Storage battery on-line state monitoring method

Abstract

The invention discloses a storage battery on-line state monitoring method, which comprises the following steps of S1, acquiring voltage, current, temperature, internal resistance, shell micro strain and electrolyte ion migration rate parameters of a storage battery in real time through a sensor array to form multi-mode original data; and S2, performing filtering noise reduction and feature extraction on the multi-modal original data acquired in the S1 by adopting data preprocessing equipment to obtain key feature values, and transmitting the key feature values to a cloud platform. According to the online state monitoring method for the storage battery, multi-mode parameters such as voltage, current, shell micro-strain and electrolyte ion migration rate are collected through the sensor array, the limitation of traditional single electrical parameter monitoring is broken through, and comprehensive characterization from macroscopic electrical characteristics to microcosmic physical and chemical states is achieved. And in combination with multi-modal feature fusion of the deep belief network, the external performance and internal degradation mechanism of the storage battery can be effectively associated, and the accuracy of state evaluation is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of battery monitoring technology, specifically to a method for online monitoring of battery status. Background Technology

[0002] As a core component of energy storage systems, the operating status of batteries directly affects the reliability and safety of the entire system, and they are widely used in communication base stations, data centers, new energy vehicles, and other fields. Currently, online battery status monitoring mainly relies on the collection of conventional electrical parameters such as voltage, current, and temperature, and assesses health status through threshold judgments or simple algorithms. However, these methods only reflect the macroscopic operating appearance of the battery and are difficult to capture microscopic changes such as internal chemical reactions, plate aging, and electrolyte deterioration. This leads to delayed fault warnings and a high false alarm rate, especially in the early stages of battery degradation, where potential risks cannot be effectively identified.

[0003] In existing technologies, some monitoring solutions attempt to introduce internal resistance detection or single physical characteristics such as casing temperature distribution to assist in assessment, but these solutions have significant limitations: First, the sensor types are limited, failing to achieve collaborative monitoring of multi-dimensional states and making it difficult to establish correlations between parameters; second, data processing largely relies on local fixed algorithms, lacking adaptability to individual differences in batteries and dynamic environmental changes, resulting in poor model generalization; and third, the collaboration mechanism between edge nodes and the cloud is imperfect, making it difficult to balance real-time performance and data integrity. Therefore, there is an urgent need for a monitoring method that integrates multimodal parameters and possesses adaptive learning capabilities to improve the accuracy of battery state assessment and the timeliness of early warnings. Summary of the Invention

[0004] The purpose of this invention is to provide an online status monitoring method for batteries to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for online status monitoring of a storage battery, characterized by comprising the following steps:

[0006] S1. Real-time acquisition of battery voltage, current, temperature, internal resistance, shell micro-strain, and electrolyte ion migration rate parameters through sensor array to form multimodal raw data;

[0007] S2. The multimodal raw data collected in S1 is filtered, denoised, and feature extracted using data preprocessing equipment. After obtaining key feature values, the data is transmitted to the cloud platform.

[0008] S3. Based on a deep belief network, multimodal fusion is performed on the key feature values ​​transmitted in S2. The fusion model is adaptively updated through reinforcement learning, and the battery health assessment result is output.

[0009] S4. Using the health assessment results output by S3, the multimodal raw data of S1, and the data obtained by filtering and noise reduction and feature extraction in S2, perform dynamic calculation in the time dimension to obtain the feature change rate, trigger the three-level early warning mechanism and locate the fault source, and generate targeted maintenance suggestions.

[0010] Preferably, the sensor array in S1 includes:

[0011] The tunnel magnetoresistive current sensor has a non-contact structure.

[0012] A fiber optic strain sensor is attached to the surface of the battery casing, and a temperature compensation fiber optic strain sensor is configured to eliminate temperature interference.

[0013] An ion-selective electrode array, comprising a metal ion electrode, a hydrogen ion electrode, a sulfate ion electrode, and a reference electrode, is used to calculate the ion migration rate of the electrolyte based on electrochemical principles. This includes obtaining the conductivity of the electrolyte, the concentration and valence state of the target ion, and calculating the ion migration rate based on the correlation between conductivity, ion concentration, and ion valence state.

[0014] Preferably, the installation process of the fiber optic strain sensor includes: using a corrosion-resistant adhesive to bond it along the axial direction of the battery casing, and pre-treating the surface of the casing before bonding to ensure strain transfer efficiency.

[0015] Preferably, the electrodes of the ion-selective electrode array have a miniaturized structure, are inserted into the electrolyte at a preset depth, are isolated from electrolyte corrosion by a corrosion-resistant shell, and the electrode potential is regulated by a potential control device.

[0016] Preferably, the preprocessing of the data in step S2 includes:

[0017] Kalman filtering is used to eliminate noise. The state vector is the voltage, the current, the temperature, the internal resistance, the strain, and the migration rate. The process noise covariance and the measurement noise covariance are set based on the parameter characteristics.

[0018] Feature extraction includes: calculating the peak-to-valley difference, root mean square, and kurtosis of the micro-strain of the shell within a preset sliding window; and extracting the energy percentage of the electrolyte ion migration rate in a preset frequency band using short-time Fourier transform.

[0019] Preferably, the data processing device in step S2 adopts a hardware platform with real-time processing capabilities, realizes time synchronization between data acquisition and preprocessing based on a real-time operating system, supports multiple sensor accesses and has a data caching function.

[0020] Preferably, the structure of the deep belief network in S3 is as follows:

[0021] The input layer is the multimodal feature vector, which includes the high-frequency ripple coefficient and dynamic response slope characteristics of the current;

[0022] The hidden layer consists of at least two layers of neural networks, using ReLU and LeakyReLU activation functions respectively;

[0023] The output layer is an abstract state vector. This abstract state vector is generated by the output layer after the hidden layer of the deep belief network performs hierarchical feature mapping and nonlinear transformation on the input multimodal feature vector. The output normalized result is generated by the Sigmoid activation function.

[0024] Preferably, the reinforcement learning in S3 employs the Q-learning algorithm, obtaining parameters that satisfy:

[0025] The state space includes the battery health status, the ambient temperature, and the charge / discharge rate;

[0026] The action space is for adjusting the weights of the output layer of the deep belief network and correcting the health threshold.

[0027] The reward function is calculated based on the error between the predicted remaining lifetime and the actual remaining lifetime, and the learning rate and discount factor are set to optimize the model update.

[0028] Preferably, the threshold of the three-level early warning mechanism in S4 satisfies:

[0029] Level 1 warning: The parameter fluctuation exceeds a preset multiple of the statistical standard deviation of a preset number of normal batteries, where the preset number is the number of battery samples of the same model and batch as the monitored object and in a healthy state;

[0030] Level 2 warning: The battery health score is lower than a preset threshold, which is determined by receiver operation characteristic curve analysis. The number of preset thresholds is set according to the warning level and parameter characteristics.

[0031] Level 3 warning: A sudden increase in internal resistance or an abnormal peak value in the micro-strain of the shell.

[0032] Preferably, the fault source localization in S4 adopts a gradient boosting model, and the input feature change rate includes the temperature rise rate, the internal resistance change, and the current dynamic fluctuation coefficient, and the fault type identification is achieved based on the feature importance ranking.

[0033] Compared with the prior art, the beneficial effects of the present invention are:

[0034] This online battery condition monitoring method uses a sensor array to collect multimodal parameters such as voltage, current, casing micro-strain, and electrolyte ion migration rate, overcoming the limitations of traditional single electrical parameter monitoring and achieving comprehensive characterization from macroscopic electrical properties to microscopic physicochemical state. By combining multimodal feature fusion with deep belief networks, it can effectively correlate external battery performance with internal degradation mechanisms, significantly improving the accuracy of condition assessment.

[0035] This online battery condition monitoring method introduces reinforcement learning to dynamically update the evaluation model. It can automatically optimize parameters based on the individual differences of batteries of different models and aging stages, solving the problem of insufficient adaptability caused by traditional algorithms. Whether it is a newly installed battery or a battery nearing the end of its life, it can maintain stable evaluation accuracy and can be widely used in battery pack monitoring in various scenarios such as communication base stations, data centers, and energy storage power stations.

[0036] This online battery status monitoring method, based on a three-level early warning mechanism constructed from the change rates of multiple parameters, can identify potential risks 3-6 months before a visible battery failure occurs, and accurately locate the fault source through a gradient boosting model. Combined with targeted maintenance recommendations, it can avoid the waste of resources caused by blindly replacing batteries, while reducing downtime losses due to sudden failures.

[0037] This online battery status monitoring method employs localized data preprocessing equipment, combined with global optimization on a cloud platform, to balance real-time performance and overall decision-making capabilities. The sensor array features a modular design, adaptable to batteries of different specifications, and installation does not compromise battery sealing, resulting in low deployment costs, strong compatibility, and ease of upgrading existing battery packs. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] Please see Figure 1 This invention provides a technical solution: a method for online status monitoring of a storage battery, comprising the following steps:

[0041] S1. Real-time acquisition of battery voltage, current, temperature, internal resistance, casing micro-strain, and electrolyte ion migration rate parameters via a sensor array to generate multimodal raw data. The sensor array in S1 includes:

[0042] The tunnel magnetoresistive current sensor is a non-contact structure with high sensitivity and wide dynamic range. Based on the magnetoresistive effect, which refers to the physical phenomenon that the resistance of a material changes significantly with the application of an external magnetic field, the sensor measures current. Its sensing element maintains a preset distance from the battery output cable to avoid mechanical contact. The shielded shell, made of highly conductive material, reduces external electromagnetic interference by blocking external electromagnetic fields from entering or leaking into the internal electromagnetic field. The sensor output integrates a signal conditioning circuit, which is an electronic circuit that amplifies, filters, and linearizes the original electrical signal to optimize the signal-to-noise ratio (SNR), which is the ratio of signal power to noise power. A higher SNR indicates less noise interference to the signal.

[0043] A fiber optic strain sensor is attached to the surface of a battery casing. A temperature-compensating fiber optic strain sensor is configured to eliminate temperature interference. The installation process of the fiber optic strain sensor includes: using corrosion-resistant adhesive to attach it along the axial direction of the battery casing; pre-treating the casing surface before attachment to ensure strain transfer efficiency; the fiber optic cable of the fiber optic strain sensor is made of bend-resistant quartz material; the grating region has a uniformly distributed periodic refractive index modulation structure, which refers to a region within the fiber formed by laser etching, where the refractive index changes periodically along the axial direction, and can produce changes in reflection or transmission characteristics for light of a specific wavelength; the temperature-compensating fiber optic strain sensor and the measurement fiber optic strain sensor are arranged parallel to each other and... Both are located in the same temperature field and connected to the demodulation device via the same optical fiber link. This device can convert the wavelength changes of the reflected / transmitted light of the fiber optic grating into electrical signals to achieve synchronous signal acquisition. The corrosion-resistant adhesive is a polymeric adhesive with acid resistance, alkali resistance, and high and low temperature resistance. It is an adhesive material mainly composed of polymeric compounds, which has good chemical stability and temperature adaptability. After curing, it forms an elastic modulus. The ratio of stress to strain in the elastic deformation stage of the material reflects the material's ability to resist deformation and is matched with the shell material. The shell surface pretreatment includes degreasing and cleaning, roughening treatment, and drying treatment in sequence to remove surface oil and oxide layer and improve the adhesion strength of the adhesive.

[0044] An ion-selective electrode array includes metal ion electrodes, hydrogen ion electrodes, sulfate ion electrodes, and a reference electrode. Based on electrochemical principles, it calculates the ion migration rate in the electrolyte. The electrodes of the ion-selective electrode array employ a miniaturized structure, inserted into the electrolyte at a predetermined depth, and isolated from electrolyte corrosion by a corrosion-resistant shell. A potential control device regulates the electrode potential. The sensitive membranes of the metal ion electrodes, hydrogen ion electrodes, and sulfate ion electrodes are thin films covering the electrode surface, selectively adsorbing or responding to specific ions. Their electrical signals change with the target ion concentration and are made of materials selective to the corresponding ions to ensure specific response to the target ion. The reference electrode is a solid reference electrode with a stable potential. An electrode whose potential remains constant during measurement serves as a reference for measuring the potential of other electrodes, and its potential is unaffected by changes in electrolyte composition. The corrosion-resistant outer shell is made of a chemically inert material that is not easily chemically reacted in a specific chemical environment and is stable in nature, with only the sensitive end of the electrode exposed in the electrolyte. The potential control device is an electronic device that can precisely adjust and stabilize the potential difference between the electrodes. It can adjust the potential difference between each working electrode and the reference electrode in real time to suppress the interference of coexisting ions and maintain the stability of the electrode reaction. The preset depth of electrode insertion into the electrolyte is determined according to the convection characteristics and reaction area distribution of the electrolyte inside the battery to ensure that the monitoring signal can reflect the overall state of the electrolyte.

[0045] S2. A data preprocessing device is deployed near the data acquisition source, enabling local data processing to reduce transmission latency and bandwidth consumption. This process filters and reduces noise in the multimodal raw data acquired in S1, including unprocessed raw signal data from various types of sensors. It involves using algorithms to reduce or eliminate useless interference components in the signals, and extracting features from the raw data to obtain key feature values. These feature parameter values, which play a decisive role in the state of the analyzed object, are then transmitted to a cloud platform. The data is centrally stored and processed via a remote server system. The data processing device integrates a data receiving module (hardware or software components for receiving external data), a data processing unit (core components for performing data calculations and processing), and a communication module (hardware or software components for data transmission). The data receiving module receives the multimodal raw data output from the sensor array. The data processing unit executes filtering, noise reduction, and feature extraction algorithms, processing the data according to specific rules. The communication module establishes a wireless or wired connection with the cloud platform to achieve targeted transmission of key feature values.

[0046] Kalman filtering, an algorithm that uses the state equations of a linear system to optimally estimate the system state based on system input and output observation data, is employed to eliminate noise. The state vector consists of multiple state variables that need to be estimated, such as voltage, current, temperature, internal resistance, strain, and migration rate. Based on parameter characteristics, process noise covariance is set, describing the covariance matrix of system model uncertainty, and measurement noise covariance is set, describing the covariance matrix of noise characteristics during measurement. The Kalman filter achieves noise elimination through a prediction step, which infers the current state based on the previous state and the system dynamic model, and an update step, which corrects the prediction result based on the current measurement value. The prediction step, based on the previous state estimate, generates the predicted state value for the current moment using the system dynamic model and a mathematical model describing the change of the system state over time. The update step corrects the predicted value based on the current measurement value to obtain the optimal state estimate, the state prediction result with the smallest statistical error. The voltage, current and other parameters in the state vector correspond to the electrical characteristics of the battery, reflecting the battery's electrical performance attributes, physical deformation, shape changes and chemical activity state caused by internal or external forces, reflecting the degree of activity of internal chemical reactions. The process noise covariance is set according to the natural fluctuation characteristics of each parameter when there is no external interference, and the measurement noise covariance is set according to the inherent noise level of the sensor, the noise level generated by the sensor itself and the characteristics of environmental interference.

[0047] Feature extraction includes: calculating the peak-to-trough difference of the shell micro-strain within a preset sliding window and a fixed-length time interval set for continuous data analysis; the difference between the maximum and minimum values ​​of the signal within a certain interval; the root mean square; the square root of the average of the squared values ​​of the signal; statistical quantities describing the energy magnitude and kurtosis of the signal; and using short-time Fourier transform to divide the signal into multiple short-time windows, performing Fourier transform on each window to obtain time-frequency information to extract the energy proportion of electrolyte ion migration rate within a preset frequency band and a specific frequency range set according to the analysis requirements. The preset sliding window is a continuous time interval, the length of which is determined according to the time scale of the battery state change and the time order describing the speed of the state change. The window slides along the time axis with a preset step size and the interval of the window movement along the time axis to achieve continuous feature extraction. The peak-valley difference reflects the maximum fluctuation amplitude of the shell micro-strain, the root mean square reflects the energy level of the strain signal, and the kurtosis reflects the impact characteristics of the strain signal and the strength of the abrupt component in the signal. The short-time Fourier transform is a mathematical transformation that converts the time domain signal into a frequency domain signal by dividing the ion migration rate signal into overlapping time windows and performing a Fourier transform on each window to obtain the time-frequency distribution, which describes the graph or data of the energy distribution of the signal at different times and frequencies. The preset frequency band is determined according to the main frequency components of electrolyte ion migration, and the energy ratio is the ratio of the signal energy in that frequency band to the total energy.

[0048] The data processing equipment adopts a hardware platform with real-time processing capabilities. It is a combination of physical devices that can complete data processing tasks within a specified time. Based on a real-time operating system, it can ensure that tasks are responded to and completed within a certain time. It realizes the time synchronization of the data acquisition and preprocessing process and the initial processing of raw data, so that the time base of different tasks is consistent. It supports the access of multiple sensors, allows multiple sensors to connect and transmit data at the same time, and has a data caching function to temporarily store data to cope with transmission or processing delays. The hardware platform includes a high-performance processor, a powerful central processing unit, a real-time clock chip, electronic components providing high-precision time information, and multi-channel interfaces that can connect to multiple external devices simultaneously. The processor is responsible for running filtering and feature extraction algorithms, while the real-time clock chip provides a high-precision time reference standard for unified time measurement. The real-time operating system uses a priority scheduling mechanism to allocate system resources based on task importance, ensuring that the time deviation between data acquisition and preprocessing tasks is controlled within a preset range. Multi-sensor access is achieved through standardized interfaces that adhere to unified technical specifications, supporting plug-and-play connections for different types of sensors without manual configuration. Data caching is implemented using non-volatile storage media, such as solid-state drives and flash memory, to temporarily store raw data and processing results, preventing data loss during communication interruptions and allowing for retransmission after communication is restored. This ensures that the monitoring signals reflect the overall state of the electrolyte.

[0049] S3. Based on a deep belief network, multimodal fusion of key feature values ​​transmitted in S2 is performed, integrating feature data of different types and sources (such as electrical, physical, and chemical features) into a unified representation. A deep learning model composed of multiple layers of restricted Boltzmann mechanisms is used to learn features through layer-by-layer pre-training and fine-tuning. The fusion model is adaptively updated through reinforcement learning. Reinforcement learning is a machine learning method that optimizes behavioral strategies based on reward signals to achieve goals through interaction between an agent and the environment. The fusion model is an algorithmic model used to integrate multimodal data; the output is the battery health assessment result. The multimodal fusion is achieved through hierarchical feature mapping of a deep belief network. Feature mapping is a process of gradually transforming the original features into more abstract and representative features through a multi-layer network structure. This transforms heterogeneous features such as voltage, current, and shell micro-strain into a unified feature space. These heterogeneous features have different data types or physical meanings, and the feature space can accommodate the common representation dimension of multiple features. It eliminates the dimensional differences between different parameters and the differences in units or numerical scales of different physical quantities. The adaptive update triggering conditions of the fusion model include the feature distribution drift threshold, the cumulative evaluation error value, and a preset time period. The drift threshold is the phenomenon that the statistical distribution of feature data changes over time, and the cumulative evaluation error value is the sum of the errors from multiple health assessments, ensuring that the model can dynamically adapt to the aging process of the battery and environmental changes.

[0050] The structure of the deep belief network is as follows: The input layer is a multimodal feature vector. The first layer of neurons in the input layer neural network receives the original data and contains a set of vectors with multiple types of features, including the high-frequency ripple coefficient of the current and the dynamic response slope feature. The high-frequency ripple coefficient reflects the fluctuation intensity of the high-frequency component in the current signal, and the dynamic response slope characterizes the rate of change of the current at the moment of switching between charging and discharging. The hidden layer includes at least two layers of neural networks, which use ReLU and LeakyReLU activation functions respectively. The ReLU activation function enhances the nonlinear expression ability of features by suppressing negative input values, and the LeakyReLU activation function avoids neuron inactivation by preserving small negative input gradients. The output layer is an abstract state vector, which uses the Sigmoid activation function to output the normalized result. Each dimension of the abstract state vector corresponds to the core health indicators such as the capacity decay degree of the battery, the internal resistance growth trend, and the shell stability. The normalized result is mapped to the [0,1] interval to quantify the health level.

[0051] The reinforcement learning algorithm employs Q-learning, with parameters satisfying the following: the state space includes battery health, ambient temperature, and charge / discharge rate; the dimensionality of the state space is determined based on the battery's operating characteristics, with health divided into continuous intervals, and ambient temperature and charge / discharge rate also defined; the state space is discretized according to actual operating conditions; the action space involves adjusting the weights of the deep belief network output layer and correcting the health threshold; weight adjustment is achieved through gradient descent, and threshold correction is based on the statistical characteristics of historical evaluation errors; the reward function is calculated based on the error between the predicted remaining lifespan and the actual remaining lifespan; a learning rate and a discount factor are set to optimize model updates; the reward function value increases as the error decreases; the learning rate controls the magnitude of each parameter update; and the discount factor balances the weights of immediate rewards and long-term cumulative rewards, achieving a balance between short-term evaluation accuracy and long-term stability.

[0052] The health assessment results output in S4 and S3 are the multimodal raw data from S1 and the feature change rate of the data obtained from filtering, noise reduction, and feature extraction in S2. The health assessment results are a quantitative evaluation of the overall state of the battery, reflecting its performance degradation and failure risk. The change range of parameters within a unit of time is used to characterize the dynamic trend of the state, trigger a three-level early warning mechanism and locate the fault source. The multi-level alarm system is divided according to the severity of the fault, with different levels corresponding to different response strategies. The specific components or causes leading to battery abnormalities are identified, and targeted maintenance suggestions are generated. Repair plans are formulated based on fault characteristics and actual needs. The health assessment results and feature change rates form a dual judgment basis. The health assessment results reflect the overall state of the battery, and the feature change rate reflects the dynamic trend of the parameters. The two trigger early warnings through logical association. Logical association refers to a conditional judgment relationship based on preset rules: when the change rate of a single parameter exceeds the normal range but the health is still within the safe range, only a low-level early warning is triggered, which is the lowest severity level in the early warning system, usually corresponding to a minor abnormality; when the health and the change rate of multiple parameters are abnormal at the same time, the early warning level is upgraded, increasing the severity level of the early warning, corresponding to more urgent handling needs. Targeted maintenance recommendations are based on fault type and categorized by fault nature, such as electrical faults, structural faults, etc., battery operating scenarios, and historical maintenance records. Battery operating scenarios refer to the application environment and working conditions of the battery, such as communication base stations and data centers. Historical maintenance records refer to past repair and maintenance records for the battery, including fault cause analysis, inferences and explanations of the root causes of the fault, operating steps, and resource allocation plans, which are the allocation plans for personnel, equipment, materials, etc. required to complete the maintenance.

[0053] The threshold for the Level 3 early warning mechanism is satisfied as follows:

[0054] Level 1 Warning: Parameter fluctuations exceed a preset multiple of the statistical standard deviation of a preset number of normal batteries. Standard deviation is a statistical measure used to describe the dispersion of data, reflecting the fluctuation range of normal battery parameters. The preset number of normal batteries refers to a sample group of the same model and batch as the monitored object, and in a healthy state. The statistical standard deviation is obtained through statistical analysis of their historical operating parameters. The preset multiple is set according to the sensitivity of the parameter to the battery state. Sensitivity refers to the significant impact of parameter changes on battery performance. Sensitive parameters (such as internal resistance and ion migration rate) use a lower multiple, while non-sensitive parameters (such as ambient temperature) use a higher multiple to balance warning sensitivity and false alarm rate. Warning sensitivity is the ability of the warning system to identify anomalies, while the false alarm rate is the probability of misjudging a normal state as an anomaly.

[0055] Level 2 Warning: The battery health score is below a preset threshold, which is the critical value for determining whether the health is abnormal. This threshold is determined through receiver operating characteristic curve analysis. The receiver operating characteristic curve is used to evaluate the classification performance of the model. The horizontal axis represents the false positive rate, and the vertical axis represents the true positive rate. By plotting the probability of correctly identifying a fault by fault identification sensitivity under different thresholds, and the probability of correctly identifying a normal state by specificity, the preset threshold is the optimal balance point on the curve that balances sensitivity and specificity, ensuring that while effectively identifying potential faults, the probability of false warnings is controlled, and the probability of misjudging a normal state as requiring a warning is minimized.

[0056] Level 3 Warning: A sudden increase in internal resistance, characterized by a rapid increase in internal resistance within a short period, or an abnormal peak value in the casing's micro-strain, where the casing strain value suddenly rises far beyond the normal range. The sudden increase in internal resistance refers to a change in internal resistance per unit time exceeding a preset proportion of the normal aging trend. The normal aging trend refers to the natural degradation of the battery over time, and this duration reaches a set threshold. The abnormal peak value in the casing's micro-strain refers to a strain value exceeding the maximum fluctuation range within a normal charge-discharge cycle, and the peak duration is shorter than the recovery period of normal mechanical deformation. Normal mechanical deformation refers to the shape change of an object under external force, indicating a possible sudden structural abnormality, or sudden damage to the casing or internal structure due to unexpected factors.

[0057] The fault source localization employs a gradient boosting model, an ensemble learning algorithm. This model iteratively trains multiple weak classifiers and weights them to form a strong classifier. Input feature change rates include temperature rise rate, internal resistance change, and current dynamic fluctuation coefficient. Fault type identification is achieved based on feature importance ranking, which is based on the contribution of features to the model's decision. The gradient boosting model, formed by iteratively ensembled multiple weak classifiers, significantly outperforms complex models with a single weak classifier. The weak classifiers slightly outperform simple classification models relying on random guessing. Each weak classifier optimizes for the error of the previous model. The input feature change rate is the parameter change amplitude per unit time, calculated using... A sliding time window is used to eliminate instantaneous noise interference. The time window is a fixed-length data interval that moves along the time axis to smooth out instantaneous fluctuations. The importance ranking of features is determined based on the contribution of each feature during model training. The contribution of a feature in reducing model error is determined by the degree of its contribution. The feature with the higher the contribution, the greater its weight in fault location. Finally, based on the feature combination pattern, the change pattern presented by multiple features, and the matching degree with the preset fault feature library, the specific fault type is output, such as plate sulfation, the phenomenon of lead sulfate crystals that are difficult to participate in the reaction forming on the surface of the battery plate, electrolyte leakage, the fault of electrolyte overflowing from the shell gaps, and shell deformation, the shape change of the battery shell due to internal pressure or external force, etc.

[0058] By employing a multimodal sensor array, including tunneling magnetoresistive current sensors, fiber optic strain sensors, and ion-selective electrode arrays, the electrical, physical, and chemical parameters of the battery are comprehensively collected, addressing the problem of incomplete state characterization caused by traditional monitoring relying solely on single parameters such as voltage and current. The collected raw data undergoes real-time filtering and noise reduction, Kalman filtering, feature extraction, sliding window statistics, and short-time Fourier transform. This local processing reduces transmission latency and bandwidth consumption while precise noise reduction and deep feature mining address signal noise interference and insufficient extraction of effective information. Finally, a deep belief network is utilized. By fusing multimodal features and combining reinforcement learning to achieve adaptive model updates, the differences in the dimensions of heterogeneous parameters are eliminated and the battery aging process is dynamically adapted, overcoming the shortcomings of traditional models such as weak generalization ability and difficulty in adapting to long-term changes. Finally, a three-level early warning mechanism is triggered based on the health assessment results and feature change rate. The model is located by gradient boosting, forming a closed loop from parameter acquisition, processing, analysis to early warning and maintenance. This solves the problems of fixed early warning thresholds, ambiguous fault location, and delayed response in traditional models, achieving comprehensive perception, accurate assessment, and intelligent early warning of battery status, and significantly improving the real-time performance, accuracy, and adaptability of monitoring.

[0059] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for online status monitoring of a storage battery, characterized in that, Includes the following steps: S1. Real-time acquisition of battery voltage, current, temperature, internal resistance, shell micro-strain, and electrolyte ion migration rate parameters through sensor array to form multimodal raw data; S2. The multimodal raw data collected in S1 is filtered, denoised, and feature extracted using data preprocessing equipment. After obtaining key feature values, the data is transmitted to the cloud platform. S3. Based on a deep belief network, multimodal fusion is performed on the key feature values ​​transmitted in S2. The fusion model is adaptively updated through reinforcement learning, and the battery health assessment result is output. S4. Using the health assessment results output by S3, the multimodal raw data of S1, and the data obtained by filtering and noise reduction and feature extraction in S2, perform dynamic calculation in the time dimension to obtain the feature change rate, trigger the three-level early warning mechanism and locate the fault source, and generate targeted maintenance suggestions.

2. The method for online status monitoring of a storage battery according to claim 1, characterized in that: The sensor array in S1 includes: The tunnel magnetoresistive current sensor has a non-contact structure. A fiber optic strain sensor is attached to the surface of the battery casing and equipped with a temperature-compensated fiber optic strain sensor to eliminate temperature interference. An ion-selective electrode array, comprising a metal ion electrode, a hydrogen ion electrode, a sulfate ion electrode, and a reference electrode, is used to calculate the ion migration rate of the electrolyte based on electrochemical principles. This includes obtaining the conductivity of the electrolyte, the concentration and valence state of the target ion, and calculating the ion migration rate based on the correlation between conductivity, ion concentration, and ion valence state.

3. The method for online status monitoring of a storage battery according to claim 2, characterized in that: The installation process of the fiber optic strain sensor includes: using corrosion-resistant adhesive to attach it along the axial direction of the battery casing, and pre-treating the surface of the casing before attachment to ensure strain transfer efficiency.

4. The method for online status monitoring of a storage battery according to claim 2, characterized in that: The electrodes of the ion-selective electrode array are miniaturized and inserted into the electrolyte at a predetermined depth. The electrolyte corrosion is isolated by a corrosion-resistant shell, and the electrode potential is regulated by a potential control device.

5. The method for online status monitoring of a storage battery according to claim 1, characterized in that: The preprocessing of the data in step S2 includes: Kalman filtering is used to eliminate noise. The state vector is the voltage, the current, the temperature, the internal resistance, the strain, and the migration rate. The process noise covariance and the measurement noise covariance are set based on the parameter characteristics. Feature extraction includes: calculating the peak-to-valley difference, root mean square, and kurtosis of the micro-strain of the shell within a preset sliding window; and extracting the energy percentage of the electrolyte ion migration rate in a preset frequency band using short-time Fourier transform.

6. The method for online status monitoring of a storage battery according to claim 5, characterized in that: The data processing device in step S2 adopts a hardware platform with real-time processing capabilities, realizes the time synchronization of data acquisition and preprocessing based on a real-time operating system, supports the access of multiple sensors and has a data caching function.

7. The method for online status monitoring of a storage battery according to claim 1, characterized in that: The structure of the deep belief network in S3 is as follows: The input layer is the multimodal feature vector, which includes the high-frequency ripple coefficient and dynamic response slope characteristics of the current; The hidden layer consists of at least two layers of neural networks, using ReLU and LeakyReLU activation functions respectively; The output layer is an abstract state vector. This abstract state vector is generated by the output layer after the hidden layer of the deep belief network performs hierarchical feature mapping and nonlinear transformation on the input multimodal feature vector. The output normalized result is generated by the Sigmoid activation function.

8. The method for online status monitoring of a storage battery according to claim 7, characterized in that: The reinforcement learning described in S3 employs the Q-learning algorithm, obtaining parameters that satisfy: The state space includes the battery health status, the ambient temperature, and the charge / discharge rate; The action space is for adjusting the weights of the output layer of the deep belief network and correcting the health threshold. The reward function is calculated based on the error between the predicted remaining lifetime and the actual remaining lifetime, and the learning rate and discount factor are set to optimize the model update.

9. The method for online status monitoring of a storage battery according to claim 1, characterized in that: The threshold values ​​of the three-level early warning mechanism in S4 satisfy the following: Level 1 warning: The fluctuation of the parameter exceeds a preset multiple of the statistical standard deviation of a preset number of normal batteries, where the preset number is the number of battery samples of the same model and batch as the monitored object and in a healthy state; Level 2 warning: The battery health score is lower than a preset threshold, which is determined by receiver operation characteristic curve analysis. The number of preset thresholds is set according to the warning level and parameter characteristics. Level 3 warning: A sudden increase in internal resistance or an abnormal peak value in the micro-strain of the shell.

10. A method for online status monitoring of a storage battery according to claim 9, characterized in that: The fault source localization in S4 adopts a gradient boosting model. The input feature change rate includes the temperature rise rate, the internal resistance change, and the current dynamic fluctuation coefficient. Fault type identification is achieved based on feature importance ranking.

Images