Low-temperature-resistant lithium battery bms monitoring method and system based on cooperative sensing

Abstract

The present application relates to the field of lithium battery intelligent monitoring, especially to a low-temperature-resistant lithium battery BMS monitoring method and system based on collaborative sensing; the present application obtains the surface temperature distribution, environmental temperature and electrochemical parameters of low-temperature-resistant lithium battery, constructs a multi-source state feature matrix through space-time alignment and data fusion; applies a composite current excitation signal, extracts impedance parameters, nonlinear distortion components and electrochemical stress factors through impedance analysis and spectrum analysis, and forms a multi-dimensional feature vector; obtains the state monitoring result by using decision-level fusion; through multi-sensing fusion and multi-dimensional feature fusion, the special internal state of low-temperature-resistant lithium battery is fully characterized, early risk identification and early warning are carried out according to its unique electrochemical behavior characteristics, compared with the ordinary lithium battery monitoring method, the accuracy, robustness and real-time performance of the state monitoring of low-temperature-resistant lithium battery in low-temperature environment are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring of lithium batteries, and in particular to a method and system for monitoring low-temperature lithium battery BMS based on collaborative sensing. Background Technology

[0002] Lithium-ion batteries, as high-energy-density energy storage devices, are widely used in electric vehicles, energy storage systems, aerospace, and other fields. However, lithium-ion batteries face challenges such as capacity decay, power reduction, and charging difficulties in low-temperature environments, affecting their performance in cold regions and extreme environments.

[0003] Low-temperature resistant lithium batteries are special lithium batteries developed for extreme low-temperature environments. Their key technical features lie in the use of modified electrolyte systems and surface-modified anode materials. The modified electrolyte, through the addition of low-temperature additives and optimized solvent formulation, significantly improves ion conductivity at low temperatures. The modified anode material, through surface coating and doping techniques, improves lithium-ion insertion / extraction kinetics at low temperatures. Due to fundamental differences in material systems, low-temperature resistant lithium batteries differ significantly from ordinary lithium batteries in electrochemical impedance spectroscopy characteristics, polarization behavior patterns, and aging degradation mechanisms.

[0004] As the core guarantee for the safe operation of lithium batteries, the Battery Management System (BMS) needs to monitor battery status parameters in real time, including voltage, current, temperature, state of charge, and state of health. Traditional BMS monitoring methods mainly rely on the measurement of basic parameters such as voltage, current, and temperature, and perform state estimation through empirical formulas or simple models. However, this method has insufficient accuracy and reliability in low-temperature environments, especially in terms of its limited ability to monitor the internal state of the battery at a deeper level.

[0005] In existing technologies, electrochemical impedance spectroscopy (EIS) is considered an effective means of characterizing the internal electrochemical processes of lithium-ion batteries, reflecting charge transfer, diffusion processes, and interface characteristics. However, traditional EIS testing typically uses a single-frequency sinusoidal excitation signal, resulting in long testing times that are difficult to meet the needs of real-time monitoring. Furthermore, existing multi-sensor fusion technologies mainly focus on simple data fusion at the data level, lacking in-depth exploration of the intrinsic correlations between different sensor information.

[0006] At low temperatures, the electrochemical reaction kinetics of lithium-ion batteries slow down significantly, ion conductivity decreases, and polarization intensifies, leading to a sharp deterioration in battery performance. Existing battery management system (BMS) monitoring methods, designed based on the electrochemical characteristics of ordinary lithium-ion batteries, cannot accurately identify and monitor the true state of low-temperature resistant lithium-ion batteries. Due to their unique material systems and electrochemical characteristics, low-temperature resistant lithium-ion batteries face the following challenges in state monitoring: the ion conduction mechanism of modified electrolytes differs from that of ordinary electrolytes, and traditional internal resistance measurement methods cannot accurately reflect their true electrochemical state; the lithium-ion insertion and extraction processes of modified anode materials have unique kinetic characteristics, which existing polarization analysis methods struggle to accurately capture; and safety risk identification at low temperatures requires consideration of the special response characteristics of the modified materials.

[0007] To address the aforementioned technical challenges, this invention proposes a low-temperature lithium battery BMS monitoring method and system based on collaborative sensing. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention provides a method and system for monitoring the BMS of low-temperature resistant lithium batteries based on collaborative sensing. This invention collects multi-source data, constructs a multi-source state feature matrix through spatiotemporal alignment and data fusion, and combines this with composite current excitation signals to perform impedance analysis and spectrum analysis, extracting impedance parameters, nonlinear distortion components, and electrochemical stress factors. A feature fusion algorithm based on a self-attention mechanism is then used to form a multi-dimensional feature vector. By fusing the multi-source state feature matrix and the multi-dimensional feature vector using a deep learning algorithm at the decision-level, the state monitoring results are obtained. This invention significantly improves the state perception and fault diagnosis capabilities of low-temperature resistant lithium batteries through multi-sensor fusion and multi-dimensional feature fusion, providing technical assurance for the safe and reliable application of low-temperature resistant lithium batteries.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] A low-temperature lithium battery BMS monitoring method based on collaborative sensing includes:

[0011] Collect surface temperature distribution data, ambient temperature data, and electrochemical parameters of low-temperature resistant lithium batteries;

[0012] A composite current excitation signal is applied to the low-temperature resistant lithium battery, and the voltage response signal of the low-temperature resistant lithium battery is acquired simultaneously. The impedance parameters at a preset characteristic frequency point are obtained through impedance analysis.

[0013] Based on the voltage response signal, the nonlinear distortion component is extracted by spectrum analysis to obtain the relaxation voltage curve. The electrochemical stress factor is obtained by differentiating the relaxation voltage curve.

[0014] Based on the surface temperature distribution data, ambient temperature data, and electrochemical parameters, a multi-source state feature matrix is ​​constructed through spatiotemporal alignment and data fusion.

[0015] Based on the impedance parameters, nonlinear distortion components, and electrochemical stress factors, a multidimensional feature vector is formed through feature fusion.

[0016] Based on the multi-source state feature matrix and multi-dimensional feature vector, a decision-level fusion is performed using a deep learning algorithm to obtain the state monitoring results.

[0017] The composite current excitation signal includes a fundamental frequency excitation signal, a high-frequency disturbance signal, and a preset current interruption sequence, wherein:

[0018] The fundamental frequency excitation signal has a preset amplitude and frequency, and is used to excite the electrochemical response of the low-temperature resistant lithium battery;

[0019] The amplitude of the high-frequency disturbance signal is smaller than that of the fundamental frequency excitation signal, and the frequency is greater than that of the fundamental frequency excitation signal, which is used to obtain the impedance parameters of the low-temperature resistant lithium battery at a preset characteristic frequency point.

[0020] The preset current interruption sequence includes multiple current interruption points of preset duration, used to obtain the relaxation voltage of the low-temperature resistant lithium battery.

[0021] The frequency range of the composite current excitation signal covers the characteristic spectrum of low-temperature resistant lithium batteries from the diffusion process to the charge transfer process under low-temperature conditions, and the temperature range of the low-temperature environment is determined based on industry standards.

[0022] The process of obtaining impedance parameters at a preset characteristic frequency point through impedance analysis includes:

[0023] Based on the composite current excitation signal and voltage response signal, Fourier transform is used to obtain current frequency domain response data and voltage frequency domain response data, respectively.

[0024] Based on the ratio of the voltage frequency domain response data to the current frequency domain response data, the complex impedance value at each frequency point is calculated.

[0025] Based on the complex impedance values ​​at each frequency point, the impedance amplitude and phase angle parameters at a preset characteristic frequency point are extracted to obtain the impedance parameters. The preset characteristic frequency point is determined based on the impedance response characteristics of the low-temperature resistant lithium battery at different frequencies.

[0026] The step of extracting the nonlinear distortion component through spectral analysis based on the voltage response signal to obtain the relaxation voltage curve includes:

[0027] Based on the voltage response signal, the fundamental frequency component and its harmonic frequency components are obtained through adaptive wavelet transform and short-time Fourier transform.

[0028] Based on the harmonic frequency components, the amplitude and phase information of the harmonic frequency components are extracted using spectral analysis methods to construct nonlinear distortion components;

[0029] Based on the nonlinear distortion components, a mapping model between electrochemical parameters and voltage response is constructed using a piecewise linear and nonlinear hybrid fitting method.

[0030] Based on the mapping relationship model, the relaxation voltage curve is obtained by extracting the transient response characteristics of the voltage response signal to the preset current interruption sequence.

[0031] The electrochemical stress factor is obtained by differentiating the relaxation voltage curve, including:

[0032] Based on the relaxation voltage curve, the voltage change rate curve, voltage acceleration curve, and voltage jump characteristic curve are obtained through multi-order differential processing.

[0033] Identify the extreme points, inflection points, and steady-state points of each curve from the voltage change rate curve, voltage acceleration curve, and voltage jump characteristic curve, and calculate the time constant and amplitude parameters corresponding to each characteristic point;

[0034] Based on the time constant and amplitude parameters, an equivalent circuit model of the lithium battery is constructed. By mapping the time constant to the electrochemical reaction rate constant and the amplitude parameters to the polarization impedance parameters, and combining the temperature compensation function, the correlation between the equivalent circuit model and the electrochemical kinetic characteristics of the low-temperature resistant lithium battery under low-temperature conditions is established.

[0035] Based on the polarization impedance parameters, electrochemical reaction rate constant, and correlation, the polarization stress coefficient, kinetic response coefficient, and temperature compensation coefficient are calculated respectively, and combined to form the electrochemical stress factor.

[0036] The process of forming a multi-dimensional feature vector through feature fusion based on the impedance parameter, nonlinear distortion component, and electrochemical stress factor includes:

[0037] Based on the impedance amplitude and phase angle parameters at preset characteristic frequencies in the impedance parameters, a nonlinear dimensionality reduction algorithm is used to perform dimensionality reduction processing to obtain the impedance characteristic feature vector.

[0038] Based on the amplitude and phase information of the harmonic frequency components in the nonlinear distortion components, feature extraction is performed using principal component analysis to obtain the nonlinear response feature vector.

[0039] Based on the polarization stress coefficient, kinetic response coefficient, and temperature compensation coefficient in the electrochemical stress factor, an electrochemical activity feature vector is obtained through mean normalization.

[0040] Based on the nonlinear response feature vector, electrochemical activity feature vector, and impedance characteristic feature vector, a weighted fusion is performed using a feature fusion algorithm based on a self-attention mechanism to obtain the fused feature vector.

[0041] Based on the fused feature vector, temperature compensation processing is performed by combining the surface temperature distribution data and ambient temperature data to form a multi-dimensional feature vector adapted to low-temperature environments.

[0042] The step involves performing temperature compensation processing based on the fused feature vector, combined with the surface temperature distribution data and ambient temperature data, to form a multi-dimensional feature vector adapted to low-temperature environments, including:

[0043] Based on the surface temperature distribution data, ambient temperature data, and fused feature vector, the temperature sensitivity coefficient of each feature component is calculated by constructing a mapping relationship between temperature and each feature component in the fused feature vector. Feature components with temperature sensitivity coefficients greater than a preset sensitivity threshold are identified as feature components to be compensated.

[0044] For the characteristic components to be compensated originating from the electrochemical activity characteristic vector, the compensation value is calculated using an exponential temperature compensation function; for the characteristic components to be compensated originating from the impedance characteristic vector, the compensation value is calculated using a piecewise linear temperature compensation function; and for the characteristic components to be compensated originating from the nonlinear response characteristic vector, the compensation value is calculated using a polynomial temperature compensation function.

[0045] Temperature compensation is performed on the feature components to be compensated based on the compensation value, and the corresponding feature components in the fused feature vector are updated to generate a multidimensional feature vector.

[0046] The process involves constructing a multi-source state feature matrix based on the surface temperature distribution data, ambient temperature data, and electrochemical parameters through spatiotemporal alignment and data fusion, including:

[0047] Based on the surface temperature distribution data, temperature gradient, hot spot distribution, and temperature uniformity indicators are extracted through thermal imaging processing to form a temperature feature sub-matrix.

[0048] Based on the ambient temperature data, time-series analysis is used to extract the ambient temperature change rate, temperature fluctuation amplitude, and temperature stability index to form an environmental feature sub-matrix.

[0049] Based on the electrochemical parameters, the state of charge, internal resistance change trend and capacity decay characteristics are extracted through incremental capacity analysis and electrochemical feature engineering to form an electrochemical feature sub-matrix.

[0050] The temperature feature sub-matrix, environmental feature sub-matrix, and electrochemical feature sub-matrix are aligned in time and space using a timestamp-based temporal alignment algorithm and a spatial registration algorithm to obtain aligned temperature feature sub-matrix, environmental feature sub-matrix, and electrochemical feature sub-matrix.

[0051] The aligned temperature feature submatrix, environmental feature submatrix, and electrochemical feature submatrix are fused through tensor decomposition and reconstruction to construct a multi-source state feature matrix.

[0052] The step of obtaining state monitoring results by performing decision-level fusion using a deep learning algorithm based on the multi-source state feature matrix and multi-dimensional feature vectors includes:

[0053] Based on the multi-source state feature matrix, temporal dependence and spatial correlation are extracted through a spatiotemporal convolutional neural network to obtain state feature representations and generate a first decision result and its confidence level.

[0054] Based on the multidimensional feature vector, a deep fully connected network is used to extract nonlinear feature relationships, obtain parametric feature representations, and generate a second decision result and its confidence level.

[0055] Based on the first decision result and its confidence level, and the second decision result and its confidence level, a weighted voting mechanism is used, combined with a preset confidence level threshold and decision rules, to achieve decision-level fusion and obtain the status monitoring results.

[0056] The low-temperature lithium battery BMS monitoring system based on collaborative sensing includes a data acquisition module, a data processing module, a data analysis module, and a monitoring decision module, wherein:

[0057] The data acquisition module is used to acquire surface temperature distribution data, ambient temperature data and electrochemical parameters of the low-temperature resistant lithium battery, as well as to apply a composite current excitation signal to the low-temperature resistant lithium battery and simultaneously acquire the voltage response signal.

[0058] The data processing module is used to preprocess the data acquired by the data acquisition module, obtain the impedance parameters at a preset characteristic frequency point through impedance analysis, extract the nonlinear distortion component through spectrum analysis, obtain the relaxation voltage curve, and extract the electrochemical stress factor.

[0059] The data analysis module is used to construct a multi-source state feature matrix and a multi-dimensional feature vector, and to perform decision-level fusion analysis through deep learning algorithms;

[0060] The monitoring decision module is used to generate status monitoring results based on the analysis results output by the data analysis module.

[0061] Compared with the prior art, the beneficial effects of the present invention are:

[0062] This invention acquires multi-dimensional electrochemical kinetic characteristics under low-temperature conditions by combining composite current excitation signals with multi-source data acquisition, impedance analysis, and spectral analysis. It employs spatiotemporal alignment and tensor decomposition reconstruction algorithms to achieve dynamic fusion of multi-source state feature matrices and multi-dimensional features. Decision-level fusion is performed using deep learning algorithms to achieve real-time and accurate intelligent monitoring of the BMS of low-temperature resistant lithium batteries. Compared to existing methods, this invention overcomes the technical bottleneck of decreased accuracy in traditional monitoring technologies under low-temperature conditions, avoids the interference of temperature changes on electrochemical parameter identification, and achieves adaptive fusion of multi-dimensional features. While ensuring high-precision state monitoring under extreme low-temperature conditions, it significantly improves the reliability of fault early warning. Attached Figure Description

[0063] Figure 1 This is a flowchart of the steps of the low-temperature lithium battery BMS monitoring method based on collaborative sensing in an embodiment of the present invention;

[0064] Figure 2 This is a flowchart illustrating the steps of obtaining impedance parameters at a preset characteristic frequency point through impedance analysis in an embodiment of the present invention.

[0065] Figure 3 This is a flowchart illustrating the steps for obtaining the electrochemical stress factor in an embodiment of the present invention. Detailed Implementation

[0066] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof.

[0067] The term "and / or" in the following text is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0068] This application provides an intelligent monitoring method, including:

[0069] Collect multi-source environmental data and electrochemical parameters of the monitored objects;

[0070] A composite excitation signal is applied to the monitored object, and the response signal is collected synchronously. The impedance parameters at the characteristic frequency point are obtained through analysis and processing.

[0071] Based on the response signal, distortion components are extracted through spectral analysis to obtain a characteristic curve, and the characteristic curve is processed to obtain the stress factor;

[0072] Based on the multi-source environmental data and electrochemical parameters, a multi-source state feature matrix is ​​constructed through spatiotemporal alignment and data fusion;

[0073] Based on the impedance parameters, distortion components, and stress factors, a multidimensional feature vector is formed through feature fusion.

[0074] Based on the multi-source state feature matrix and multi-dimensional feature vector, a decision-level fusion is performed using a deep learning algorithm to obtain the state monitoring results.

[0075] For example, this embodiment takes the intelligent monitoring of a low-temperature resistant lithium battery BMS in a low-temperature environment as an example to specifically illustrate the above-mentioned intelligent monitoring method.

[0076] In related technologies, the monitoring of lithium battery BMS is usually designed only based on the ambient temperature environment, lacking consideration for low temperature environment. On the one hand, it does not take into account the impact of ambient temperature changes on different parameter characteristics; on the other hand, it does not take into account the significant changes in electrochemical reaction characteristics caused by temperature changes. The multi-source information fusion strategy is fixed and lacks the ability to adaptively adjust the sensor weight allocation and data fusion strategy according to environmental characteristics, and cannot cope with the problem of different sensor accuracy changes and response time differences at low temperatures.

[0077] Therefore, it is often difficult for related technologies to achieve a perfect balance between improving the accuracy of low-temperature monitoring, reducing the impact of temperature interference, and enhancing the robustness of the system. This is especially true in application scenarios such as new energy vehicles in cold regions, where the low-temperature reliability requirements of lithium battery monitoring systems are extremely high. Existing lithium battery BMS monitoring methods are unable to meet the dual needs of high-precision condition assessment and rapid fault early warning in low-temperature environments.

[0078] To avoid the adverse effects of temperature changes on monitoring accuracy and to achieve accurate acquisition of multi-dimensional electrochemical characteristics under low-temperature conditions, this application provides a low-temperature resistant lithium battery BMS monitoring method based on collaborative sensing. Its application scenarios include, but are not limited to, new energy vehicle power battery systems in cold regions, large-scale battery arrays in low-temperature energy storage power stations, and lithium batteries for portable electronic devices in polar or high-altitude regions.

[0079] It should be noted that the intelligent monitoring method proposed in this application does not depend on a fixed type of battery or a single application scenario, but is aimed at lithium battery monitoring scenarios with multi-dimensional state changes, complex failure modes and high reliability requirements under low temperature environment.

[0080] It is worth noting that the method of this application is not only applicable to low-temperature resistant lithium battery BMS systems, but can also be applied to other complex systems that require intelligent state monitoring, such as fuel cell health management, supercapacitor state assessment, or monitoring systems for other electrochemical energy storage devices.

[0081] In this application, low-temperature environment specifically refers to the operating condition where the ambient temperature is lower than room temperature when the lithium battery is working. Based on the common characteristics of the electrochemical properties of low-temperature resistant lithium batteries, this application defines the low-temperature environment temperature range as -40℃ to 5℃. This range can meet the low-temperature performance requirements of lithium batteries in multiple industries, specifically complying with the core indicator requirements for low-temperature performance testing of low-temperature resistant lithium batteries in relevant standards and industry specifications in fields such as energy storage systems, electric vehicles, consumer electronics, and industrial equipment. There is no need to adjust the basic temperature range due to differences in application scenarios.

[0082] Within the low-temperature range of -40℃ to 5℃, the electrochemical characteristics of low-temperature resistant lithium batteries deteriorate significantly, specifically manifested as a sharp decrease in ion conductivity, a substantial increase in charge transfer impedance, and a reduction in the utilization rate of electrode active materials. These changes directly lead to battery capacity decay, a decrease in power output capability, and an increase in safety risks such as lithium plating and internal short circuits. This is the core issue that this application focuses on addressing.

[0083] Furthermore, the monitoring method of this application possesses wide temperature adaptability and cross-industry compatibility. It is not only applicable to the aforementioned low-temperature environments, but also adaptable to normal temperature environments (such as 23℃±2℃, 25℃±2℃) and high-temperature environments (such as 45℃, 55℃, 60℃) defined by various industry standards. Its multimodal sensing fusion module and intelligent analysis algorithm can adaptively adjust core parameters based on the differences in the characteristics of lithium batteries in different industries and changes in temperature conditions, referring to the performance indicators of corresponding industry standards. These parameters include the spatiotemporal alignment weights of multi-source sensor data, the frequency and amplitude of composite current excitation, feature extraction thresholds, and deep learning model decision parameters. This ensures that high-precision condition assessment and fault early warning functions that meet the requirements of corresponding industry standards can be achieved in different application fields such as energy storage, automobiles, consumer electronics, and industrial equipment, as well as in the full temperature range of low temperature, normal temperature, and high temperature.

[0084] It is understood that in this application, the low-temperature resistant lithium battery BMS monitoring based on collaborative sensing specifically refers to the comprehensive and multi-dimensional intelligent state monitoring of low-temperature resistant lithium battery systems through multi-sensor data fusion technology. Under normal operating conditions, the battery system continuously collects surface temperature distribution data, ambient temperature data, electrochemical parameters, and voltage response signals under composite current excitation through a multi-sensor array. The intelligent monitoring system performs real-time analysis using impedance analysis, spectrum analysis, and deep learning algorithms. Its main function is to maintain continuous tracking of the battery's internal state to achieve early fault warning. Due to the extremely high safety requirements of battery systems, the multi-sensor fusion intelligent monitoring method must not only possess conventional state assessment functions but also intelligent identification of complex failure modes and predictive control functions for risk events.

[0085] It is worth noting that although the implementation of the multi-sensor data acquisition, composite current excitation signal generation, deep learning network model construction, and decision-level fusion technologies of this invention requires certain hardware investment and algorithm development costs, it can avoid economic losses and safety risks caused by sudden failures of the battery system through early fault warning. At the same time, the predictive maintenance mode based on multi-dimensional feature vectors and state feature matrices can significantly extend battery life and reduce unplanned downtime. The safety guarantee value and operational efficiency improvement it brings far exceed the cost of technology implementation. Moreover, with the large-scale application and the improvement of technology maturity, the unit implementation cost will continue to decrease, and the investment return period is reasonable.

[0086] To achieve multi-sensor data fusion and deep learning decision-making, and significantly improve fault warning accuracy and battery safety protection capabilities, such as... Figure 1 As shown, the execution steps of the low-temperature lithium battery BMS monitoring method based on collaborative sensing in this embodiment include:

[0087] S1: Collect surface temperature distribution data, ambient temperature data, and electrochemical parameters of low-temperature resistant lithium batteries;

[0088] In this step, a multi-sensor array is used to collect comprehensive data from the low-temperature resistant lithium-ion battery system. Surface temperature distribution data is acquired using an infrared thermal imaging sensor, enabling real-time monitoring of the temperature field distribution and hotspot changes on the battery surface. Ambient temperature data is collected using a high-precision temperature sensor to assess the impact of the external environment on battery performance. Electrochemical parameters, including key indicators such as voltage, current, and state of charge, are acquired synchronously using sensors built into the electrochemical workstation and BMS. The multi-sensor data acquisition system ensures the time synchronization of various data types through a unified timestamp mechanism, laying the foundation for subsequent data fusion and analysis.

[0089] Specifically, an infrared thermal imaging sensor is used to scan and measure the surface of the low-temperature resistant lithium battery, obtaining the temperature values ​​of various regions on the surface of the low-temperature resistant lithium battery, forming surface temperature distribution data. This surface temperature distribution data records the spatial distribution characteristics of the battery surface temperature in the form of a two-dimensional temperature field, which can reflect the non-uniformity of heat generation and conduction inside the battery.

[0090] Meanwhile, high-precision temperature sensors deployed around the battery measure the ambient temperature in real time, acquiring environmental temperature data. This data, recorded as time-series data, tracks the trend of ambient temperature changes, providing fundamental information for assessing the impact of the external environment on battery performance.

[0091] In addition, key electrochemical indicators such as voltage, current, and state of charge of the low-temperature resistant lithium battery are simultaneously collected using the voltage and current sensors built into the electrochemical workstation and BMS, forming electrochemical parameters. These electrochemical parameters contain real-time operating status information of the battery, reflecting its electrochemical activity and performance level.

[0092] To ensure data temporal consistency, all sensors employ a unified timestamp mechanism for data tagging, keeping surface temperature distribution data, ambient temperature data, and electrochemical parameters synchronized over time. The acquired multi-source data undergoes filtering, noise reduction, and format standardization via a data preprocessing module, providing high-quality input data for subsequent data fusion and analysis. This multi-sensor data acquisition method comprehensively acquires multi-dimensional state information of the battery system, laying a reliable data foundation for intelligent monitoring in low-temperature environments.

[0093] S2: Apply a composite current excitation signal to the low-temperature resistant lithium battery, simultaneously acquire the voltage response signal of the low-temperature resistant lithium battery, and obtain the impedance parameters at a preset characteristic frequency point through impedance analysis;

[0094] In this step, the frequency domain characteristics of the battery are obtained through a composite current excitation technique and impedance analysis. The composite current excitation signal includes a fundamental frequency excitation, high-frequency perturbations, and a current interruption sequence, covering the complete characteristic spectrum of the battery under low-temperature conditions. Impedance analysis technology processes the voltage and current responses through Fourier transform to extract impedance parameters at preset characteristic frequencies. This method can deeply mine the internal electrochemical state information of the battery, providing high-precision frequency domain characteristics for health status assessment and effectively enhancing the monitoring accuracy under low-temperature conditions.

[0095] like Figure 2The diagram shows the steps for obtaining impedance parameters at a preset characteristic frequency point through impedance analysis in this embodiment of the application. Parallelograms represent specific operational steps, and rectangles represent intermediate or final data. First, a composite current excitation signal is constructed and applied to a low-temperature resistant lithium battery. The composite current excitation signal consists of three components: a fundamental frequency excitation signal, a high-frequency perturbation signal, and a preset current interruption sequence. The fundamental frequency excitation signal has a preset amplitude and frequency, used to excite the electrochemical response of the low-temperature resistant lithium battery, providing the main excitation energy for impedance testing. The high-frequency perturbation signal has a smaller amplitude and a higher frequency than the fundamental frequency excitation signal, used to obtain the impedance parameters of the low-temperature resistant lithium battery at the preset characteristic frequency point, enabling the detection of rapid electrochemical processes within the battery. The preset current interruption sequence contains multiple current interruption points of preset duration, used to obtain the relaxation voltage of the low-temperature resistant lithium battery, providing a data basis for subsequent electrochemical stress factor analysis. The frequency range of the composite current excitation signal covers the characteristic spectrum of low-temperature resistant lithium batteries from diffusion to charge transfer processes under low-temperature conditions, ensuring that it can fully excite various electrochemical response characteristics of low-temperature resistant lithium batteries under low-temperature conditions.

[0096] Simultaneously, the voltage response signal generated by the low-temperature resistant lithium battery under the action of a composite current excitation signal is acquired. High-precision voltage measurement technology is employed to monitor the changes in the terminal voltage of the low-temperature resistant lithium battery in real time, converting the analog voltage signal into a digital voltage response signal. The voltage response signal contains complete response information of the low-temperature resistant lithium battery to the composite current excitation signal, including the main response to the fundamental frequency excitation signal, the impedance response to high-frequency disturbance signals, and the relaxation response under the action of a current interruption sequence. To ensure test accuracy, strict time synchronization control is used for voltage acquisition and current excitation to guarantee the precise correspondence between the composite current excitation signal and the voltage response signal on the time axis.

[0097] Next, impedance parameters at preset characteristic frequencies are obtained through impedance analysis. First, frequency domain transformation is performed using Fourier transform based on the composite current excitation signal and voltage response signal. The time-domain composite current excitation signal is converted into current frequency domain response data, and the time-domain voltage response signal is converted into voltage frequency domain response data. The current frequency domain response data reflects the amplitude and phase distribution characteristics of the composite current excitation signal at various frequencies, and the voltage frequency domain response data reflects the amplitude and phase distribution characteristics of the voltage response signal at various frequencies.

[0098] Then, based on the ratio of voltage frequency domain response data to current frequency domain response data, the complex impedance values ​​at each frequency point are calculated. The complex impedance values ​​at each frequency point can fully describe the impedance characteristics of low-temperature resistant lithium batteries at different frequencies, reflecting the frequency dependence of electrochemical parameters such as resistance, capacitance, and inductance inside the battery.

[0099] Finally, based on the complex impedance values ​​at each frequency point, the impedance amplitude and phase angle parameters at preset characteristic frequencies are extracted to obtain the impedance parameters. The preset characteristic frequencies are determined based on the impedance response characteristics of low-temperature resistant lithium batteries at different frequencies. Typically, characteristic frequency points that reflect key electrochemical processes within the battery are selected, such as ohmic impedance frequencies, charge transfer impedance frequencies, and diffusion impedance frequencies. The impedance amplitude and phase angle parameters corresponding to the preset characteristic frequencies are extracted from the complex impedance values ​​at each frequency point to form the impedance parameters. These impedance parameters contain the impedance amplitude and phase angle information of the low-temperature resistant lithium battery at the preset characteristic frequencies, quantitatively reflecting key electrochemical parameters such as the battery's internal resistance characteristics, charge transfer kinetics, and ion diffusion characteristics. This provides important frequency domain feature data for subsequent multidimensional feature vector construction.

[0100] S3: Based on the voltage response signal, extract the nonlinear distortion component through spectrum analysis to obtain the relaxation voltage curve, and perform differential processing on the relaxation voltage curve to obtain the electrochemical stress factor.

[0101] In this step, spectral analysis and differential processing methods are used to extract the electrochemical stress factor. Nonlinear distortion components are extracted using adaptive wavelet transform and short-time Fourier transform. The relaxation voltage curve is obtained based on transient response characteristics, and feature points are identified and the time constant is calculated through multi-order differential processing. This technique can quantify the electrochemical kinetic characteristics inside the battery, providing a sensitive stress index for fault early warning and significantly improving the early identification capability of abnormal states.

[0102] S4: Based on the surface temperature distribution data, ambient temperature data, and electrochemical parameters, a multi-source state feature matrix is ​​constructed through spatiotemporal alignment and data fusion;

[0103] In this step, a multi-source state feature matrix is ​​constructed using spatiotemporal alignment technology and tensor fusion methods. Key indicators from each data source are extracted through feature engineering, and spatiotemporal synchronization of multi-source data is achieved based on temporal alignment and spatial registration algorithms. Deep fusion is realized through tensor decomposition and reconstruction. This method effectively integrates multi-dimensional heterogeneous information, constructs a feature matrix containing rich state information, and provides a comprehensive state representation foundation for intelligent decision-making.

[0104] S5: Based on the impedance parameters, nonlinear distortion components, and electrochemical stress factors, a multidimensional feature vector is formed through feature fusion.

[0105] In this step, multi-level feature fusion technology and temperature compensation methods are employed to form multi-dimensional feature vectors. Different types of feature vectors are obtained through dimensionality reduction algorithms, principal component analysis, and normalization. Weighted fusion is achieved based on a self-attention mechanism, and temperature robustness is enhanced by incorporating multiple types of temperature compensation functions. This technique achieves effective integration and temperature adaptability optimization of electrochemical multi-dimensional features, significantly improving the stability and expressive power of feature vectors in low-temperature environments.

[0106] S6: Based on the multi-source state feature matrix and multi-dimensional feature vector, a decision-level fusion is performed using a deep learning algorithm to obtain the state monitoring result.

[0107] In this step, a dual-path deep learning network and decision-level fusion technology are employed to obtain the final monitoring results. A spatiotemporal convolutional neural network and a deep fully connected network are used to process the multi-source state feature matrix and multi-dimensional feature vector, respectively, and decision fusion is achieved based on a weighted voting mechanism. This method fully leverages the nonlinear modeling capabilities of deep learning and the complementary advantages of decision fusion, achieving high-precision state monitoring and reliable fault early warning, providing intelligent decision support for battery safety management.

[0108] Please see Figure 3 This is a flowchart illustrating the steps for obtaining the electrochemical stress factor according to an embodiment of this application. Parallelograms represent specific operational steps, and rectangles represent intermediate or final data. The specific implementation steps of S3 are as follows:

[0109] S3.1: Based on the voltage response signal, extract the nonlinear distortion component through spectrum analysis to obtain the relaxation voltage curve;

[0110] In this embodiment, the voltage response signal is analyzed in depth using spectrum analysis technology to extract the nonlinear distortion component and obtain the relaxation voltage curve.

[0111] First, based on the voltage response signal, the fundamental frequency components and their harmonic frequency components are obtained through adaptive wavelet transform and short-time Fourier transform. Specifically, the voltage response signal is used as input data and multi-resolution decomposition is performed through adaptive wavelet transform. The wavelet basis function is adaptively selected for the voltage response signal, and the scale and translation parameters of the wavelet basis function are automatically adjusted by minimizing the reconstruction error criterion to obtain the optimal wavelet basis function. Second, based on the optimal wavelet basis function, multi-scale decomposition is performed on the voltage response signal, and decomposition analysis is performed at different time and frequency scales to obtain the wavelet coefficient matrix. Finally, a transient feature recognition algorithm is applied to the wavelet coefficient matrix, and combined with the time-varying characteristics of the electrochemical process, the transient response and local frequency components in the signal are identified to obtain the transient feature vector.

[0112] Based on the voltage response signal, segmented spectral analysis is performed using a short-time Fourier transform. A sliding window function is used to segment the voltage response signal, resulting in segmented signals. An independent short-time Fourier transform is performed on each segmented signal within a time window to obtain a time-frequency matrix. The variation of the voltage response signal's spectrum over time in the time-frequency matrix is ​​identified, yielding a dynamic spectral feature matrix. The wavelet coefficient matrix, transient eigenvectors, and dynamic spectral feature matrix are then comprehensively processed to obtain the fundamental frequency components and their corresponding harmonic frequency components.

[0113] Furthermore, based on the harmonic frequency components, the amplitude and phase information of the harmonic frequency components are extracted using spectral analysis methods to construct nonlinear distortion components. Specifically, the harmonic frequency components are used as input data to perform multi-scale spectral decomposition: in the low-frequency band, a high-frequency resolution analysis window is used to perform fine spectral analysis on the harmonic frequency components to obtain a low-frequency fine spectrum; in the high-frequency band, a high-time resolution analysis window is used to perform fast spectral analysis on the harmonic frequency components to obtain a high-frequency fast spectrum; the low-frequency fine spectrum and the high-frequency fast spectrum are processed collaboratively to extract fine spectral features of each harmonic to obtain a multi-scale harmonic feature matrix.

[0114] Based on the multi-scale harmonic feature matrix, amplitude and phase separation are performed through harmonic feature extraction. Specifically, the multi-scale harmonic feature matrix is ​​processed using complex decomposition to extract the real and imaginary parts of each harmonic, calculate the amplitude of each harmonic, and obtain the harmonic amplitude vector. The phase difference between each harmonic and the fundamental frequency component is calculated using the arctangent function to obtain the harmonic phase vector. Based on the harmonic amplitude vector and harmonic phase vector, nonlinear distortion components are constructed by calculating the amplitude ratio and phase relationship of each harmonic relative to the fundamental frequency.

[0115] Furthermore, based on the nonlinear distortion components, a mapping model between electrochemical parameters and voltage response is constructed using a piecewise linear and nonlinear hybrid fitting method. Specifically, based on the nonlinear distortion components, a segmentation point determination process is performed by dividing the working interval. A preset linear threshold is determined based on prior knowledge of the electrochemical characteristics of low-temperature resistant lithium batteries; the amplitude of the nonlinear distortion components is analyzed through comparative operations; when the amplitude is less than the preset linear threshold, it is divided into a linear interval; when the amplitude is greater than the preset linear threshold, it is divided into a nonlinear interval; a continuity constraint condition ensures a smooth transition between each interval, resulting in a complete interval division result.

[0116] Based on the interval division results and the nonlinear distortion components, a corresponding fitting strategy is executed through hybrid fitting. In the linear interval, least squares linear fitting is performed on the nonlinear distortion components to obtain linear mapping parameters; in the nonlinear interval, Newton-Raphson method or gradient descent method is performed on the nonlinear distortion components for polynomial fitting to obtain nonlinear mapping parameters; the linear and nonlinear mapping parameters are combined using a piecewise function construction method to obtain a complete mapping relationship model.

[0117] Finally, based on the mapping relationship model, the relaxation voltage curve is obtained by extracting the transient response characteristics of the voltage response signal to the preset current interruption sequence. Using the voltage response signal and the preset current interruption sequence as input data, the current interruption time in the preset current interruption sequence is identified using a threshold detection method to obtain the interruption time vector; the response of the voltage response signal at the interruption time is analyzed using differential operations to obtain the initial jump characteristics; the dynamic changes of the voltage response signal after the interruption are analyzed using the sliding window averaging method to obtain the transient response characteristics; and the final stable value of the voltage response signal is identified using statistical convergence judgment to obtain the steady-state response characteristics.

[0118] Through relaxation process analysis, initial jump characteristics, transient response characteristics, steady-state response characteristics, and a mapping relationship model are used as input data. Relaxation characteristic processing is then performed: the transient response characteristics are fitted with multiple exponential functions using the least squares method to obtain the decay time constant vector; combined with the mapping relationship model, the decay time constant vector is processed by solving the diffusion equation to obtain the electrochemical process time constant; and the steady-state response characteristics are processed by fitting an exponential decay model to obtain the relaxation completion time. Based on the initial jump characteristics, transient response characteristics, steady-state response characteristics, electrochemical process time constant, and relaxation completion time, a piecewise function construction method is used for comprehensive processing to obtain the relaxation voltage curve.

[0119] S3.2: Differentiate the relaxation voltage curve to obtain the electrochemical stress factor;

[0120] In this embodiment, by performing multi-order differential processing and feature analysis on the relaxation voltage curve, the electrochemical stress factor that can reflect the electrochemical kinetic characteristics of low-temperature resistant lithium batteries is finally obtained.

[0121] First, based on the relaxation voltage curve, the voltage change rate curve, voltage acceleration curve, and voltage jump characteristic curve are obtained through multi-order differential processing. Specifically, the relaxation voltage curve is used as input data through a first-order differential operation, and first-order derivative calculation is performed. The first-order derivative of the relaxation voltage curve is calculated using the central difference method to obtain the original voltage change rate. A Butterworth low-pass filter is used to filter the original voltage change rate to eliminate high-frequency noise introduced by the differential operation, resulting in a filtered voltage change rate. The filtered voltage change rate is then processed through peak detection and slope change analysis to identify the dynamic response stage of the electrochemical reaction, thus obtaining the voltage change rate curve.

[0122] Using the voltage rate of change curve as input data, a second-order derivative calculation is performed via second-order differential operations. The five-point central difference method is used to calculate the second derivative of the voltage rate of change curve to obtain the original voltage acceleration. Kalman filtering is then used to remove noise from the original voltage acceleration, resulting in the filtered voltage acceleration. Zero-point detection and sign change analysis are then used to process the filtered voltage acceleration, identifying the acceleration and deceleration phases of the electrochemical process, determining the transition moments of the electrochemical reaction mechanism, and obtaining the voltage acceleration curve.

[0123] In practical implementation, acquiring the voltage jump characteristic curve involves three stages: data acquisition, signal processing, and feature extraction. First, the relaxation voltage curve is sampled at a high frequency to record the voltage change process with millisecond-level time resolution, ensuring complete capture of the rapid voltage fluctuations during the electrochemical reaction. Subsequently, the acquired raw voltage signal is filtered and smoothed. A Butterworth low-pass filter algorithm is used to effectively remove high-frequency noise and transient interference generated in the test system, and the voltage amplitude is standardized using a normalization method to ensure the comparability of data from different experimental batches. Second, in the signal processing stage, the voltage change rate and voltage acceleration are calculated using numerical differentiation methods to obtain the dynamic trend of voltage change over time. The rate characteristics of voltage change are extracted using the central difference method, and the acceleration characteristics of voltage change are further analyzed using a multi-point difference algorithm, thereby revealing the stage-specific characteristics of voltage change during the electrochemical reaction. Finally, in the feature extraction stage, by identifying the extreme points, inflection points, and steady-state points of the rate of change curve and acceleration curve, the response differences of voltage in the reaction initiation, transition, and steady-state stages are analyzed, and key parameters such as voltage abrupt change points, response delay time, and steady-state convergence time are extracted. Combining the above analysis results, a complete voltage jump characteristic curve can be obtained, which reflects the kinetic response law of low-temperature resistant lithium batteries in different electrochemical stages, providing a reliable basis for subsequent calculation of the electrochemical stress factor.

[0124] Furthermore, based on the voltage change rate curve, voltage acceleration curve, and voltage jump characteristic curve, the extreme points, inflection points, and steady-state points of each curve are identified, and the time constants and amplitude parameters corresponding to each characteristic point are calculated. Through extreme point identification, the voltage change rate curve, voltage acceleration curve, and voltage jump characteristic curve are used as input data, and point detection processing with zero first derivative is performed: the voltage change rate curve is processed using a local maximum / minimum search method to identify the maximum and minimum times of the electrochemical reaction rate, obtaining a set of rate extreme points; the voltage acceleration curve is processed using a threshold comparison and neighborhood search method to identify the peak times of the electrochemical process acceleration, obtaining a set of acceleration extreme points; the set of rate extreme points and the set of acceleration extreme points are processed using linear interpolation and numerical integration, and the time position and amplitude of the extreme points are measured to obtain the extreme characteristic parameters.

[0125] Through inflection point identification, the voltage change rate curve, voltage acceleration curve, and voltage jump characteristic curve are used as input data. Point detection processing where the second derivative is zero is performed: the second derivative sign change detection method is used to process each curve, identifying the inflection points where the curve's concavity / convexity changes, thus obtaining an inflection point set; the inflection point set is processed through time series segmentation and statistical analysis to determine the critical moments of electrochemical reaction mechanism transitions, obtaining a transition time vector; the inflection point set and transition time vector are processed through exponential fitting and time constant calculation to obtain time constant boundary information for different electrochemical processes.

[0126] The steady-state point identification process uses voltage change rate curves, voltage acceleration curves, and voltage jump characteristic curves as input data, and performs constant value approach point detection processing: A sliding window variance analysis method is used to process each curve; when the signal variance within a continuous time window is less than a variance threshold determined based on historical data statistics, it is identified as a steady-state point, resulting in a set of steady-state points. The steady-state point set is then processed using exponential decay fitting and a convergence criterion based on engineering experience to determine the completion time of electrochemical equilibrium establishment, obtaining the equilibrium time. Statistical analysis and time series processing are then used to process the steady-state point set and the equilibrium time to obtain the time parameters for the electrochemical system to reach dynamic equilibrium. Based on the extreme value characteristic parameters, time constant boundary information, and time parameters, precise calculations are performed through numerical integration and parameter estimation to obtain the time constant vector and amplitude parameter vector.

[0127] Furthermore, based on the time constant vector and amplitude parameter vector, an equivalent circuit model of the lithium battery is constructed. By mapping the time constant to the electrochemical reaction rate constant and the amplitude parameter to the polarization impedance parameter, and combining this with a temperature compensation function, the correlation between the equivalent circuit model and the electrochemical kinetic characteristics of the low-temperature resistant lithium battery under low-temperature conditions is established. Specifically, based on the time constant vector and amplitude parameter vector, an electrochemical impedance equivalent network is constructed using a multi-time constant decomposition method. First, the time constant vector is processed using the equivalent circuit, i.e., the time constant decomposition method is used. Combined with the impedance amplitude constraint conditions derived from circuit theory, the resistance and capacitance values ​​of each parallel branch in the electrochemical impedance equivalent network are calculated. The resistance and capacitance values ​​are then combined using a multi-level parallel circuit topology to construct an equivalent circuit structure with multiple parallel branches, thus obtaining the equivalent circuit model.

[0128] Based on the time constant vector, amplitude parameter vector, and preset current temperature parameter, parameter transformation is performed using an electrochemical parameter mapping method. Specifically, the Arrhenius equation is used to process the time constant vector, considering the activation energy and temperature dependence determined based on material properties, mapping the time constant to the electrochemical reaction rate constant to obtain the rate constant vector. The amplitude parameter vector is processed according to electrochemical impedance spectroscopy theory and complex impedance relations, mapping the amplitude parameter to polarization impedance parameters to obtain the polarization impedance vector, including activation polarization impedance, concentration polarization impedance, and ohmic polarization impedance.

[0129] Based on the rate constant vector, polarization impedance vector, and current temperature parameters, parameter correction is performed using a temperature effect compensation method. The Arrhenius temperature compensation algorithm is used to process the rate constant vector and polarization impedance vector, establishing the variation of electrochemical parameters with temperature, resulting in a temperature-dependent function. A low-temperature correction algorithm is then used to process the temperature-dependent function, considering factors such as decreased ion mobility and reduced electrochemical reactivity under low-temperature conditions, resulting in a low-temperature compensation function. Finally, an activation energy parameter calculation algorithm is used to process the temperature-dependent function and the low-temperature compensation function, ensuring the model's accuracy under low-temperature conditions, resulting in a temperature compensation function. Based on the equivalent circuit model, rate constant vector, polarization impedance vector, and temperature compensation function, a correlation relationship is established to comprehensively process the data, obtaining the correlation between the equivalent circuit model and the electrochemical kinetic characteristics of low-temperature tolerant lithium batteries under low-temperature conditions.

[0130] Finally, based on the polarization impedance vector, rate constant vector, and correlation relationships, the polarization stress coefficient, kinetic response coefficient, and temperature compensation coefficient are calculated respectively, and combined to form the electrochemical stress factor. Specifically, the polarization stress coefficient is calculated based on the polarization impedance vector using a polarization degree quantification method. A weighted average algorithm is used to process the activation polarization impedance, concentration polarization impedance, and ohmic polarization impedance, comprehensively considering the relative contributions of different types of polarization impedance to obtain the comprehensive polarization impedance. A normalization algorithm is then used to process the comprehensive polarization impedance, converting it into a dimensionless parameter reflecting the polarization degree and polarization recovery capability of the low-temperature resistant lithium battery, thus obtaining the polarization stress coefficient.

[0131] Based on the rate constant vector, the kinetic response coefficient is calculated using a reactivity quantification method. A weighted summation algorithm is used to process the rate constants of multiple electrochemical processes in the rate constant vector, comprehensively considering the rate contributions of each process to obtain the comprehensive reaction rate. A standardization algorithm is then used to process the comprehensive reaction rate, converting it into a dimensionless parameter reflecting the electrochemical reaction response rate and reactivity, thus obtaining the kinetic response coefficient.

[0132] Based on the correlation, current temperature parameters, and preset standard temperatures, a temperature compensation coefficient is calculated using a temperature influence quantification method. A temperature comparison algorithm is employed to compare the standard temperature and current temperature parameters, calculating the differential impact of temperature on different electrochemical processes to obtain a temperature influence factor. A compensation coefficient calculation algorithm is then used to process the temperature influence factor, converting it into a dimensionless parameter reflecting the degree of influence of the low-temperature environment, thus obtaining the temperature compensation coefficient.

[0133] The electrochemical stress factor is calculated using a weighted combination method based on the polarization stress coefficient, kinetic response coefficient, and temperature compensation coefficient. A weighting determination algorithm is employed to determine the weighting coefficients based on their impact on the overall battery performance. A weighted summation algorithm is then used to process the three coefficients, resulting in a comprehensive electrochemical stress factor. This factor fully reflects the electrochemical stress state and performance characteristics of low-temperature resistant lithium batteries under low-temperature conditions.

[0134] The specific implementation steps of S4 are as follows:

[0135] S4.1: Based on the surface temperature distribution data, temperature gradient, hot spot distribution and temperature uniformity indicators are extracted through thermal imaging processing to form a temperature feature sub-matrix;

[0136] In this embodiment, firstly, based on the surface temperature distribution data, a temperature gradient feature vector is extracted using the finite difference gradient calculation method. Specifically, the surface temperature distribution data is spatially differentiated using a central difference scheme to calculate the rate of temperature change in the x and y directions at each spatial location, obtaining the original temperature gradient field. The original temperature gradient field is then smoothed using a Gaussian filtering method, with appropriate filter kernel size and standard deviation parameters set to eliminate measurement noise and outliers, resulting in a filtered temperature gradient field. Finally, the filtered temperature gradient field is processed using gradient magnitude and direction calculation methods. The region with the most drastic temperature change is identified by calculating the magnitude of the gradient vector, and the gradient direction is determined by calculating the arctangent function, thus obtaining the temperature gradient feature vector.

[0137] Then, based on the surface temperature distribution data, hotspot distribution feature vectors are extracted using threshold segmentation and connected component analysis. An adaptive thresholding method or percentile method is used to determine the hotspot identification threshold. A binarization segmentation method is used to process the surface temperature distribution data, marking pixels with temperatures exceeding the hotspot identification threshold as foreground regions, thus obtaining candidate hotspot regions. A connected component labeling algorithm is used to process the candidate hotspot regions, merging adjacent high-temperature pixels into the same hotspot region based on pixel connectivity, resulting in a hotspot distribution map. Finally, a region attribute calculation method is used to process the hotspot distribution map, obtaining the hotspot location coordinates through centroid calculation, the temperature peak value through pixel value statistics, and the coverage area through pixel counting, thus obtaining the hotspot distribution feature vector.

[0138] Finally, based on the surface temperature distribution data, descriptive statistical analysis and spatial autocorrelation analysis methods were used to extract uniformity quantification indicators and spatial correlation coefficients. The surface temperature distribution data was statistically analyzed using the mean, variance, and standard deviation calculation methods to obtain temperature statistical parameters. The temperature statistical parameters were processed using the coefficient of variation calculation method, and the relative dispersion of the temperature distribution was calculated by the ratio of the standard deviation to the mean. The uniformity of the temperature distribution was quantified using the uniformity index calculation method based on the coefficient of variation, resulting in a uniformity quantification indicator. The surface temperature distribution data was processed using the spatial autocorrelation index calculation method, and the correlation between temperature values ​​at adjacent locations was analyzed based on the spatial weight matrix to obtain a spatial correlation coefficient. Based on the temperature gradient feature vector, hotspot distribution feature vector, uniformity quantification indicator, and spatial correlation coefficient, a temperature feature submatrix was obtained through feature vector concatenation.

[0139] S4.2: Based on the ambient temperature data, extract the ambient temperature change rate, temperature fluctuation amplitude, and temperature stability index through time series analysis to form an environmental feature sub-matrix.

[0140] In this embodiment, firstly, based on the ambient temperature data, the temperature change rate feature is extracted through time series differentiation processing. Then, based on the ambient temperature data, the temperature change at adjacent time points is calculated using the first-order difference method to obtain the instantaneous temperature change rate. Based on the instantaneous temperature change rate, short-term fluctuations are smoothed using a simple moving average method to obtain the smoothed temperature change rate. Finally, based on the smoothed temperature change rate, threshold segmentation and sign determination methods are used to identify time periods of temperature rise, fall, and stability, resulting in a temperature change rate feature vector.

[0141] Then, based on the ambient temperature data, a temperature fluctuation amplitude feature vector is extracted through fluctuation analysis. Specifically, based on the ambient temperature data, the temperature is segmented using a sliding window method to obtain the temperature sequence within each time window; based on the temperature sequence within each time window, the difference between the highest and lowest temperatures is calculated using maximum and minimum value calculation methods to obtain the local fluctuation amplitude; based on the local fluctuation amplitude, the mean, maximum, and standard deviation of the fluctuation amplitude are calculated using descriptive statistical methods to obtain the temperature fluctuation amplitude feature vector.

[0142] Finally, based on the ambient temperature data, temperature stability features are extracted through stability assessment. Specifically, based on the ambient temperature data, the long-term temperature change trend is identified using least squares linear regression to obtain temperature trend parameters; based on the ambient temperature data and temperature trend parameters, the degree of temperature deviation around the trend line is calculated using the root mean square error method to obtain the stability coefficient; based on the ambient temperature data, the duration of temperature remaining within a preset range is analyzed using threshold judgment and counting statistics methods to obtain the temperature persistence index. The temperature change rate feature vector, temperature fluctuation amplitude feature vector, stability coefficient, and temperature persistence index are combined using a feature vector concatenation method to obtain an environmental feature submatrix.

[0143] S4.3: Based on the electrochemical parameters, the state of charge, internal resistance change trend and capacity decay characteristics are extracted through incremental capacity analysis and electrochemical feature engineering to form an electrochemical feature sub-matrix.

[0144] In this embodiment, firstly, based on electrochemical parameters, incremental capacity analysis is used to extract state-of-charge (POC) characteristics. Specifically, based on the electrochemical parameters, an interpolation fitting method is used to establish the correspondence between voltage and capacity, resulting in a voltage-capacity curve. Based on the voltage-capacity curve, a numerical differentiation method is used to calculate the capacity change corresponding to a unit voltage change, resulting in an incremental capacity curve. Based on the incremental capacity curve, a local extremum detection method is used to identify the characteristic peak positions and amplitudes, obtaining POC characteristic parameters. Based on the POC characteristic parameters, a lookup table method and linear interpolation method are used, combined with the battery's nominal capacity information, to obtain the current POC.

[0145] Furthermore, based on the electrochemical parameters, an internal resistance evolution analysis is performed to extract the characteristic vector of internal resistance change trend. Specifically, based on the electrochemical parameters, the internal resistance values ​​at different times are calculated using Ohm's law to obtain the internal resistance time series; based on the internal resistance time series, the change trend of internal resistance over time is fitted using the least squares method to obtain the internal resistance trend function; based on the internal resistance trend function, the rate of change and acceleration of internal resistance are calculated using the first and second derivatives to obtain the characteristic vector of internal resistance change trend.

[0146] Finally, based on the electrochemical parameters, capacity degradation feature vectors are extracted through capacity degradation analysis. Specifically, based on the electrochemical parameters, the actual usable capacity is calculated using the coulomb counting method to obtain a capacity time series. Based on the capacity time series, a mathematical model of capacity degradation is fitted using an exponential decay model fitting method to obtain a capacity degradation function. Based on the capacity degradation function, features such as degradation rate, remaining lifetime prediction, and degradation mode are extracted using parameter extraction and extrapolation calculation methods to obtain a capacity degradation feature vector. The current state of charge, internal resistance change trend feature vectors, and capacity degradation feature vectors are combined using a feature vector concatenation method to obtain an electrochemical feature sub-matrix.

[0147] S4.4: The temperature feature sub-matrix, environmental feature sub-matrix, and electrochemical feature sub-matrix are aligned in time and space dimensions using a time-series alignment algorithm and a spatial registration algorithm based on timestamps to obtain the aligned temperature feature sub-matrix, environmental feature sub-matrix, and electrochemical feature sub-matrix.

[0148] In this embodiment, firstly, time dimension alignment is achieved through time synchronization processing based on the temperature feature sub-matrix, environmental feature sub-matrix, electrochemical feature sub-matrix, and corresponding timestamp information. Specifically, based on the timestamp information of each sub-matrix, the time format and time zone settings are unified through string parsing and format conversion methods to obtain standardized timestamps. Based on the standardized timestamps, a common time reference point and sampling interval are selected through least common multiple calculation and time window division methods to obtain a unified time reference. Based on each sub-matrix and the unified time reference, data with different sampling frequencies are interpolated onto the unified time reference through linear interpolation and spline interpolation methods to obtain a preliminary time-aligned matrix.

[0149] Then, based on the preliminary time-aligned matrix and spatial location information, spatial registration is performed to achieve spatial dimension alignment. Specifically, based on the spatial coordinates of each sub-matrix, the coordinate systems of different sensors are transformed to a unified spatial reference system using coordinate transformation matrices and affine transformation methods, resulting in a unified coordinate system. Based on each sub-matrix and the unified coordinate system, data with different spatial resolutions are interpolated onto a unified spatial grid using bilinear interpolation and Kriging interpolation methods, resulting in a spatially aligned intermediate matrix. Based on the spatially aligned intermediate matrix, the accuracy of spatial registration is verified using root mean square error calculation and correlation analysis methods, yielding a registration quality assessment result. Based on the registration quality assessment result, the spatially aligned intermediate matrix is ​​finely adjusted using an iterative nearest-point algorithm and a weighted average method, resulting in aligned temperature feature sub-matrices, environmental feature sub-matrices, and electrochemical feature sub-matrices.

[0150] S4.5: The aligned temperature feature submatrix, environmental feature submatrix, and electrochemical feature submatrix are fused through tensor decomposition and reconstruction to construct a multi-source state feature matrix.

[0151] In this embodiment, the aligned temperature feature submatrix, environmental feature submatrix, and electrochemical feature submatrix are first processed through multidimensional data organization to construct a tensor structure. Specifically, based on each submatrix, the two-dimensional feature matrix is ​​expanded into a three-dimensional tensor structure through matrix reshaping and dimension expansion methods, resulting in a three-dimensional feature tensor. Based on the three-dimensional feature tensor, tensors from different data sources are stacked along the third dimension through array concatenation and axial stacking methods, resulting in a multi-source data tensor. Based on the multi-source data tensor, the numerical range and dimensions of different data sources are unified through standardization and min-max normalization methods, resulting in a standardized multi-source tensor.

[0152] Furthermore, based on the standardized multi-source tensor, tensor decomposition is achieved through low-rank decomposition. Specifically, based on the standardized multi-source tensor, the high-dimensional tensor is decomposed into the outer product of multiple low-dimensional factor matrices using a multivariate decomposition method, resulting in a set of factor matrices. Based on the standardized multi-source tensor, the core structure and main components of each mode are extracted using a decomposition method, resulting in a core tensor and a mode matrix. Based on the set of factor matrices, the core tensor, and the mode matrix, the reconstruction error and feature preservation of the decomposition are evaluated using Frobenius norm calculation and relative reconstruction error analysis, resulting in a decomposition quality index.

[0153] Finally, based on the factor matrix set, core tensor, pattern matrix, and decomposition quality index, tensor reconstruction is achieved through feature fusion processing. Specifically, based on the decomposition quality index, the fusion weights of each data source are determined using gradient descent optimization and cross-validation methods to obtain a fusion weight vector. Based on the factor matrix set, core tensor, and fusion weight vector, the fused feature representation is reconstructed using weighted tensor multiplication and Kronecker product operations to obtain a fused feature tensor. Based on the fused feature tensor, the high-dimensional fused features are projected onto a suitable low-dimensional space using principal component analysis and singular value decomposition methods to obtain a multi-source state feature matrix. This matrix integrates state information from three dimensions: temperature, environment, and electrochemistry, providing a comprehensive feature foundation for subsequent intelligent monitoring.

[0154] The specific implementation steps of S5 are as follows:

[0155] S5.1: Based on the impedance amplitude and phase angle parameters at the preset characteristic frequency point in the impedance parameters, a nonlinear dimensionality reduction algorithm is used to perform dimensionality reduction processing to obtain the impedance characteristic feature vector.

[0156] In this embodiment, impedance extraction at characteristic frequencies is achieved through frequency domain feature selection processing based on impedance parameters. Specifically, a set of characteristic frequencies is obtained by selecting preset characteristic frequencies using a bandpass filtering method based on the impedance parameters; impedance amplitude and phase angle parameters are extracted using a fast Fourier transform method based on the characteristic frequency set to obtain impedance amplitude vector and phase angle vector; and complex impedance features are formed by combining the impedance amplitude vector and phase angle vector using a complex number construction method to obtain the original impedance feature vector.

[0157] Then, based on the original impedance feature vector, dimensionality compression is achieved through nonlinear dimensionality reduction. Specifically, based on the original impedance feature vector, the nonlinear manifold structure is identified using an equidistant mapping method to obtain manifold embedding parameters; the high-dimensional impedance features are mapped to a low-dimensional space using a t-distributed random neighborhood embedding method to obtain the dimensionality-reduced impedance features; based on the dimensionality-reduced impedance features, the dimensionality reduction quality is evaluated using a reconstruction error calculation method to obtain a dimensionality reduction quality index; based on the dimensionality reduction quality index, the dimensionality-reduced features are adjusted using a gradient optimization method to obtain the impedance characteristic feature vector.

[0158] S5.2: Based on the amplitude and phase information of the harmonic frequency components in the nonlinear distortion components, feature extraction is performed using principal component analysis to obtain the nonlinear response feature vector.

[0159] In this embodiment, harmonic features are first extracted based on the nonlinear distortion components through frequency domain decomposition. Specifically, the frequency components of each harmonic are identified using Fourier transform based on the nonlinear distortion components; the amplitude and phase information of each harmonic are extracted using amplitude detection and phase detection methods to obtain harmonic amplitude vectors and harmonic phase vectors; and a harmonic feature matrix is ​​formed based on the harmonic amplitude vectors and harmonic phase vectors using matrix construction methods.

[0160] Then, based on the harmonic characteristic matrix, dimensionality reduction is achieved through principal component analysis. Specifically, based on the harmonic characteristic matrix, the dimensional differences between different harmonic components are eliminated using the Z-score normalization method to obtain the standardized harmonic characteristic matrix; principal component parameters are obtained through covariance calculation and eigenvalue decomposition; and the number of principal components to be retained is selected based on the cumulative variance contribution rate to obtain the principal component selection results.

[0161] Finally, based on the standardized harmonic characteristic matrix and the principal component selection results, feature projection is achieved through principal component projection processing. Specifically, based on the principal component selection results, a PCA projection matrix is ​​constructed using matrix transformation methods; based on the standardized harmonic characteristic matrix and the PCA projection matrix, the original harmonic features are projected onto the principal component space through matrix multiplication operations to obtain the principal component scores; based on the principal component scores, the contribution of each principal component to the nonlinear response is evaluated using contribution analysis methods to obtain the nonlinear response feature vector.

[0162] S5.3: Based on the polarization stress coefficient, kinetic response coefficient and temperature compensation coefficient in the electrochemical stress factor, the electrochemical activity feature vector is obtained through mean normalization.

[0163] In this embodiment, coefficients are extracted through coefficient separation based on the electrochemical stress factor. Specifically, based on the electrochemical stress factor, the polarization stress coefficient, kinetic response coefficient, and temperature compensation coefficient are separated using a numerical decomposition method, resulting in three independent coefficient vectors. Based on these three coefficient vectors, the rationality of each coefficient is verified through numerical range testing and physical constraint testing, yielding an effective set of coefficient vectors.

[0164] Furthermore, based on the set of effective coefficient vectors, mean normalization is achieved through numerical standardization. Specifically, the mean and standard deviation of each coefficient vector are calculated using statistical methods based on the polarization stress coefficient, kinetic response coefficient, and temperature compensation coefficient. Based on each coefficient vector and its statistical parameters, each coefficient value is converted to a zero-mean, unit-variance form using Z-score normalization to obtain a normalized coefficient vector. Finally, based on the three normalized coefficient vectors, a unified feature representation is formed through vector concatenation, resulting in an electrochemical activity feature vector.

[0165] S5.4: Based on the nonlinear response feature vector, electrochemical activity feature vector, and impedance characteristic feature vector, a weighted fusion is performed using a feature fusion algorithm based on a self-attention mechanism to obtain the fused feature vector.

[0166] In this embodiment, feature vector preprocessing is achieved through feature alignment based on the nonlinear response feature vector, electrochemical activity feature vector, and impedance characteristic feature vector. Specifically, the three feature vectors are processed using a dimension unification algorithm, which zero-padding or truncation is used to make them have the same dimension, resulting in a unified-dimensional feature vector; the unified-dimensional feature vectors are then processed using a feature standardization algorithm, which standardizes feature vectors from different sources to the same numerical range, resulting in a standardized feature vector set; finally, the standardized feature vector set is processed using a feature matrix construction algorithm, which arranges the three feature vectors in rows to form a feature matrix, resulting in the input feature matrix.

[0167] Furthermore, based on the input feature matrix, self-attention computation is achieved through attention weight calculation. Specifically, based on the input feature matrix, a query matrix Q, a key matrix K, and a value matrix V are generated using a linear transformation method; based on the query matrix Q and the key matrix K, the similarity score between the query and the key is calculated through dot product operation to obtain the original attention score matrix; based on the original attention score matrix, the score is converted into a probability distribution using the Softmax function to obtain the normalized attention weight matrix; based on the normalized attention weight matrix and the value matrix V, the weighted feature representation is calculated through matrix multiplication to obtain the self-attention output matrix.

[0168] Finally, feature fusion is achieved through multi-feature integration processing based on the self-attention output matrix. Specifically, the importance weights of each feature component are calculated using analysis of variance based on the self-attention output matrix to obtain a feature importance vector; the features are combined using a weighted average method based on the self-attention output matrix and the feature importance vector to obtain a preliminary fused feature vector; the fusion effect is evaluated using information entropy calculation and correlation analysis based on the preliminary fused feature vector, and fine-tuning is performed using gradient optimization to obtain the final fused feature vector.

[0169] S5.5: Based on the fused feature vector, temperature compensation processing is performed by combining the surface temperature distribution data and ambient temperature data to form a multi-dimensional feature vector adapted to low-temperature environments.

[0170] In this embodiment, temperature compensation for the fused feature vectors is achieved through temperature robustness enhancement processing. This temperature compensation strategy effectively eliminates the impact of temperature changes on the stability of the feature vectors, significantly improves the system's feature representation capability and monitoring accuracy in low-temperature environments, and provides a reliable feature foundation for subsequent intelligent analysis.

[0171] The specific implementation steps of S5.5 are as follows:

[0172] S5.5.1: Based on the surface temperature distribution data, ambient temperature data, and fused feature vector, the temperature sensitivity coefficient of each feature component is calculated by constructing a mapping relationship between temperature and each feature component in the fused feature vector, and the feature components with temperature sensitivity coefficients greater than a preset sensitivity threshold are determined as feature components to be compensated.

[0173] In this embodiment, a temperature-feature mapping relationship is constructed based on surface temperature distribution data, ambient temperature data, and the fused feature vector. Specifically, based on the surface temperature distribution data and ambient temperature data, spatial distribution features and temporal variation features are extracted using temperature data preprocessing methods to obtain a comprehensive temperature feature vector. The fused feature vector is then decomposed into independent feature components using a feature component analysis method. A function mapping relationship is established based on the comprehensive temperature feature vector and each feature component using a mapping function construction method. The effectiveness and stability of the mapping relationship are then verified using a mapping relationship verification method, resulting in a verified mapping relationship model.

[0174] Furthermore, based on the validated mapping model, the sensitivity is quantified by calculating the temperature sensitivity coefficient. Specifically, based on the validated mapping model, the partial derivatives of each feature component with respect to temperature change are calculated using the partial derivative calculation method to obtain the instantaneous sensitivity vector; based on the instantaneous sensitivity vector, statistical characteristics within different temperature ranges are calculated using sensitivity statistics methods to obtain sensitivity statistical parameters; based on the sensitivity statistical parameters, the temperature sensitivity coefficient of each feature component is calculated by comprehensively considering the mean, variance, and extreme values ​​using the sensitivity coefficient generation method.

[0175] Finally, the feature components to be compensated are determined based on the temperature sensitivity coefficient. Specifically, a preset sensitivity threshold is determined based on the statistical distribution characteristics of the sensitivity coefficient. The temperature sensitivity coefficient of each feature component is compared with the preset sensitivity threshold one by one, and feature components with temperature sensitivity coefficients greater than the preset sensitivity threshold are identified as high-sensitivity feature components. Based on the high-sensitivity feature components, the necessity and feasibility of temperature compensation for each feature component are evaluated using a compensation necessity assessment method, and the compensation necessity assessment result is obtained. Based on the compensation necessity assessment result, the high-sensitivity feature components are screened to determine the feature components to be compensated.

[0176] S5.5.2: For the characteristic components to be compensated originating from the electrochemical activity characteristic vector, the compensation value is calculated using an exponential temperature compensation function; for the characteristic components to be compensated originating from the impedance characteristic vector, the compensation value is calculated using a piecewise linear temperature compensation function; and for the characteristic components to be compensated originating from the nonlinear response characteristic vector, the compensation value is calculated using a polynomial temperature compensation function.

[0177] In this embodiment, the source of the feature components is identified based on the composition information of the feature components to be compensated and the fused feature vector. Specifically, based on the feature components to be compensated, the original source of each feature component is analyzed inversely using matrix factorization based on the transformation matrix in the feature fusion process to obtain the feature source identifier; based on the feature source identifier, the correlation coefficient between the feature component and the three types of original features (electrochemical activity, impedance characteristics, and nonlinear response) is calculated using correlation analysis to identify the specific source type of each feature component to be compensated; based on the classification results, the classification accuracy is verified using cross-validation to obtain the verified classification result.

[0178] Furthermore, based on the validated classification results and the comprehensive temperature feature vector, a differentiated compensation function is used to perform classification compensation calculations. Specifically, an exponential compensation function construction algorithm is used to process the feature components to be compensated originating from the electrochemical activity feature vector. Based on the Arrhenius relation of the electrochemical reaction, an exponential temperature compensation function is used to calculate the compensation value, thus obtaining the electrochemical activity compensation value. A piecewise linear compensation function construction algorithm is used to process the feature components to be compensated originating from the impedance characteristic feature vector. Based on the piecewise characteristics of impedance changing with temperature, a piecewise linear temperature compensation function is used to calculate the compensation value. Different linear compensation strategies are used in different temperature ranges to obtain the impedance characteristic compensation value. A polynomial compensation function construction algorithm is used to process the feature components to be compensated originating from the nonlinear response feature vector. Based on the complex relationship between nonlinear response and temperature, a polynomial temperature compensation function is used to calculate the compensation value, capturing the higher-order influence of temperature on nonlinear features, thus obtaining the nonlinear response compensation value.

[0179] Finally, based on various compensation values, adaptive parameter adjustment is achieved through compensation function parameter optimization. Specifically, a historical data fitting algorithm is used to process various compensation values, and the parameters of the compensation function are optimized using historical temperature and feature data to obtain optimized compensation parameters. A compensation effect prediction algorithm is then used to process the optimized compensation parameters to predict the performance of the compensated feature vector at different temperatures, yielding a compensation effect prediction result. Finally, a parameter adjustment algorithm is used to fine-tune the optimized compensation parameters based on the compensation effect prediction result, ensuring optimal compensation effect and obtaining the final compensation value.

[0180] S5.5.3: Perform temperature compensation on the feature component to be compensated according to the compensation value, update the corresponding feature component in the fused feature vector, and generate a multi-dimensional feature vector with temperature robustness.

[0181] In this embodiment, real-time compensation is performed based on the final compensation value and the current temperature conditions. Specifically, the current temperature state vector is obtained based on the surface temperature distribution data and ambient temperature data; the electrochemical activity compensation value, impedance characteristic compensation value, and nonlinear response compensation value are applied to the corresponding feature components to be compensated, and temperature compensation is performed on the feature components to be compensated to obtain the compensated feature components; the compensated feature components are processed to update the corresponding feature components in the fused feature vector to obtain the updated feature vector.

[0182] Furthermore, robustness assessment is performed based on the updated eigenvectors. Specifically, based on the updated eigenvectors, the performance of the eigenvectors under different temperature conditions is simulated using Monte Carlo simulation to obtain temperature response characteristics; based on the temperature response characteristics, the stability index of the eigenvectors under temperature changes is calculated using analysis of variance to obtain robustness assessment results; based on the robustness assessment results, the updated eigenvectors are optimized and adjusted using gradient descent to obtain optimized temperature-compensated eigenvectors.

[0183] Finally, based on the optimized temperature-compensated feature vector, the final feature vector is constructed through multi-dimensional feature vector generation. Specifically, based on the optimized temperature-compensated feature vector, the integrity and consistency of the feature vector are verified using an integrity check method; based on the verified feature vector, the numerical range and distribution characteristics of the feature vector are unified using a normalization method to obtain a standardized feature vector; based on the standardized feature vector, its applicability in low-temperature environments is evaluated using a cross-validation method to generate a temperature-robust multi-dimensional feature vector.

[0184] The specific implementation steps of S6 are as follows:

[0185] S6.1: Based on the multi-source state feature matrix, extract temporal dependence and spatial correlation through a spatiotemporal convolutional neural network to obtain state feature representation and generate the first decision result and its confidence level;

[0186] In this embodiment, firstly, the network architecture is initialized by constructing a spatiotemporal convolutional network based on the multi-source state feature matrix. Specifically, based on the dimension and feature distribution of the multi-source state feature matrix, a network architecture including temporal convolutional layers, spatial convolutional layers, and feature fusion layers is designed using a multi-level network topology construction algorithm to obtain the network architecture configuration. Based on the network architecture configuration, an adaptive weight initialization algorithm is used to calculate the initial variance based on the input-output dimension ratio and set the parameters of each layer to obtain the initialized network model. Based on the initialized network model, a hierarchical activation function adaptive configuration algorithm is used to configure the ReLU activation function for the temporal convolutional layers and the LeakyReLU activation function for the spatial convolutional layers to obtain the complete spatiotemporal convolutional network.

[0187] Then, based on the multi-source state feature matrix, temporal dependency extraction is performed to learn the temporal dimension dependency. Specifically, based on the temporal sequence dimension of the multi-source state feature matrix, temporal features are extracted in parallel using multi-scale one-dimensional convolutional kernels to obtain multi-scale temporal dependencies. Based on the multi-scale temporal dependencies, importance weights for each time step are calculated using an attention mechanism, and the weights are weighted and summed with the corresponding temporal features to obtain weighted temporal dependencies. Based on the weighted temporal dependencies, a bidirectional LSTM network is used to model the forward and backward temporal information respectively, and a gating mechanism is used to control the information flow to capture long-term and short-term dependencies, thus obtaining the complete temporal dependencies.

[0188] Furthermore, based on the multi-source state feature matrix, spatial correlations are extracted to learn spatial dimensional relationships. Specifically, based on the sensor spatial layout dimension of the multi-source state feature matrix, local and global spatial correlation features are extracted through multi-scale two-dimensional convolution operations to obtain multi-level spatial correlations. Based on the multi-level spatial correlations, importance weights of each sensor position are calculated through an attention mechanism and weighted fusion is performed to obtain enhanced spatial correlations. Based on the enhanced spatial correlations, topological relationships between sensors are modeled through graph convolution operations and neighborhood feature information is aggregated to obtain aggregated spatial correlations.

[0189] Next, based on temporal dependence and spatial correlation, spatiotemporal joint modeling is achieved through spatiotemporal feature fusion. Specifically, based on temporal dependence and spatial correlation, the dimensions and scales of the two types of features are unified through dimensionality transformation and normalization operations to obtain aligned feature pairs. Based on the aligned feature pairs, the interaction relationship between spatiotemporal features is modeled through separable convolution and cross-attention mechanisms to obtain spatiotemporal interactive features. Based on the spatiotemporal interactive features, spatiotemporal information is fused through residual connections and feature concatenation operations to obtain fused spatiotemporal features. Based on the fused spatiotemporal features, high-level feature abstraction processing is performed through state feature representation. Specifically, based on the fused spatiotemporal features, feature representations are further abstracted through a multi-layer fully connected network to obtain abstract feature representations. Based on the abstract feature representations, the feature dimensionality is reduced through principal component analysis and linear discriminant analysis to obtain compact feature representations. Based on the compact feature representations, the numerical range and distribution characteristics of the features are unified through numerical normalization operations to obtain standardized state feature representations.

[0190] Finally, based on the state feature representation, the probability distribution of each state category is output through the Softmax classifier to obtain the state classification probability; based on the state classification probability, the final state category is determined by the maximum probability principle to obtain the first decision result; based on the state classification probability, the confidence of the decision is calculated by the entropy value and the maximum probability value to obtain the confidence of the first decision result.

[0191] S6.2: Based on the multidimensional feature vector, extract nonlinear feature relationships through a deep fully connected network to obtain parametric feature representations and generate a second decision result and its confidence level;

[0192] In this embodiment, firstly, a deep fully connected network is constructed to extract nonlinear feature relationships from the multidimensional feature vectors. Specifically, based on the dimensionality information of the multidimensional feature vectors, an input layer, multiple hidden layers, and an output layer are designed through a multi-layer fully connected network to obtain a network hierarchical structure. Based on the network hierarchical structure, a decreasing configuration strategy is used to configure the number of neurons in each hidden layer to obtain a neuron configuration scheme. Based on the neuron configuration scheme, a deep fully connected network is constructed by introducing Dropout layers and batch normalization layers. Based on the deep fully connected network, multi-layer nonlinear transformations are performed, with each layer using the ReLU activation function to enhance the nonlinear modeling capability, resulting in hierarchical nonlinear feature relationships. Based on the hierarchical nonlinear feature relationships, an attention mechanism is used to model the interaction relationships between features, resulting in interaction-enhanced nonlinear feature relationships. Based on the interaction-enhanced nonlinear feature relationships, residual connections are used to alleviate the gradient vanishing problem, resulting in nonlinear feature relationships.

[0193] Next, based on the nonlinear feature relationship, multi-layer feature information is aggregated by combining weighted averaging and max pooling to obtain aggregated feature representation; based on aggregated feature representation, low variance and redundant features are removed by variance thresholding and mutual information calculation to obtain selected feature representation; based on selected feature representation, parametric feature representation is obtained by autoencoder compression and abstract feature representation.

[0194] Then, based on the parameter feature representation, a decision space mapping process is performed using a decision mapping method. Specifically, based on the parameter feature representation, the features are mapped to the class probability distribution through a fully connected layer and a Softmax activation function to obtain the initial probability output; based on the initial probability output, the confidence of the probability distribution is adjusted using a temperature scaling method to obtain a calibrated probability model; based on the calibrated probability model, the class index corresponding to the highest probability is selected through an argmax operation to obtain the class mapping result.

[0195] Finally, based on the category mapping results, the predicted category label is determined by selecting the maximum probability value to obtain the second decision result; based on the category mapping results, the degree of uncertainty of the prediction is calculated by the probability distribution entropy to obtain the uncertainty measure; based on the second decision result and the uncertainty measure, the prediction confidence is calculated by the difference between the maximum probability value and the second largest probability value to obtain the confidence level corresponding to the second decision result.

[0196] S6.3: Based on the first decision result and its confidence level and the second decision result and its confidence level, a decision-level fusion is achieved through a weighted voting mechanism, combined with a preset confidence level threshold and decision rules, to obtain the status monitoring result;

[0197] In this embodiment, firstly, decision information standardization processing is performed based on the first decision result and its confidence level, and the second decision result and its confidence level. Specifically, based on the two decision results, category labels of different formats are converted into unified codes through category label unification to obtain aligned decision results; based on the confidence levels of the two decision results, the confidence level values ​​are mapped to the [0,1] interval through min-max normalization to obtain normalized confidence levels; based on the aligned decision results and normalized confidence levels, the category consistency and confidence level difference of the two decision results are calculated through consistency checks to obtain decision quality indicators.

[0198] Then, based on the normalized confidence level and decision quality index, a dynamic weight allocation is performed through a weighted voting mechanism. Specifically, the normalized confidence level is processed using a confidence weight calculation algorithm, and the basic weight of each decision result is calculated based on the magnitude and distribution characteristics of the confidence level to obtain the confidence weight; the quality weight is obtained by calculating the quality adjustment weight based on the historical accuracy and stability of the decision results; the confidence weight and the quality weight are combined by weighted averaging to obtain the final voting weight; the final voting weight is then normalized to obtain the normalized voting weight.

[0199] Furthermore, based on expert experience and domain knowledge, pre-set reliability thresholds and pre-defined decision rules are determined. By comparing the normalized confidence scores of each decision outcome with the pre-set reliability thresholds, normalized confidence scores greater than the pre-set reliability thresholds are considered high-quality decision outcomes, resulting in a high-confidence decision set. For decision outcomes with normalized confidence scores lower than the pre-set reliability thresholds, their weights in the fusion process are reduced, resulting in adjusted decision weights. All decision outcomes are then classified according to the pre-set reliability thresholds, and different processing strategies and fusion rules are set to obtain a tiered decision strategy.

[0200] Next, a weighted voting mechanism is used, with normalized voting weights, high-confidence decision sets, and hierarchical decision strategies as input data, to perform multi-decision fusion processing. Specifically, a weighted probability calculation algorithm is used to process each decision result, calculating the weighted probability of each state category based on the normalized voting weights to obtain a fusion probability distribution. Based on the fusion probability distribution, the final state category is determined through preset decision rules, including the maximum probability rule, threshold judgment rule, and consistency check rule, to obtain a preliminary fusion result. When conflicts exist between different decision results in the preliminary fusion result, evidence theory or fuzzy logic methods are used to resolve the conflicts, resulting in a consistent fusion result.

[0201] Subsequently, based on the consistency fusion results and preset decision rules, the applicable set of decision rules is determined by pattern matching, resulting in an applicable rule set; based on the applicable rule set, rule conflicts are identified by logical consistency checks, resulting in conflict detection results; based on rule priorities, the applicable rules are sorted by priority ranking, resulting in a sorted rule set; based on the sorted rule set, the consistency fusion results are constrained and adjusted by rule application, resulting in a rule-constrained decision result.

[0202] Finally, based on the decision results constrained by the rules, the rationality of the decision is checked through consistency verification to obtain the verified decision results; based on the uncertainty propagation in the fusion process, the decision confidence is recalculated through confidence update to obtain the final confidence; based on the verified decision results and the final confidence, a complete fusion result containing state category, confidence and decision basis is generated through result encapsulation to obtain the state monitoring results.

[0203] This application also provides a low-temperature lithium battery BMS monitoring system based on collaborative sensing, including a data acquisition module, a data processing module, a data analysis module, and a monitoring decision module, wherein:

[0204] The data acquisition module is used to acquire surface temperature distribution data, ambient temperature data and electrochemical parameters of the low-temperature resistant lithium battery, as well as to apply a composite current excitation signal to the low-temperature resistant lithium battery and simultaneously acquire the voltage response signal.

[0205] The data processing module is used to preprocess the data acquired by the data acquisition module, obtain the impedance parameters at a preset characteristic frequency point through impedance analysis, extract the nonlinear distortion component through spectrum analysis, obtain the relaxation voltage curve, and extract the electrochemical stress factor.

[0206] The data analysis module is used to construct a multi-source state feature matrix and a multi-dimensional feature vector, and to perform decision-level fusion analysis through deep learning algorithms;

[0207] The monitoring decision module is used to generate status monitoring results based on the analysis results output by the data analysis module.

[0208] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0209] 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, as well as 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. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0210] 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.

[0211] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A low-temperature lithium battery BMS monitoring method based on collaborative sensing, characterized in that, include: Collect surface temperature distribution data, ambient temperature data, and electrochemical parameters of low-temperature resistant lithium batteries; A composite current excitation signal is applied to the low-temperature resistant lithium battery, and the voltage response signal of the low-temperature resistant lithium battery is acquired simultaneously. The impedance parameters at a preset characteristic frequency point are obtained through impedance analysis. Based on the voltage response signal, the nonlinear distortion component is extracted by spectrum analysis to obtain the relaxation voltage curve. The electrochemical stress factor is obtained by differentiating the relaxation voltage curve. Based on the surface temperature distribution data, ambient temperature data, and electrochemical parameters, a multi-source state feature matrix is ​​constructed through spatiotemporal alignment and data fusion. Based on the impedance parameters, nonlinear distortion components, and electrochemical stress factors, a multidimensional feature vector is formed through feature fusion. Based on the multi-source state feature matrix and multi-dimensional feature vector, a decision-level fusion is performed using a deep learning algorithm to obtain the state monitoring results. The composite current excitation signal includes a fundamental frequency excitation signal, a high-frequency disturbance signal, and a preset current interruption sequence, wherein: The fundamental frequency excitation signal has a preset amplitude and frequency, and is used to excite the electrochemical response of the low-temperature resistant lithium battery; The amplitude of the high-frequency disturbance signal is smaller than that of the fundamental frequency excitation signal, and the frequency is greater than that of the fundamental frequency excitation signal, which is used to obtain the impedance parameters of the low-temperature resistant lithium battery at a preset characteristic frequency point. The preset current interruption sequence includes multiple current interruption points of preset duration, used to obtain the relaxation voltage of the low-temperature resistant lithium battery. The frequency range of the composite current excitation signal covers the characteristic spectrum of low-temperature resistant lithium batteries from the diffusion process to the charge transfer process under low-temperature conditions, and the temperature range of the low-temperature environment is determined based on industry standards. The step of extracting the nonlinear distortion component through spectral analysis based on the voltage response signal to obtain the relaxation voltage curve includes: Based on the voltage response signal, the fundamental frequency component and its harmonic frequency components are obtained through adaptive wavelet transform and short-time Fourier transform. Based on the harmonic frequency components, the amplitude and phase information of the harmonic frequency components are extracted using spectral analysis methods to construct nonlinear distortion components; Based on the nonlinear distortion components, a mapping model between electrochemical parameters and voltage response is constructed using a piecewise linear and nonlinear hybrid fitting method. Based on the mapping relationship model, the relaxation voltage curve is obtained by extracting the transient response characteristics of the voltage response signal to the preset current interruption sequence. The electrochemical stress factor is obtained by differentiating the relaxation voltage curve, including: Based on the relaxation voltage curve, the voltage change rate curve, voltage acceleration curve, and voltage jump characteristic curve are obtained through multi-order differential processing. Identify the extreme points, inflection points, and steady-state points of each curve from the voltage change rate curve, voltage acceleration curve, and voltage jump characteristic curve, and calculate the time constant and amplitude parameters corresponding to each characteristic point; Based on the time constant and amplitude parameters, an equivalent circuit model of a lithium battery is constructed. By mapping the time constant to an electrochemical reaction rate constant and the amplitude parameters to polarization impedance parameters, and combining the temperature compensation function, the correlation between the equivalent circuit model and the electrochemical kinetic characteristics of a low-temperature resistant lithium battery under low-temperature conditions is established. Based on the polarization impedance parameters, electrochemical reaction rate constant, and correlation, the polarization stress coefficient, kinetic response coefficient, and temperature compensation coefficient are calculated respectively, and combined to form the electrochemical stress factor.

2. The low-temperature lithium battery BMS monitoring method based on collaborative sensing according to claim 1, characterized in that, The process of obtaining impedance parameters at a preset characteristic frequency point through impedance analysis includes: Based on the composite current excitation signal and voltage response signal, Fourier transform is used to obtain current frequency domain response data and voltage frequency domain response data, respectively. Based on the ratio of the voltage frequency domain response data to the current frequency domain response data, the complex impedance value at each frequency point is calculated. Based on the complex impedance values ​​at each frequency point, the impedance amplitude and phase angle parameters at a preset characteristic frequency point are extracted to obtain the impedance parameters. The preset characteristic frequency point is determined based on the impedance response characteristics of the low-temperature resistant lithium battery at different frequencies.

3. The low-temperature lithium battery BMS monitoring method based on collaborative sensing according to claim 1, characterized in that, The process of forming a multi-dimensional feature vector through feature fusion based on the impedance parameter, nonlinear distortion component, and electrochemical stress factor includes: Based on the impedance amplitude and phase angle parameters at preset characteristic frequencies in the impedance parameters, a nonlinear dimensionality reduction algorithm is used to perform dimensionality reduction processing to obtain the impedance characteristic feature vector. Based on the amplitude and phase information of the harmonic frequency components in the nonlinear distortion components, feature extraction is performed using principal component analysis to obtain the nonlinear response feature vector. Based on the polarization stress coefficient, kinetic response coefficient, and temperature compensation coefficient in the electrochemical stress factor, an electrochemical activity feature vector is obtained through mean normalization. Based on the nonlinear response feature vector, electrochemical activity feature vector, and impedance characteristic feature vector, a weighted fusion is performed using a feature fusion algorithm based on a self-attention mechanism to obtain the fused feature vector. Based on the fused feature vector, temperature compensation processing is performed by combining the surface temperature distribution data and ambient temperature data to form a multi-dimensional feature vector adapted to low-temperature environments.

4. The low-temperature lithium battery BMS monitoring method based on collaborative sensing according to claim 3, characterized in that, The step involves performing temperature compensation processing based on the fused feature vector, combined with the surface temperature distribution data and ambient temperature data, to form a multi-dimensional feature vector adapted to low-temperature environments, including: Based on the surface temperature distribution data, ambient temperature data, and fused feature vector, the temperature sensitivity coefficient of each feature component is calculated by constructing a mapping relationship between temperature and each feature component in the fused feature vector. Feature components with temperature sensitivity coefficients greater than a preset sensitivity threshold are identified as feature components to be compensated. For the characteristic components to be compensated originating from the electrochemical activity characteristic vector, the compensation value is calculated using an exponential temperature compensation function; for the characteristic components to be compensated originating from the impedance characteristic vector, the compensation value is calculated using a piecewise linear temperature compensation function; and for the characteristic components to be compensated originating from the nonlinear response characteristic vector, the compensation value is calculated using a polynomial temperature compensation function. Temperature compensation is performed on the feature components to be compensated based on the compensation value, and the corresponding feature components in the fused feature vector are updated to generate a multidimensional feature vector.

5. The low-temperature lithium battery BMS monitoring method based on collaborative sensing according to claim 1, characterized in that, The process involves constructing a multi-source state feature matrix based on the surface temperature distribution data, ambient temperature data, and electrochemical parameters through spatiotemporal alignment and data fusion, including: Based on the surface temperature distribution data, temperature gradient, hot spot distribution, and temperature uniformity indicators are extracted through thermal imaging processing to form a temperature feature sub-matrix. Based on the ambient temperature data, time-series analysis is used to extract the ambient temperature change rate, temperature fluctuation amplitude, and temperature stability index to form an environmental feature sub-matrix. Based on the electrochemical parameters, the state of charge, internal resistance change trend and capacity decay characteristics are extracted through incremental capacity analysis and electrochemical feature engineering to form an electrochemical feature sub-matrix. The temperature feature sub-matrix, environmental feature sub-matrix, and electrochemical feature sub-matrix are aligned in time and space using a timestamp-based temporal alignment algorithm and a spatial registration algorithm to obtain aligned temperature feature sub-matrix, environmental feature sub-matrix, and electrochemical feature sub-matrix. The aligned temperature feature submatrix, environmental feature submatrix, and electrochemical feature submatrix are fused through tensor decomposition and reconstruction to construct a multi-source state feature matrix.

6. The low-temperature lithium battery BMS monitoring method based on collaborative sensing according to claim 1, characterized in that, The step of obtaining state monitoring results by performing decision-level fusion using a deep learning algorithm based on the multi-source state feature matrix and multi-dimensional feature vectors includes: Based on the multi-source state feature matrix, temporal dependence and spatial correlation are extracted through a spatiotemporal convolutional neural network to obtain state feature representations and generate a first decision result and its confidence level. Based on the multidimensional feature vector, a deep fully connected network is used to extract nonlinear feature relationships, obtain parametric feature representations, and generate a second decision result and its confidence level. Based on the first decision result and its confidence level, and the second decision result and its confidence level, a weighted voting mechanism is used, combined with a preset confidence level threshold and decision rules, to achieve decision-level fusion and obtain the status monitoring results.

7. A low-temperature lithium battery BMS monitoring system based on collaborative sensing, used to implement the low-temperature lithium battery BMS monitoring method based on collaborative sensing as described in any one of claims 1-6, characterized in that, It includes a data acquisition module, a data processing module, a data analysis module, and a monitoring and decision-making module, among which: The data acquisition module is used to acquire surface temperature distribution data, ambient temperature data and electrochemical parameters of the low-temperature resistant lithium battery, as well as to apply a composite current excitation signal to the low-temperature resistant lithium battery and simultaneously acquire the voltage response signal. The data processing module is used to preprocess the data acquired by the data acquisition module, obtain the impedance parameters at a preset characteristic frequency point through impedance analysis, obtain the relaxation voltage curve and extract the electrochemical stress factor by extracting the nonlinear distortion component. The data analysis module is used to construct a multi-source state feature matrix and a multi-dimensional feature vector, and to perform decision-level fusion analysis through deep learning algorithms; The monitoring decision module is used to generate status monitoring results based on the analysis results output by the data analysis module.

Images