intelligence-based intelligent early warning and physical constraint correction lifespan system for batteries.

[0017] Figure 2 The diagram illustrates the implementation steps of an artificial intelligence-based intelligent early warning and physical constraint correction lifespan system for batteries, as provided by this invention.

[0018] Figure 3 The present invention provides a flowchart of SOH prediction for an artificial intelligence-based intelligent early warning and physical constraint correction life system for batteries.

[0019] Figure 4 The present invention provides a flowchart of a battery dynamic adjustment grouping system based on artificial intelligence for intelligent early warning and physical constraint correction of battery life.

[0020] Figure 5 The present invention provides a flowchart for an anomaly detection and fault early warning system for an intelligent early warning and physical constraint correction battery life system based on artificial intelligence. Figure 6 The present invention provides a flowchart of the RUL prediction process for an artificial intelligence-based intelligent early warning and physical constraint correction battery life system. Detailed Implementation

[0021] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0022] Please refer to Figure 1 This embodiment provides an artificial intelligence-based intelligent early warning and physical constraint correction lifespan system for batteries, including: The intelligent acquisition unit is used to collect the voltage, current, temperature, and SOC of a single battery. The artificial intelligence unit includes a data processing module and a grouping early warning module. The data processing module uses artificial intelligence algorithms to dynamically fuse and analyze various parameters collected and calculated by the intelligent acquisition unit, and outputs the predicted value of SOH. The grouping early warning module groups the batteries in the battery pack according to the artificial intelligence equalization strategy, uses five-dimensional feature anomaly detection to realize anomaly identification and fault early warning, and uses an algorithm model to predict RUL. The communication unit is used for local data interaction and remote platform data upload, and supports wireless or wired communication; A human-computer interaction platform for visual early warning and operation and maintenance management.

[0023] In this embodiment, a total of 32 intelligent acquisition units, 1 artificial intelligence unit, 1 communication unit, and 1 human-computer interaction platform are employed. All 32 intelligent acquisition units adopt a parallel bus architecture, with physical wiring following a chain-cascade principle, and are wired to the artificial intelligence unit. The artificial intelligence unit is connected to the communication unit via a network cable. The human-computer interaction platform is deployed remotely and is wired to the communication unit.

[0024] refer to Figure 2 In this embodiment, the installation and deployment steps of an artificial intelligence-based intelligent early warning and physical constraint correction lifespan system for batteries are as follows: Step S1, Hardware Installation and Deployment: Install the intelligent data acquisition unit, configure and deploy the artificial intelligence unit, configure and debug the communication unit, and deploy and debug the human-machine interaction platform. Ensure the equipment is working properly.

[0025] Step S1 includes: Step S11: Unscrew the bolts securing the battery busbars. Place the voltage probe of the intelligent acquisition unit onto the unscrewed bolts, with the current probe passing through the busbars. The voltage probe on the negative terminal integrates temperature monitoring functionality. Then, screw the bolts back on, restoring the battery connection to its pre-installation state. Insert the device connection cable into the device, following the chain-cascade principle, to connect all 32 intelligent units, ultimately connecting them to the artificial intelligence unit.

[0026] Step S12: Configure and deploy the network parameters for the artificial intelligence unit and the communication unit, and install and deploy the human-computer interaction platform. Perform network debugging to ensure smooth device interconnection.

[0027] Step S2: Load the algorithms and models of the data processing module and group early warning module of the artificial intelligence unit, and fill in the relevant parameters.

[0028] Step S2 includes: Step S21: Load the data processing module algorithm: CNN-BiLSTM-SelfAttention, for predicting SOH.

[0029] In existing technologies, the prediction of battery state of harm (SOH) mainly relies on the following types of methods, and their drawbacks are as follows: (1) Statistical regression methods based on a single or few features, such as linear regression and support vector regression, cannot make full use of multi-source time series data and have limited prediction accuracy; (2) Time series prediction methods based on single neural network structures such as LSTM and GRU have the problem of weak feature extraction ability and difficulty in capturing local spatiotemporal features; (3) Some literature uses CNN or RNN, but most of them do not achieve multi-structure fusion and are insufficient for modeling charging and discharging data under complex working conditions; (4) Most data-driven models rely solely on historical sample fitting and lack physical interpretability, which can easily lead to physical inconsistencies and poor generalization ability across operating conditions.

[0030] This algorithm achieves innovation through structural optimization, physical fusion, and parameter adaptation, specifically addressing the following issues: traditional CNNs can only extract spatial features, BiLSTM lacks spatial dimensionality reduction capabilities, and single SelfAttention is computationally redundant; data-driven models deviate from the physical laws of battery capacity decay, leading to large prediction biases under extreme conditions (high and low temperatures, high current); and the problem of noisy collected data and weak feature correlation, resulting in insufficient model generalization ability.

[0031] The core improvements of this algorithm are: Hierarchical collaborative structure: CNN-BiLSTM-SelfAttention concatenation optimization to achieve full-link coverage of "spatial feature extraction → temporal dependency capture → key feature focusing"; Physical constraint fusion: The physical equation of battery degradation is embedded into the loss function to correct the prediction bias of the pure data-driven model; Engineering adaptation: Optimize network parameters and preprocessing logic based on the characteristics of actual collected data to improve applicability in industrial scenarios.

[0032] The algorithm is implemented through a complete process of "preprocessing → feature extraction → temporal modeling → attention focusing → physical correction → output". The flowchart of the entire algorithm is shown below. Figure 3 The practical application details of each part are as follows: Data preprocessing: The core functions are noise reduction and feature processing, with the input current collected by the intelligent acquisition unit ( I ),Voltage( V ),temperature( T ), number of cycles ( N Multidimensional parameters such as σ are used to remove outliers based on the 3σ criterion, and the time step is set according to the acquisition frequency to obtain the sample.

[0033] CNN Layers: Their core functions are spatial feature extraction and dimensionality reduction. A CNN is a multi-layer feedforward neural network that uses convolutional computation, typically consisting of an input layer, convolutional layers, pooling layers, fully connected layers, and an output layer. The formula for a convolutional layer in a CNN is as follows: (1) in, C This is the output of the convolutional layer. X This is the input to the previous layer, and "*" represents the convolution operation. W For convolution kernel, B For offset vectors, fis the activation function. The role of convolutional layers is to extract features from a local region. Different convolutional kernels are equivalent to different feature extractors. The role of pooling layers is to perform feature selection, reduce feature dimensionality, and avoid overfitting.

[0034] BiLSTM layer: Its core function is long-term time dependency capture. LSTM selectively filters and retains state information through three special gating structures: input gate, output gate, and forget gate. The specific calculation formula is as follows: (2) (3) (4) (5) (6) (7) in, For the Sigmoid function, i t The input gate output is [output], and the forget gate output is [output]. o t For output gate output, This represents the candidate state at the current moment. c t The internal state at the current moment. h t The external state at the current moment. W and b These are the weight matrix and offset of the corresponding gate, respectively.

[0035] Specifically, the Gate of Oblivion f t Used to control the internal state of the previous time step. c t-1 The part that needs to be forgotten, the input gate i t Determine the candidate state at the current moment How much information is written to the new internal state, and what is the output gate? o t This is used to adjust the internal state at the current moment. How much information needs to be output to the external state? h t .

[0036] Building upon this foundation, BiLSTM consists of two LSTM subnetworks, one forward and one backward, capable of simultaneously modeling the forward and backward dependencies of a sequence. Compared to unidirectional LSTM, BiLSTM can more fully uncover the implicit abstract features in time series data, leveraging bidirectional information to enhance the understanding of feature evolution patterns. This significantly improves prediction performance and effectively overcomes the lag problem caused by unidirectional models that can only capture one-sided dependencies.

[0037] The SelfAttention layer's core function is to dynamically allocate weights and focus on key features. SelfAttention is an improvement on the attention mechanism, selectively focusing on certain important information while ignoring other information received at the same time. Its core idea is to dynamically allocate weights based on the relevance of different features, giving higher weights to important information for accurate modeling and lower weights to less important information, thereby improving the effectiveness of information processing and the efficiency of system resource utilization. Through this mechanism, the model can capture the dependencies between features globally, thus enhancing the ability to extract sequence features. The specific calculation formula is as follows: (8) In the formula, Q , K and V These represent three types of vectors: query, key, and value. d k for Q and V The second dimension, T, represents the transpose operation. The output of this expression is the attention value.

[0038] Fully connected layer: Its core function is to output preliminary SOH prediction values. It mainly fuses the deep features extracted by the three layers of CNN, BiLSTM, and SelfAttetion to output preliminary SOH prediction values.

[0039] Physical constraint correction layer: Its core function is to solve the problem of data-driven models deviating from physical laws by constructing constraints in conjunction with the physical equations of battery capacity decay.

[0040] The effective capacity of a battery gradually decreases with use, primarily due to both cycle aging and calendar aging. Cycle aging is caused by the charging and discharging process, while calendar aging accumulates over time under static conditions. The combined effect of these two factors determines the overall degradation trend of the battery's performance. The battery's State of Harmony (SOH) is essentially the percentage of its effective capacity relative to its rated capacity. (9) In the formula C rated The rated capacity of the battery. Cactual This represents the battery's currently available effective capacity. Therefore, effective capacity directly characterizes the battery's state of equilibrium (SOH).

[0041] Studies have shown that calendar aging accumulates over time, gradually reducing the effective capacity of a battery. This aging process is primarily driven by slow-moving side reactions within the battery, is irreversible, and its degradation rate is typically lower than that of cycle aging. The capacity change caused by calendar aging can be described by the following formula: (10) (11) In the formula, E cap This represents the percentage of the battery's remaining effective capacity. α cap For an open-circuit voltage of Vol (unit: volt) and a temperature of K The attenuation coefficient at (unit: Kelvin) t It refers to the number of days.

[0042] Studies have shown that battery cycle aging is affected by factors such as the degree of discharge, the number of cycles, the discharge rate, and temperature, and is also related to the battery's current SOC and its own degree of aging.

[0043] To assess the degree of battery life degradation, the cumulative throughput of the battery is used to reflect its degradation rate, and the depth of discharge is defined. D The lifetime degradation factor is: (12) in, E m This refers to the total amount of electricity that the battery can handle during its lifespan. E mB This represents the total amount of energy the battery can handle at its rated depth of discharge. E R The rated capacity of the battery. N For the battery at a depth of discharge of D The number of charge-discharge cycles before reaching the end of its lifespan.

[0044] The discharge rate C of a battery refers to the current value output by the battery when it discharges its rated capacity within a specified time. It is numerically equal to the rated capacity. Studies have shown that the battery discharge capacity and the battery discharge rate are roughly negatively correlated, and the discharge rate decreases as the discharge rate increases.

[0045] The effect of temperature on a battery can be characterized by the Arrhenius equation: (13) In the formula, KAbsolute temperature k Temperature T The reaction rate constant at that time, A Pre-exponential factor, E a For the experimental activation energy, R is the molar gas constant, with a value of 8.314.

[0046] This formula shows that temperature is positively correlated with the rate of chemical reaction, and changing the temperature can significantly change the discharge capacity of the battery.

[0047] This algorithm predicts an MAE of 0.012 and an RMSE of 0.018, significantly improving accuracy. Furthermore, it enhances robustness by more than 30% in complex scenarios such as temperature fluctuations (-10℃ to 55℃) and sudden current changes. This solves the practical problems of existing models, such as large prediction bias under complex conditions and weak generalization ability due to the lack of integration with physical laws.

[0048] Step S22: Load the AI ​​equalization strategy of the grouping early warning module: dynamic grouping equalization based on impedance spectrum clustering. Based on PCA and XGBoost algorithms, the principal components of the impedance spectrum are clustered to divide the batteries into healthy groups, degraded groups, and fault early warning groups. The algorithm's flowchart is as follows: Figure 4 The specific steps are as follows: The AC impedance spectrum data of the battery under different operating conditions are collected, and key characteristic quantities such as impedance amplitude and phase angle are extracted within the preset characteristic frequency band to reflect the internal electrochemical reaction characteristics and aging degree of the battery.

[0049] PCA algorithm is used to reduce the dimensionality of multidimensional impedance features in order to extract the key feature components that best distinguish differences in battery health status. The basic idea of ​​PCA is to map the high-dimensional feature set to a low-dimensional space while preserving the overall change information of the original features, that is, to... M 3D feature projection to N (< M On the dimension, this N This entirely new feature is also called the principal component. Given a data matrix A=[a1,…,a…] D ]∈R M×D It records the M A characteristic random variable D PCA finds a single observation by minimizing the error function. N ≤min{ M , D The linear subspace of dimension} is calculated using the following formula: (14) In the formula, X∈R M×N It is a projection matrix, C∈R D×N It is a coefficient matrix.

[0050] Cluster analysis of the extracted impedance spectrum principal components is performed using the XGBoost algorithm to achieve adaptive grouping of battery health status, including healthy, deteriorating, and fault warning groups. The XGBoost algorithm is an augmented form of gradient boosting tree models, primarily optimizing the objective function, optimization method, missing value handling, and overfitting prevention. The core of the XGBoost algorithm is gradient boosting, which involves iteratively training a series of weak learners (usually decision trees). Each iteration attempts to correct the error of the previous iteration, ultimately combining these weak learners into a strong learner and using gradient descent to minimize the loss function, thereby continuously improving the model's predictive ability. For an n-item m-dimensional dataset, the XGBoost calculation formula is as follows: (15) in, . q A tree structure that maps samples to leaf nodes. T The number of leaf nodes. ω For each leaf node, there is a real fraction.

[0051] To find the optimal parameters and establish the optimal model, the objective function of the XGBoost algorithm is a combination of the loss function and the regularization term, expressed as follows: (16) in, γT yes L The regular term of 1, yes L The regular term for 2.

[0052] The algorithm first analyzes AC impedance spectroscopy data using PCA to extract the main feature components reflecting differences in battery health status. Then, it performs cluster analysis on the obtained principal components based on XGBoost to adaptively classify batteries into healthy, deteriorating, and fault warning groups. In practical applications, impedance spectroscopy data may be affected by sensor errors and environmental noise. Compared to traditional methods, the combination of PCA and XGBoost has stronger anti-interference capabilities, effectively tolerating and correcting noise and abnormal data during feature extraction and classification, thereby reducing the impact of external interference on the grouping results and ensuring the stability and reliability of health status grouping and warning determination.

[0053] Step S23: Load the anomaly detection and fault warning method of the grouped early warning module: five-dimensional feature anomaly detection. Algorithm flow reference. Figure 5 The details are as follows: First, a five-dimensional feature space including battery voltage, internal resistance, temperature (SOC), and SOH is constructed to comprehensively characterize the static and dynamic operating characteristics of the battery.

[0054] Next, based on the One-Class SVM algorithm, the sample data under normal working conditions is trained in a five-dimensional feature space. By introducing RBF, the data is mapped to the five-dimensional space, and a "hypersphere" or "hyperplane" with the smallest volume that encloses the normal data distribution is constructed in this space. Samples that deviate from this region are then identified as anomalies. The specific calculation process is as follows: Assuming the generated hypersphere parameters are center o and corresponding hypersphere radius r > 0, the hypersphere volume VI is minimized, and center o is a linear combination of support lines; it can be required that the distance from all training data points xi to the center is strictly less than r. However, a slack variable ζi with a penalty coefficient C is constructed simultaneously, and the optimization problem is as follows:

[0055]

[0056]

[0057] After using Lagrange duality to solve, we can determine whether a new data point z is inside the hypersphere. If the distance from z to the center is less than or equal to the radius r, it is not an outlier. If it is outside the hypersphere, it is an outlier.

[0058] Step S24: Load the algorithm model of the grouped early warning module: AE-CNN-BiLSTM model, used to calculate the remaining useful life (RUL). Existing technical challenges: Traditional models struggle to handle the "high noise – strong temporal sequence – nonlinearity" characteristics of battery monitoring data (such as voltage spikes caused by charging and discharging fluctuations, and feature distortion caused by temperature drift); battery lifecycle data exhibits nonlinear distribution differences of "early redundancy – mid-term complexity – late-term sparsity," making it difficult for traditional models to balance local spatiotemporal features (such as voltage fluctuations) with long-term forward and backward dependencies (such as cycle decay trends); model parameters (such as convolution kernel size and hidden layer dimensions) rely on manual tuning, easily getting trapped in local optima, resulting in poor generalization across different models or operating conditions.

[0059] This algorithm achieves a breakthrough through an integrated design of "encoder-predictor-Bayesian adaptive optimization," specifically addressing the aforementioned pain points. Its structure, calculation method, and parameter explanations are as follows: Data preprocessing: The core functions are data cleaning and sample construction. Input: Multi-dimensional core parameters collected by the battery monitoring unit; anomaly handling based on thresholds and filtering; obtaining samples by setting the time step according to the collection frequency. Algorithm flow reference. Figure 6 The details are as follows: The encoder module's core functions are feature extraction, noise reduction and compression, and dimension adaptation. It mainly consists of CNN layers, BiLSTM layers, and AE layers.

[0060] CNN Layers: Their core functions are spatial feature extraction and dimensionality reduction. A CNN is a multi-layer feedforward neural network that uses convolutional computation, typically consisting of an input layer, convolutional layers, pooling layers, fully connected layers, and an output layer. The formula for a convolutional layer in a CNN is as follows: (17) in, C This is the output of the convolutional layer. X This is the input to the previous layer, and "*" represents the convolution operation. W For convolution kernel, B For offset vectors, f is the activation function. The role of convolutional layers is to extract features from a local region. Different convolutional kernels are equivalent to different feature extractors. The role of pooling layers is to perform feature selection, reduce feature dimensionality, and avoid overfitting.

[0061] 2. BiLSTM Layer: Its core function is long-term time dependency capture. LSTM selectively filters and retains state information through three special gating structures: input gates, output gates, and forget gates. The specific calculation formula is as follows: (18) (19) (20) (twenty one) (twenty two) (twenty three) in, For the Sigmoid function, i t The input gate output is [output], and the forget gate output is [output]. o t For output gate output, This represents the candidate state at the current moment. c t The internal state at the current moment. h t The external state at the current moment. W and b These are the weight matrix and offset of the corresponding gate, respectively.

[0062] Specifically, the Gate of Oblivion f t Used to control the internal state of the previous time step.c t-1 The part that needs to be forgotten, the input gate i t Determine the candidate state at the current moment How much information is written to the new internal state, and what is the output gate? o t This is used to adjust the internal state at the current moment. How much information needs to be output to the external state? h t .

[0063] Building upon this foundation, BiLSTM consists of two LSTM subnetworks, one forward and one backward, capable of simultaneously modeling the forward and backward dependencies of a sequence. Compared to unidirectional LSTM, BiLSTM can more fully uncover the implicit abstract features in time series data, leveraging bidirectional information to enhance the understanding of feature evolution patterns. This significantly improves prediction performance and effectively overcomes the lag problem caused by unidirectional models that can only capture one-sided dependencies.

[0064] 3. AE: Primarily compresses dimensions while removing noise and redundant information. The process mainly consists of two parts.

[0065] (1) Encoding layer: The input feature vector is encoded. x Transform into latent feature vectors h The calculation formula is as follows: (twenty four) In the formula, W e This is the weight matrix in the encoder; b e This is the bias vector in the encoder; σ e This is the activation function in the encoder.

[0066] (2) Decoding layer: converts the latent feature vectors h Reconstructed into output feature vectors z The calculation formula is as follows: (25) In the formula, W d This is the weight matrix in the decoder; b d This is the bias vector in the decoder; σ d This is the activation function in the decoder.

[0067] By training the encoder-decoder (AE) to reconstruct the input features and adjusting the encoder and decoder parameters, the error between the reconstructed result and the input is minimized. Reconstruction error function. Jas follows: (26) The encoder module connects the After Effects (AE) module with the CNN-BiLSTM module, rather than using them independently, thus solving the problem that "traditional AEs can only reduce noise but cannot capture temporal correlations," and through... J Constraints are applied to remove sensor noise (such as temperature fluctuation interference) and redundant information (such as invalid data from repeated charge and discharge cycles).

[0068] Predictor module: Introduces a temporal attention mechanism (for key impact nodes of battery RUL), optimizes the long-term dependency capture capability of LSTM, and achieves accurate mapping through a fully connected layer.

[0069] 1. Time-Attention Mechanism: Addressing the characteristic of "sudden capacity drop in the later stages of battery cycle life," attention weight is shifted towards time steps with high impact (e.g., stages with more than 80% of the battery's lifespan in cycles). Calculation logic: Attention Score , Where v is the attention vector, W a and b a Learnable parameters; weighted features: ( The weighted global features highlight key time-series nodes. The time-series attention mechanism can automatically identify the "inflection point" of battery capacity decay (such as the rate of capacity decay increasing from 0.5% / cycle to 2% / cycle after a certain cycle), improving prediction sensitivity.

[0070] 2. BiLSTM Layer: Its core function is long-term time dependency capture. LSTM selectively filters and retains state information through three special gating structures: input gates, output gates, and forget gates. The specific calculation formula is as follows: (27) (28) (29) (30) (31) (32) in, For the Sigmoid function, i t The input gate output is [output], and the forget gate output is [output]. o t For output gate output, This represents the candidate state at the current moment. c t The internal state at the current moment. ht The external state at the current moment. W and b These are the weight matrix and offset of the corresponding gate, respectively.

[0071] Specifically, the Gate of Oblivion f t Used to control the internal state of the previous time step. c t-1 The part that needs to be forgotten, the input gate i t Determine the candidate state at the current moment How much information is written to the new internal state, and what is the output gate? o t This is used to adjust the internal state at the current moment. How much information needs to be output to the external state? h t .

[0072] (33) Building upon this foundation, BiLSTM consists of two LSTM subnetworks, one forward and one backward, capable of simultaneously modeling the forward and backward dependencies of a sequence. Compared to unidirectional LSTM, BiLSTM can more fully uncover the implicit abstract features in time series data, leveraging bidirectional information to enhance the understanding of feature evolution patterns. This significantly improves prediction performance and effectively overcomes the lag problem caused by unidirectional models that can only capture one-sided dependencies.

[0073] 3. Fully Connected Layer: Its core function is to output preliminary RUL predictions. It mainly fuses the temporal attention mechanism in the encoder module (composed of CNN, BiLSTM, and AE layers) and the deep features extracted by BiLSTM, to output preliminary RUL predictions.

[0074] Bayesian optimization: Its core function is adaptive parameter adjustment. Addressing the issue that traditional manual parameter tuning cannot adapt to battery data under different operating conditions (such as high / normal temperature, high / low frequency charging and discharging), Bayesian optimization is introduced to achieve a globally optimal search for model parameters. It automatically adapts to the optimal parameters for different usage scenarios without manual intervention, thus improving the model's generalization ability.

[0075] Gaussian process regression is generally used as a core component of Bayesian optimization surrogate models. The formula for a Gaussian process is: (34) in, m ( x ) is the mean function, k ( x , x ') is the covariance function.

[0076] Each optimization iteration of the Bayesian algorithm updates its understanding of the objective function based on previous test results, gradually approaching the optimal parameter settings. The mathematical expression for the acquisition function is: (35) in, f ( x (This refers to information about hyperparameters) x The objective function, x * This is the optimal parameter set.

[0077] Step S3: System testing, adjustment and optimization.

[0078] Step S3 includes: Step S31: Check whether each unit can work properly. For example, check whether the intelligent acquisition unit 11 can collect data such as battery voltage, current, and temperature. The accuracy of these data needs to be compared with other tools such as multimeters and infrared thermometers.

[0079] Step S32: Check whether the entire system can output the battery's State of Health (SOH) and Remaining Life (RUL) as expected. If the output results deviate too much or have obvious errors, the algorithm parameters set in step S22 need to be adjusted.

[0080] Step S4, the Human-Machine Interface Platform 14, mainly includes three stages in the field deployment phase: system installation, configuration and debugging, and functional verification. Its core objective is to achieve a closed-loop process integrating digital twin modeling of the entire battery lifecycle, real-time monitoring and alarms, near-life (RUL) prediction, alarms, and operation and maintenance, thereby constructing a visualized decision-making platform for intelligent battery operation and maintenance. The installation of the Human-Machine Interface Platform 14 interface system is completed first in the initial stage of system deployment. This system is the core of the visualized operation of this invention, adopts a distributed architecture design, and is installed on the field operation and maintenance terminal or the control center host.

[0081] Step S4 includes: Step S41: The installation process includes the following steps: Environment initialization: Deploy the front-end visualization module and back-end interface service on the maintenance host, install the necessary runtime environment (Java runtime, Node service framework, and Web container), and complete the initial connection configuration with the database. The system installation package includes a digital twin engine module, a real-time monitoring module, a lifetime (RUL) prediction display module, and a CMDB data visualization module. The installer automatically identifies the hardware environment and completes the initialization of the graphics engine, 3D rendering library, and communication driver, as well as multi-terminal adaptation deployment. The human-computer interaction platform 14 interface supports three deployment methods: Web, PC, and mobile App. During the installation phase, the system automatically detects the screen resolution, browser type, and network bandwidth, and loads the digital twin scene with an adaptive rendering strategy to achieve a unified human-computer interaction experience on-site.

[0082] Step S42: Enter the system parameter configuration and model mapping stage. The focus of this stage is to establish the correspondence between the physical entity of the battery and the digital twin model, and to complete the real-time data stream access and intelligent algorithm binding. The system automatically generates the corresponding digital twin model in the background by importing the topology information, location information, and equipment codes of the battery packs on site. Each battery cell is mapped as an independent node in the 3D interface, containing characteristic parameters such as temperature, voltage, internal resistance, SOH, and SOC, and can be bound to real-time sensor data. After binding, the system accesses the data stream from the field acquisition device through a standardized communication protocol. In the configuration stage, the sampling period, alarm threshold, and data verification strategy are set to ensure that the real-time monitoring data and the twin model are updated synchronously. The unique identifier, operating status, installation time, manufacturer and model, and historical maintenance records of the battery are entered into the CMDB module. The system automatically establishes a battery life cycle index table to support lifespan (RUL) prediction and health trend analysis. The parameters of the health model are configured through the system management interface to configure the artificial intelligence algorithm model parameters. After the parameter configuration is completed, the system deploys the model to the local server to realize intelligent prediction of individual cells.

[0083] Step S43, System Debugging and Verification Phase: After configuration, the system enters the field debugging phase to ensure stable operation of real-time monitoring, data synchronization, and algorithm prediction. After loading field data, the system automatically generates a digital twin scene, which can display the operating status of each battery group in real time. When voltage, current, or temperature parameters are abnormal, the node status in the interface automatically changes to a warning color and triggers an alert, verifying the system's response delay and alarm accuracy.

[0084] Step S44: The Lifetime (RUL) prediction verification system calls the algorithm to perform historical trend backtracking and lifetime (RUL) prediction output on the access data, displaying the prediction curve and remaining usage days. During the debugging phase, the actual operating data and prediction results are compared, and the physical constraint correction coefficient is adjusted to improve the consistency between the prediction results and the actual degradation curve. During debugging, the CMDB data linkage is verified to be normal. The system allows users to click on any battery cell on the interface to access its historical operating records, maintenance logs, and lifetime (RUL) prediction reports, achieving bidirectional traceability from equipment to data.

[0085] Step S45: To verify the system's response capability and intelligent predictive analysis effect during the alarm and prediction phases, this invention designed a closed-loop verification process for battery alarm triggering and response in on-site deployment. This process includes five steps: data acquisition layer triggering, algorithm recognition layer analysis, visualization interface response, alarm linkage, and handling confirmation. Alarm triggering simulation simulates various alarm and prediction scenarios such as battery overvoltage, overtemperature, sudden increase in internal resistance, and sudden drop in SOC in the on-site system. The system's data acquisition module reports abnormal signals to the monitoring host in real time, forming a trigger event flow. The algorithm module then automatically analyzes the abnormal characteristics and calls upon a multi-parameter fusion early warning model to determine the risk level (e.g., Level I warning, Level II alarm, Level III severe alarm). Verification includes model recognition accuracy, response latency, level classification accuracy, and false alarm rate. In the human-computer interaction platform 14 interface, the digital twin model nodes automatically change color, flash, or jitter, and the system simultaneously pops up an early warning dialog box and a traceable information panel. The system's display consistency, interface refresh latency, and user interaction response speed are verified across different terminals (Web, maintenance host, App). The system automatically writes alarm events to the CMDB database, generating complete alarm logs, equipment status snapshots, and field parameter sets. Log backtracking verifies the correspondence between alarm information and battery operating status and prediction curves, ensuring data consistency and event traceability.

[0086] Step S46: The closed-loop verification process simulates maintenance personnel executing the entire closed-loop process of "confirming the alarm—maintenance suggestion—performing retest—closing the alarm" on the visual interface. The system automatically records the operation path and compares it with the subsequent predicted changes of the AI ​​model to verify the model's self-recovery and relearning functions after the alarm is cleared.

[0087] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0088] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0089] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0090] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

[0091] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A battery intelligent early warning and life prediction system integrating physical constraints, characterized in that, include: Intelligent acquisition unit, used to collect battery parameters; The artificial intelligence unit, connected to the intelligent acquisition unit, is used for data fusion analysis and early warning. The communication unit connects to the artificial intelligence unit for data interaction; the human-computer interaction platform connects to the communication unit for visualized operation and maintenance management.

2. The system as described in claim 1, characterized in that, The intelligent acquisition unit is configured to acquire the voltage, current, temperature and remaining power (SOC) of a single battery, and is connected to the artificial intelligence unit using a parallel bus architecture.

3. The system as described in claim 1, characterized in that, The artificial intelligence unit includes a data processing module and a grouping early warning module. The data processing module calculates the state of health (SOH) using the CNN-BiLSTM-SelfAttention algorithm, while the grouping early warning module performs battery grouping, anomaly detection, and RUL calculation.

4. The system as described in claim 3, characterized in that, The data processing module includes a Convolutional Neural Network (CNN) layer, a Bidirectional Long Short-Term Memory (BiLSTM) layer, a self-attention mechanism layer, a fully connected layer, and a physical constraint correction layer.

5. The system as described in claim 3, characterized in that, The grouping early warning module adopts a dynamic grouping equalization strategy based on impedance spectrum clustering, and uses the XGBoost algorithm to divide the battery into a healthy group, a degradation group, and a fault early warning group.

6. The system as described in claim 1, characterized in that, The communication unit supports both wired and wireless communication methods, enabling local data interaction and remote platform data uploading.

7. A method for intelligent early warning and life prediction of a battery based on the system described in any one of claims 1-6, incorporating physical constraints, characterized in that, Includes the following steps: S1. The intelligent acquisition unit collects the voltage, current, temperature and SOC data of the battery in real time, and performs data preprocessing to remove outliers. S2. Use artificial intelligence algorithms to process data, including predicting the health status SOH based on the CNN-BiLSTM-SelfAttention model and predicting the remaining useful life RUL based on the AE-CNN-BiLSTM model, while combining a physical constraint correction layer to optimize the prediction accuracy. S3. Perform anomaly detection and fault warning. Use a dynamic grouping equalization strategy based on impedance spectrum clustering to classify batteries and use a five-dimensional feature anomaly detection algorithm to identify potential faults. S4. Achieve result visualization, digital twin modeling, and closed-loop operation and maintenance management through a human-computer interaction platform.

8. The method based on claim 7, characterized in that, Step S2, which uses the CNN-BiLSTM-SelfAttention model to predict SOH, includes the following sub-steps: S21. Perform outlier removal and standardization on the collected voltage, current, temperature and time series data to generate a sample set; S22. Use convolution kernels to extract local spatial features and reduce dimensionality, and reduce input redundancy through convolution operations; S23. Capture the long-term dependencies of the charging and discharging process through input gate, forget gate and output gate mechanisms; S24. Calculate the correlation between each part of the sequence and the SOH state, and dynamically allocate attention weights to focus on key features; S25. After fusing the deep features, output the preliminary SOH prediction value; S26. Calculate the residuals based on the physical equation of battery capacity decay, and construct a comprehensive loss function to correct the predicted values.

9. The method based on claim 7, characterized in that, The anomaly detection and fault warning in step S3 includes the following sub-steps: S31. Collect battery AC impedance spectrum data, extract impedance amplitude and phase angle features, use principal component analysis algorithm for dimensionality reduction, and then use extreme gradient boosting algorithm for clustering to divide the battery into healthy group, degraded group and fault warning group. S32. A five-dimensional feature space is constructed based on voltage, internal resistance, temperature, SOC, and SOH to comprehensively characterize the static and dynamic operating characteristics of the battery; S33. Using a single-class support vector machine algorithm, a radial basis kernel function is introduced to construct the minimum hypersphere that encloses normal data in five-dimensional space; S34. When the real-time sample features exceed the boundary of the hypersphere, it is judged as an anomaly and an early warning is triggered.

10. The method based on claim 7, characterized in that, Step S2, which uses the AE-CNN-BiLSTM model to predict RUL, includes the following sub-steps: S41. Local features of time series are extracted through convolutional neural network layers, bidirectional long short-term memory network layers are used to capture forward and backward time dependencies, and autoencoders are used for feature compression and noise reduction. S42. A time attention mechanism is introduced to assign weights to different time steps, highlighting key moments of influence. Then, a bidirectional long short-term memory network layer is used to mine long-term trends, and a fully connected layer outputs the RUL prediction value. S43. Using Gaussian process regression as a surrogate model, the optimal parameters are searched globally through the acquisition function to adaptively optimize the model performance.

Images