Motor train unit battery health state evaluation and life prediction method and system

Abstract

The invention relates to the technical field of motor train unit battery pack evaluation, and provides a motor train unit battery health state evaluation and life prediction method and system, and the method comprises the steps: building a database through voltage fluctuation and temperature data in a floating charge state, and analyzing a correlation mechanism between target parameters, such as a battery capacity retention ratio and active lithium stock, and internal chemical substance loss; respectively extracting driving behavior characteristics, temperature parameters and electrochemical impedance spectroscopy by adopting second-level, minute-level and hour-level time granularity, and aligning multi-frequency data by utilizing a dynamic time warping algorithm; constructing a graph structure model taking the electrochemical parameters as nodes and cross-level association as edges, and extracting a battery aging characteristic topological relation through a graph neural network; and finally, establishing a prediction model fused with multi-dimensional features, and realizing accurate prediction of the residual life and the capacity recession inflection point of the battery. And the multi-scale characteristics are combined with an electrochemical mechanism, so that the accuracy of battery health state evaluation under a complex working condition is improved.

Description

Technical Field

[0001] This application relates to the technical field of evaluation of EMU battery packs. Specifically, it relates to a method and system for evaluating the health status and predicting the life of EMU batteries. Background Technique

[0002] The content of this part only provides background information related to this application, and it may not constitute prior art.

[0003] In the field of EMU operation, as a key power source, the health status of EMU batteries is directly related to the operation safety and performance of EMUs. In the management and supervision of such public facilities as EMUs, accurately evaluating the health status of batteries and predicting their life is of great significance for ensuring the reliable operation of EMUs.

[0004] With the development of technology, some existing methods attempt to predict battery life. For example, the patent with the publication number CN113393064A proposes a method and terminal device for predicting the life of nickel-cadmium batteries of EMUs, including an algorithm that combines particle filtering and extended Kalman filtering, establishing a degradation model based on historical capacity data and discharge depth data of the battery, etc., and then realizing life prediction.

[0005] However, such methods have many limitations. On the one hand, the extraction of their characteristic data is not comprehensive enough, only focusing on limited characteristics and ignoring various key characteristic information of the battery at different time scales, such as the frequency of rapid acceleration or deceleration events during EMU operation, the fine changes in battery temperature data, the frequency domain characteristics of voltage signals and current signals, etc. This makes the evaluation of the battery health status lack sufficient comprehensiveness and accuracy. On the other hand, in the face of the complex and changeable driving modes and working conditions in the actual operation of EMUs, this method has insufficient adaptability for predicting battery life under different driving modes and is difficult to accurately reflect the differential impact of different working conditions on battery life.

[0006] Therefore, there is an urgent need for a method and system for evaluating the health status and predicting the life of EMU batteries to evaluate the battery pack more comprehensively, accurately and adaptively. Summary of the Invention

[0007] To solve the above technical problems, the purpose of this application is to provide a method and system for evaluating the health status and predicting the life of EMU batteries. By extracting characteristic data at different time levels to obtain various key characteristic information of the battery at different time scales, the comprehensiveness and accuracy of battery health status evaluation are improved; at the same time, by using the dynamic time warping algorithm to align data with different sampling frequencies and identify the driving mode of the EMU, the adaptability of battery life prediction under different driving modes is improved.

[0008] The purpose of this application is achieved through the following technical solutions: In a first aspect, the present invention provides a method for evaluating the health status and predicting the life of a multiple unit train battery, including: Based on the running state of the multiple unit train at a preset speed within a preset time under the floating charge state, collect the floating charge voltage fluctuation data and the multiple unit train battery temperature data, and establish a database; Through the data corresponding to the charge and discharge cycles in the database, monitor the target parameters of the battery. The target parameters include the capacity retention rate, the active lithium stock, and the phase change amount of the positive electrode material; establish an associated mapping relationship between the target parameters and the loss of internal chemical active substances in the battery, the deterioration of the electrode structure, and the interfacial side reaction, and obtain the aging mechanism information; Extract feature data at different time levels, including extracting the frequency of rapid acceleration or rapid deceleration events during the running of the multiple unit train at a second-level time granularity, using the dynamic time warping algorithm to align data with different sampling frequencies and identify the driving mode of the multiple unit train; extracting the temperature data of the multiple unit train battery at a minute-level time granularity; extracting the voltage signal and current signal of the multiple unit train at an hour-level time granularity; performing frequency domain analysis on the voltage signal and current signal to obtain the electrochemical impedance spectrum; Taking the aging mechanism information as a graph structure, where the nodes in the graph structure represent the electrochemical impedance spectrum or the feature data corresponding to different time levels, and the edges in the graph structure represent the association relationships between the feature data at different time levels; extract the graph topology features reflecting the battery aging state based on the graph structure, and establish a corresponding graph neural network model; Using the feature data as the input and the remaining life of the battery sample and the capacity degradation inflection point as the output, train the graph neural network model to obtain a prediction model; input the new voltage data, current data, temperature data, and electrochemical impedance spectrum into the prediction model to obtain the prediction result.

[0009] Further, the step of extracting the voltage signal and current signal of the multiple unit train at an hour-level time granularity specifically includes: Select a preset decomposition level, and perform multi-scale decomposition extraction on the voltage signal and current signal through the wavelet packet decomposition method.

[0010] Further, the step of using the dynamic time warping algorithm to align data with different sampling frequencies and identify the driving mode of the multiple unit train specifically includes: Obtain the test sequence and standard sequence of the driving mode of the multiple unit train; Construct the distance matrix of the test sequence and the standard sequence; Based on the distance matrix, find an alignment path from the starting point to the ending point of the matrix. The alignment path satisfies the preset boundary conditions, preset continuity conditions, and preset monotonic conditions. By accumulating the distances of each point on the path, the alignment distance is obtained. The preset boundary condition is that the alignment path starts from the starting point of the distance matrix and terminates at the other end of the diagonal of the distance matrix. The preset continuity condition is that only the matrix adjacent to the current path point can be selected for alignment. The preset monotonic condition is that the alignment path is monotonic on the time axis. If the alignment distance is less than or equal to the threshold, it indicates that the driving mode in the test sequence is similar to that in the standard sequence, and it is determined to be the same driving mode, and the driving mode corresponding to the standard sequence at this time is output. If the alignment distance is greater than the threshold, the driving mode corresponding to the test sequence is obtained by comparing the known driving modes in the standard sequence.

[0011] Furthermore, based on the graph structure, extract the graph topology features reflecting the battery aging state, specifically including: Taking the voltage, current, and electrochemical impedance spectrum of the battery as input parameters, representing the influence relationship of voltage changing the electrode polarization state on the electrochemical impedance spectrum through the Butler-Volmer equation, and obtaining the graph topology features.

[0012] Furthermore, the formula corresponding to the influence relationship is:

[0013] Among them, the boundary conditions are:

[0014] In the formula, 、 are the electronic current of the battery external circuit and the ionic current of the battery internal circuit respectively; 、 are the electronic potential of the battery external circuit and the ionic potential of the battery internal circuit respectively, 、 are the conductivities of the electrode solid skeleton and the electrolyte respectively; is the working current output by the electrode; is the electrode thickness; is the product of the specific surface area of the porous electrode and the exchange current density; is the transfer coefficient; is the number of electrons transferred by the electrode reaction; is a constant; is the corresponding position of the electrode.

[0015] Furthermore, after establishing the corresponding graph neural network model, it also includes: Based on the node features and adjacency relationships of the graph structure, embed a spatio-temporal attention module between each graph neural network layer. The spatio-temporal attention module includes: For any node, a temporal attention weight matrix and a spatial attention weight matrix are respectively generated according to the characteristic change trend of its adjacent nodes in the time series and the spatial topological connection strength; the temporal attention weight matrix and the spatial attention weight matrix are coupled to obtain a spatio-temporal fusion attention coefficient; based on the spatio-temporal fusion attention coefficient, the features of the adjacent nodes are weighted and averaged to update the feature representation of the current node.

[0016] Further, the capacity decay inflection point is obtained by the knee algorithm; the knee algorithm includes: Extract the time series data of the capacity retention rate during the full life cycle of the battery from the database, smooth the time series data to obtain a denoised aging trajectory curve; Connect the starting point and the ending point of the aging trajectory curve to generate a reference straight line, calculate the perpendicular distance from each sampling point on the curve to the reference straight line to form a distance sequence; Traverse the distance sequence and select the point with the maximum perpendicular distance as the capacity decay inflection point.

[0017] Further, after obtaining the prediction result, it also includes: Taking temperature as the acceleration stress and the thickening of the negative electrode SEI film as the main attenuation mechanism, describe the capacity decay law through the Arrhenius model to obtain the target model; According to the relative capacity decay amount of the battery at different cycle numbers, use the target model to calculate the capacity value of the battery after different cycle numbers.

[0018] Further, the formula corresponding to the target model is:

[0019] Among them, is the relative capacity decay amount after n cycles; is a constant greater than 0; is the activation energy; is the gas constant; is the absolute temperature; is the number of cycles; is the exponent; is the capacity value estimated by the model after n cycles of the battery; is the initial capacity of the battery.

[0020] In a second aspect, the present invention provides a battery health state evaluation and life prediction system for a multiple unit train, including: A data establishment module, based on the running state of the multiple unit train at a preset speed within a preset time under the floating charge state, collects floating charge voltage fluctuation data and multiple unit train battery temperature data, and establishes a database; An aging mechanism information acquisition module is used to monitor the target parameters of the battery through the data corresponding to charge and discharge cycles in the database. The target parameters include the capacity retention rate, the inventory of active lithium, and the phase change amount of the positive electrode material. An association mapping relationship between the target parameters and the loss of internal chemical active substances in the battery, the deterioration of the electrode structure, and the interfacial side reactions is established to obtain aging mechanism information. A characteristic data extraction module is used to extract characteristic data at different time levels, including extracting the frequency of sudden acceleration or deceleration events during the operation of the EMU at a second-level time granularity, aligning data with different sampling frequencies using the dynamic time warping algorithm and identifying the driving mode of the EMU; extracting the temperature data of the EMU battery at a minute-level time granularity; extracting the voltage signal and current signal of the EMU at an hour-level time granularity; performing frequency domain analysis on the voltage signal and current signal to obtain the electrochemical impedance spectrum. A neural network establishment module is used to take the aging mechanism information as a graph structure, where the nodes in the graph structure represent the electrochemical impedance spectrum or the characteristic data corresponding to different time levels, and the edges in the graph structure represent the association relationships between the characteristic data at different time levels; extract the graph topology features reflecting the battery aging state based on the graph structure and establish a corresponding graph neural network model. A result module is used to take the characteristic data as the input and the remaining life of the battery sample and the capacity fade inflection point as the output to train the graph neural network model to obtain a prediction model; input the new voltage data, current data, temperature data, and electrochemical impedance spectrum into the prediction model to obtain a prediction result.

[0021] In summary, the technical solution of the embodiment of the present application has at least the following advantages and beneficial effects: In the present invention, a battery operation database is established through floating charge voltage fluctuation monitoring and temperature data acquisition, and then the association relationships between parameters such as capacity retention rate and active lithium inventory and electrode deterioration and side reactions are extracted to form an aging mechanism model. To solve the problem of incomplete feature extraction, for the second-level dynamic events, minute-level temperature change trends, and hour-level electrical signal cycles during the operation of the EMU, the sudden acceleration frequency, temperature fluctuation curve, and frequency domain impedance spectrum are respectively extracted, and the dynamic time warping algorithm is used to eliminate data deviations under different driving modes. By constructing a graph neural network model with aging mechanism information as the graph neural network model and multi-time scale features as nodes, using the adaptive topological association characteristics of the graph neural network, the internal chemical attenuation process of the battery and the dynamic characteristics of the external working conditions are cross-level coupled and analyzed to realize the deep integration of microscopic mechanism and macroscopic data. The adaptability and accuracy of battery life prediction under complex working conditions are improved, and the capacity fade inflection point and the remaining life decay trajectory can be accurately identified. Description of the Drawings

[0022] Figure 1 It is a flowchart of a method for evaluating the health state and predicting the life of an EMU battery provided by the present invention. Figure 2 In the present invention, the voltage, temperature, current, and EIS of the battery system are used to describe the input relationship diagram of the feature data of the graph neural network; Figure 3 It is the adjacency matrix of the input relationship diagram of the feature data of the graph neural network of the present invention; Figure 4 It is the spatio-temporal fusion attention module of the graph neural network of the present invention; Figure 5 It is a schematic structural diagram of a battery health state assessment and life prediction system for EMUs provided by the present invention. Detailed implementation manners

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are some, but not all, of the embodiments of the present application. Usually, the components of the embodiments of the present application described and illustrated in the accompanying drawings here can be arranged and designed in various different configurations.

[0024] As Figure 1 shown, a method for assessing the battery health state and predicting the life of an EMU proposed in an embodiment of the present application includes: S101, based on the running state of the EMU at a preset speed within a preset time under the floating charge state, collect the floating charge voltage fluctuation data and the EMU battery temperature data, and establish a database.

[0025] Specifically, by real-time monitoring the running parameters of the EMU under the floating charge state, a basic data set reflecting the dynamic characteristics of the battery is constructed. The floating charge state refers to a charging mode in which, after the battery reaches full charge, its voltage is maintained stable by continuously inputting a small current. At this time, the electrochemical reaction inside the battery is in a dynamic equilibrium state, but the voltage fluctuation and temperature change still imply the key information of battery aging and performance degradation. Based on this, the present invention first synchronously collects the floating charge voltage fluctuation data of the battery and the battery body temperature data through on-vehicle sensors under the running conditions of the EMU at a high speed (such as 300 km / h) for a long time (such as continuously for one month). Among them, the floating charge voltage fluctuation data reflects the polarization characteristics and internal resistance change of the battery, while the temperature data characterizes the thermodynamic characteristics of the side reactions inside the battery. Specifically, the slight fluctuation of the floating charge voltage (for example, within the range of 52.5 V to 53.5 V) is directly related to the loss of battery active materials and the intensification of interfacial side reactions, while the temperature change (for example, in the range of 25°C to 45°C) affects the electrochemical reaction rate and lithium ion diffusion kinetics through the Arrhenius effect. During this process, all the collected data is dynamically associated with the running state of the EMU (such as traction or braking, etc.) through timestamp marking and stored in the database in a standardized format.

[0026] S102. Monitor the target parameters of the battery through the data corresponding to charge-discharge cycles in the database. The target parameters include the capacity retention rate, the inventory of active lithium, and the phase change amount of the cathode material. Establish the correlation mapping relationship between the target parameters and the loss of internal chemical active substances, the deterioration of the electrode structure, and the interfacial side reactions in the battery to obtain the aging mechanism information.

[0027] Specifically, taking the capacity retention rate, the inventory of active lithium, and the phase change amount of the cathode material of the battery as the core monitoring targets, combined with the voltage-capacity curve, differential capacity analysis, and electrochemical impedance spectroscopy (EIS), reveal the coupling relationship between the battery performance degradation and the loss of internal chemical active substances, the deterioration of the electrode structure, and the interfacial side reactions. The decrease in the capacity retention rate not only reflects the irreversible loss of the active material during the lithium-ion insertion / extraction process but is also directly related to the lattice distortion of the cathode material and the consumption of lithium inventory caused by the thickening of the solid electrolyte interface (SEI) film on the anode. By comparing the changes in the phase change amount of the cathode material before and after cycling, phenomena such as material structure collapse or phase separation can be identified, such as the local enrichment of nickel elements or the dissolution and loss of cobalt elements in ternary materials. These structural deteriorations will exacerbate the obstruction of the lithium-ion diffusion path, thereby accelerating the capacity decay.

[0028] Meanwhile, the dynamic monitoring of the inventory of active lithium combined with the quantitative analysis of the temperature sensitivity of the Arrhenius equation can locate the thermodynamic dominant mechanism of interfacial side reactions (such as electrolyte decomposition and lithium metal precipitation): when the battery temperature fluctuates beyond the stable range of 25°C to 45°C, the desolvation energy barrier of lithium ions at the electrode / electrolyte interface decreases, resulting in increased competition between the SEI film repair reaction and lithium dendrite growth, further consuming active lithium and causing an increase in internal resistance. By synchronously analyzing the voltage hysteresis phenomenon during the charge-discharge process and the evolution law of the charge transfer resistance (Rct) in the EIS spectrum, a multi-parameter collaborative aging fingerprint spectrum can be established. For example, when the increase in Rct exceeds 20% and the peak of the differential capacity shifts by 0.1V, it indicates that a significant kinetic barrier has been formed on the electrode surface passivation layer. This mapping relationship based on the multi-dimensional correlation of electrochemistry-material-thermodynamics not only realizes the accurate classification of battery aging modes (such as cathode-dominated decay, lithium inventory loss, or electrolyte dry-out) but also can reverse-derive the quantitative relationship between the electrode polarization voltage and the side reaction rate, providing physically interpretable feature inputs for the subsequent graph neural network model, thereby significantly improving the generalization ability of the life prediction model for complex operating conditions.

[0029] S103. Extract feature data by time levels, including extracting the frequency of rapid acceleration or deceleration events during the operation of the EMU at a second-level time granularity, aligning data with different sampling frequencies using the dynamic time warping algorithm and identifying the driving mode of the EMU; extracting the temperature data of the EMU battery at a minute-level time granularity; extracting the voltage signal and current signal of the EMU at an hour-level time granularity; performing frequency domain analysis on the voltage signal and current signal to obtain the electrochemical impedance spectrum Specifically, this step constructs a holographic representation of the battery's dynamic behavior through feature extraction and fusion technology with multi-level time granularity. First, for the driving events at the second-level time granularity, the system monitors the instantaneous action frequency of rapid acceleration or deceleration during the operation of the EMU in real time. Such high-dynamic events will cause severe fluctuations in the battery load, resulting in an instantaneous sharp increase in the lithium-ion concentration gradient at the electrode interface, and further accelerating the structural stress accumulation of the active material. To accurately correlate the driving behavior with the battery response, the dynamic time warping algorithm (DTW) is used to align the time series of sensor data with different sampling frequencies.

[0030] Specifically, the dynamic time warping algorithm mainly constructs a distance matrix between the test sequence (such as the current driving operation data) and the standard sequence (preset driving mode template). Based on the distance matrix, an alignment path from the starting point to the ending point of the matrix is found. The alignment path satisfies the preset boundary conditions, preset continuity conditions, and preset monotonic conditions. By accumulating the distances of each point on the path, the alignment distance is obtained. The corresponding calculation process is as follows: Suppose there are two time series and , is the test sequence with a length of ; is the standard sequence with a length of : (1) (2) where represents the operation data at the nth time point in the time series , represents the operation data at the mth time point in the time series .

[0031] To align the time series and , for the sequences and a distance matrix M of n×m is constructed. The element M i,j in the i-th row and j-th column of this matrix is the distance between the point and ​The Euclidean distance is usually adopted. After T and S are aligned, an alignment path W formed by matrix frames can be expressed as: (3) (4) Where represents the intermediate node of the alignment path W, represents the last node of the alignment path W.

[0032] The constraint conditions of the alignment path W are: Preset boundary conditions: The alignment path starts from the starting point of the distance matrix and ends at the other end of the diagonal of the distance matrix, that is , .

[0033] Preset continuity conditions: Only the matrices adjacent to the current path point can be selected for alignment, that is, if , then it must satisfy and .

[0034] Preset monotonicity conditions: The alignment path is monotonic on the time axis. That is , , then it satisfies and .

[0035] Starting from the matrix initial point (1,1), each subsequent point is the accumulation of the calculated distances of the points on the previous path. When reaching the end point , the cumulative distance is the final alignment distance , and the expression is as follows: (5) (6) If the alignment distance is less than or equal to the threshold, it indicates that the driving pattern in the test sequence is similar to that in the standard sequence, and it is judged as the same driving pattern and the corresponding driving pattern in the standard sequence is output at this time; if the alignment distance is greater than the threshold, the driving pattern corresponding to the test sequence is obtained by comparing the known driving patterns in the standard sequence.

[0036] In addition, the temperature data acquisition at the minute-level time granularity focuses on the thermal dynamic characteristics of the battery body, especially on the influence of temperature fluctuations on the electrochemistry reaction kinetics. For example, when the temperature rises from 25°C to 45°C, the lithium-ion diffusion coefficient increases exponentially with the Arrhenius relationship, but the overheated state (such as exceeding 50°C) will trigger a chain reaction of electrolyte decomposition. This level of data is processed by sliding window mean filtering to effectively suppress the interference of environmental noise on the temperature trend analysis.

[0037] Further, at the hourly time granularity, the wavelet packet decomposition method is used to extract multi-scale features from voltage and current signals: the original signal is decomposed into different frequency bands by presetting the decomposition level (such as 5 layers). The low-frequency components reflect the macroscopic evolution of the battery polarization process, and the high-frequency components capture the characteristic harmonic components of the microscopic side reactions on the electrode surface. Combining the charge and discharge cycle data collected synchronously, a fast Fourier transform is performed on the reconstructed sub-signals to generate an electrochemical impedance spectrum (EIS). The Cole-Cole plot (i.e., the Cole-Cole plot) of the real part and the imaginary part can quantify the collaborative growth law of the charge transfer resistance (Rct) and the ohmic internal resistance. For example, when the increase in Rct exceeds 20% and the phase angle in the low-frequency region decreases significantly, it indicates that the formation of the passivation layer at the electrode interface has seriously hindered the migration of lithium ions. This multi-time level feature extraction framework provides full-dimensional data support from transient shocks to long-term degradation for the subsequent graph neural network model through cross-scale fusion of second-level event-driven, minute-level thermodynamic tracking, and hourly impedance spectrum analysis. Among them, the wavelet decomposition expression is as follows: (7) In the formula, is the wavelet packet coefficient on the j-th sub-frequency band of the i-th layer after signal decomposition; i represents the layer number, j represents the number, 、 are the low-pass and high-pass filter coefficients respectively; is the number of sub-frequency bands. By iterating the above formula, the wavelet packet coefficients of the current signal and the voltage signal in different frequency bands can be obtained. Then, multi-threshold processing is performed on the wavelet packet coefficients to extract useful harmonic information, and the expression is: (8) In the formula, is the wavelet packet coefficient after multi-threshold processing; is the n-th threshold.

[0038] S104. Taking the aging mechanism information as the graph structure, the nodes in the graph structure represent the electrochemical impedance spectrum or the characteristic data corresponding to different time levels, and the edges in the graph structure represent the correlation relationships between the characteristic data at different time levels; based on the graph structure, extract the graph topological features reflecting the battery aging state and establish the corresponding graph neural network model; as Figure 2 shown in the figure, EIS in the figure is the electrochemical impedance spectrum, I is the current, U1, U2... Un are the voltage values extracted at different hourly time granularities, and T1, T2... Tn are the temperature values extracted at different minute-level time granularities.

[0039] Specifically, based on the aging mechanism information of the battery, a graph structure is constructed. In this graph structure, nodes represent Electrochemical Impedance Spectroscopy (EIS) or characteristic data corresponding to different time levels (second level, minute level, hour level), and these characteristic data include the battery voltage, current, temperature, and other key parameters derived from these data (such as the frequency of rapid acceleration / rapid deceleration events, battery temperature and humidity, etc.). The edges in the graph structure represent the correlation relationships between the characteristic data at different time levels, and these correlation relationships reflect the interaction and influence between various parameters of the battery under different operating conditions.

[0040] Next, based on the above graph structure, graph topological features reflecting the battery aging state are extracted. This process is achieved by introducing the Butler-Volmer equation (i.e., the Butler-Volmer equation), which describes how voltage changes affect Electrochemical Impedance Spectroscopy (EIS) by changing the polarization state of the electrode. Specifically, the Butler-Volmer equation reveals the internal connection between voltage and electrochemical impedance by expressing the relationships between parameters such as the electronic current in the external circuit of the battery, the ionic current in the internal circuit of the battery, the electrons in the external circuit of the battery, the ionic potential in the internal circuit of the battery, and conductivity. By solving this equation, the graph topological features of each node in the graph structure can be obtained, and these features contain key information about the battery aging state. The corresponding formula is: (9) Among them, the boundary conditions are: (10) In the formula, , are respectively the electronic current in the external circuit of the battery and the ionic current in the internal circuit of the battery; , are respectively the electronic potential in the external circuit of the battery and the ionic potential in the internal circuit of the battery, , are respectively the conductivity of the electrode solid skeleton and the electrolyte; is the working current output by the electrode; is the electrode thickness; is the product of the specific surface area of the porous electrode and the exchange current density; is the transfer coefficient; is the number of electrons transferred by the electrode reaction; is a constant; is the corresponding position of the electrode.

[0041] In addition, after constructing the graph neural network model, considering that the battery aging process is a complex spatio-temporal dynamic process, a spatio-temporal attention module is embedded between the graph neural network layers, and its principle is as Figure 4As shown. For any node, this module generates a temporal attention weight matrix and a spatial attention weight matrix respectively according to the characteristic change trend (temporal attention) of its adjacent nodes in the time series and the spatial topological connection strength (spatial attention). Subsequently, these two weight matrices are coupled to obtain a spatio-temporal fusion attention coefficient. Based on this coefficient, the features of adjacent nodes (i.e., the adjacent node matrix, such as Figure 3 as shown) are weighted and averaged to update the feature representation of the current node. This process not only considers the temporal dynamics during battery aging but also takes into account the spatial correlation between parameters, significantly improving the model's generalization and reasoning ability for battery aging paths.

[0042] S105, using the feature data as input and the remaining life of the battery sample and the capacity degradation inflection point as output, trains the graph neural network model to obtain a prediction model; inputs the new voltage data, current data, temperature data, and electrochemical impedance spectroscopy into the prediction model to obtain a prediction result.

[0043] Specifically, by introducing a graph neural network (GNN) improved with multi-dimensional feature fusion and spatio-temporal attention mechanism, the dynamic modeling of the characteristics of EMU battery aging and the life decay path is realized. Its core principle is to abstract the characteristics at different time levels during battery operation (such as second-level rapid acceleration events, minute-level temperature fluctuations, and hour-level electrochemical impedance spectroscopy) as nodes in the graph structure. After aligning the time series data through dynamic time warping (DTW), the spatio-temporal correlation between nodes is extracted using a graph convolutional network. For example, a sudden increase in temperature and a mutation in the charge transfer resistance (Rct) are mapped as strongly correlated edges between nodes. The spatio-temporal attention mechanism further dynamically adjusts the feature weights: in the time dimension, by analyzing the time correlation between voltage hysteresis effect and the increase in Rct, the non-linear acceleration characteristics of lithium dendrite growth under high-temperature conditions are quantified (such as when the temperature increases by 10 °C, the dendrite growth rate increases by 1.8 times); in the spatial dimension, based on the coupling relationship between diffusion impedance and differential capacity peak shift in the electrochemical impedance spectroscopy (EIS), the dominant role of abnormal phase change amount of the positive electrode material (such as a 30% decrease in lithium diffusion coefficient caused by nickel element enrichment) in capacity decay is identified. Through end-to-end training, this model learns the complex laws of the co-evolution of multiple parameters during battery aging and finally outputs the predicted remaining life value and the capacity inflection point (such as when the capacity suddenly drops to 80% ± 2% after 500 cycles).

[0044] Among them, the inflection point of capacity fade is obtained by the knee algorithm. The knee algorithm includes: extracting the time-series data of the capacity retention rate during the entire life cycle of the battery from the database, smoothing the time-series data to obtain the denoised aging trajectory curve; connecting the starting point and the ending point of the aging trajectory curve to generate a reference straight line, calculating the vertical distance from each sampling point on the curve to the reference straight line to form a distance sequence; traversing the distance sequence and selecting the point with the maximum vertical distance as the inflection point of capacity fade.

[0045] Specifically, first, extract the time-series data of the capacity retention rate of the battery during multiple charge and discharge cycles from the database. The capacity retention rate is defined as the percentage of the current cycle discharge capacity to the rated capacity. Due to the influence of sensor noise and operating condition fluctuations in actual operation, there are local outliers in the original time-series data. Therefore, the Savitzky-Golay filter is used to smooth the time-series data to obtain the denoised aging trajectory curve. For example, for a certain type of lithium-ion battery, its original capacity retention rate shows a fluctuating downward trend during 0 to 1000 cycles. After Savitzky-Golay filtering with a window width of 15 and a polynomial order of 3, the high-frequency noise is effectively suppressed, and a smooth decay curve is obtained.

[0046] Connect the starting point (the initial capacity corresponding to the cycle number of 0) and the ending point (the cycle number corresponding to the capacity retention rate decaying to a preset threshold such as 80%) of the aging trajectory curve to generate a reference straight line, which represents the theoretical aging path of the battery in an ideal linear decay mode. By calculating the vertical distance from each sampling point on the curve to the reference straight line, a distance sequence is formed. Specifically, for the cycle number corresponding to the i-th sampling point and its capacity retention rate , its coordinates in the rectangular coordinate system are C, where the slope k and the intercept b are determined by the starting point and the ending point. According to the formula for the vertical distance from a point to a straight line, the distance of the i-th sampling point can be expressed as: (11) After traversing all sampling points, select the sampling point corresponding to the maximum value in the distance sequence as the inflection point of capacity fade. Take a certain EMU battery as an example. The vertical distance between its aging trajectory curve and the reference straight line reaches the maximum value at the 500th cycle. At this time, the capacity retention rate drops from the initial 100% to 92.8%. The decay rate after the inflection point is about 3.5 times higher than that before the inflection point, indicating that irreversible electrode structure collapse or intensified interfacial side reactions have occurred inside the battery.

[0047] Further, after obtaining the prediction result, it further includes: using temperature as the acceleration stress and the thickening of the negative electrode SEI film as the main attenuation mechanism, describing the capacity attenuation law through the Arrhenius model to obtain the target model; calculating the capacity value of the battery after different numbers of cycles by using the target model according to the relative capacity attenuation of the battery at different numbers of cycles.

[0048] Specifically, a quantitative analysis model of capacity attenuation based on temperature acceleration stress is introduced. This model takes the thickening of the solid electrolyte interface (SEI) film on the negative electrode as the dominant attenuation mechanism, and establishes the temperature-aging rate correlation relationship by combining with the Arrhenius thermodynamic equation. That is, when the EMU battery operates under high-temperature conditions, the interfacial side reaction activity between the electrolyte and the negative electrode graphite is significantly enhanced, resulting in the dynamic balance of SEI film repair and growth being broken. At this time, additional active lithium needs to be consumed during the process of lithium ions embedding into the negative electrode to compensate for the lithium inventory loss caused by the thickening of the SEI film. By taking temperature as the acceleration aging factor and establishing the capacity attenuation activation energy parameters corresponding to different temperature ranges, the non-linear effect of temperature fluctuation on the long-term decline process of the battery can be quantified.

[0049] Specifically in implementation, taking a certain type of ternary lithium-ion power battery as an example, an accelerated aging test at 55 °C is carried out under laboratory conditions. By regularly disassembling the battery to measure the thickness of the negative electrode SEI film, it is found that when the number of cycles reaches 300 times, the thickness of the SEI film increases from the initial 2.1 nm to 5.8 nm, and the corresponding capacity retention rate drops from 100% to 82.3%. Based on the Arrhenius model, the capacity attenuation activation energy at this temperature is calculated to be 45 kJ / mol. Compared with 35 kJ / mol under the 25 °C reference condition, it shows that high temperature significantly accelerates the kinetic process of the interfacial side reaction. After inputting the corrected activation energy parameter into the target model, the capacity prediction error of the battery at the 150th cycle during actual vehicle operation is reduced from 8.7% before compensation to 2.3%, verifying the effectiveness of this method.

[0050] Among them, the formula corresponding to the target model is: (12) (13) Among them, is the relative capacity attenuation after n cycles; is a constant greater than 0; is the activation energy; is the gas constant; is the absolute temperature; is the number of cycles; is the exponent; is the capacity value estimated by the model after n cycles of the battery; is the initial capacity of the battery.

[0051] Based on the same inventive concept, as Figure 5 shown, the present invention provides a system for evaluating the health state and predicting the life of a multiple unit train battery, including: A data establishment module 201, which collects floating charge voltage fluctuation data and multiple unit train battery temperature data based on the running state of the multiple unit train at a preset speed within a preset time under the floating charge state, and establishes a database; An aging mechanism information acquisition module 202, which is used to monitor the target parameters of the battery through the data corresponding to the charge and discharge cycles in the database. The target parameters include the capacity retention rate, the stock of active lithium, and the phase change amount of the positive electrode material; establish an associated mapping relationship between the target parameters and the loss of internal chemical active substances of the battery, the deterioration of the electrode structure, and the interfacial side reaction, and obtain the aging mechanism information; A feature data extraction module 203, which is used to extract feature data at different time levels, including extracting the frequency of rapid acceleration or rapid deceleration events during the running of the multiple unit train at a second-level time granularity, aligning data with different sampling frequencies using the dynamic time warping algorithm and identifying the driving mode of the multiple unit train; extracting the temperature data of the multiple unit train battery at a minute-level time granularity; extracting the voltage signal and current signal of the multiple unit train at an hour-level time granularity; performing frequency domain analysis on the voltage signal and current signal to obtain the electrochemical impedance spectrum; A neural network establishment module 204, which uses the aging mechanism information as a graph structure. In the graph structure, nodes represent the electrochemical impedance spectrum or feature data corresponding to different time levels, and edges in the graph structure represent the associated relationship between feature data at different time levels; extract graph topological features reflecting the battery aging state based on the graph structure, and establish a corresponding graph neural network model; A result module 205, which uses the feature data as input and the remaining life of the battery sample and the capacity decline inflection point as output to train the graph neural network model to obtain a prediction model; input new voltage data, current data, temperature data, and electrochemical impedance spectrum into the prediction model to obtain a prediction result.

[0052] The above are only the preferred embodiments of the present application and are not used to limit the present application. For those skilled in the art, various changes and modifications can be made to the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for evaluating the health state and predicting the remaining useful life of the batteries of multiple unit trains, characterized in that, Including: Based on the running state of the motor car of the multiple unit train at a preset speed within a preset time under the floating charge state, collect the floating charge voltage fluctuation data and the battery temperature data of the multiple unit train, and establish a database; Monitor the target parameters of the battery through the data corresponding to the charge and discharge cycles in the database, where the target parameters include the capacity retention rate, the inventory of active lithium, and the phase change amount of the positive electrode material; establish the correlation mapping relationship between the target parameters and the loss of internal chemical active substances of the battery, the deterioration of the electrode structure, and the interfacial side reactions, and obtain the aging mechanism information; Extract feature data at different time levels, including extracting the frequency of rapid acceleration or rapid deceleration events during the running of the motor car at the second-level time granularity, using the dynamic time warping algorithm to align data with different sampling frequencies and identify the driving mode of the motor car; extracting the temperature data of the multiple unit train battery at the minute-level time granularity; extracting the voltage signal and current signal of the motor car at the hour-level time granularity; Perform frequency domain analysis on the voltage signal and current signal to obtain the electrochemical impedance spectrum; Take the aging mechanism information as a graph structure, where the nodes in the graph structure represent the electrochemical impedance spectrum or the feature data corresponding to different time levels, and the edges in the graph structure represent the correlation relationships between the feature data at different time levels; extract the graph topology features reflecting the battery aging state based on the graph structure, and establish a corresponding graph neural network model; Use the feature data as the input and the remaining life and capacity decline inflection point of the battery sample as the output to train the graph neural network model to obtain a prediction model; input the new voltage data, current data, temperature data, and electrochemical impedance spectrum into the prediction model to obtain the prediction result.

2. The method for evaluating the health state and predicting the remaining useful life of the EMU battery according to claim 1, characterized in that, The step of extracting the voltage signal and current signal of the motor car at the hour-level time granularity specifically includes: Select a preset decomposition layer number, and perform multi-scale decomposition and extraction on the voltage signal and current signal through the wavelet packet decomposition method.

3. The method for evaluating the health state and predicting the life of the multiple unit train battery according to claim 1, wherein The step of using the dynamic time warping algorithm to align data with different sampling frequencies and identify the driving mode of the motor car specifically includes: Obtain the test sequence and standard sequence of the motor car driving mode; Construct the distance matrix of the test sequence and the standard sequence; Based on the distance matrix, find an alignment path from the starting point to the end point of the matrix, where the alignment path satisfies the preset boundary conditions, preset continuous conditions, and preset monotonic conditions. By accumulating the distances of each point on the path, obtain the alignment distance; the preset boundary condition is that the alignment path starts from the starting point of the distance matrix and terminates at the other end of the diagonal of the distance matrix; the preset continuous condition is that only the matrix adjacent to the current path point can be selected for alignment; the preset monotonic condition is that the alignment path is monotonic on the time axis; If the alignment distance is less than or equal to the threshold, it indicates that the driving mode in the test sequence is similar to that in the standard sequence, and it is judged as the same driving mode and the driving mode corresponding to the standard sequence at this time is output; if the alignment distance is greater than the threshold, the driving mode corresponding to the test sequence is obtained by comparing the known driving modes in the standard sequence.

4. A method for evaluating the health state and predicting the remaining useful life of the battery of a multiple unit train according to claim 1, characterized in that, The graph topology features reflecting the battery aging state are extracted based on the graph structure, specifically including: Using the voltage, current, and electrochemical impedance spectrum of the battery as input parameters, the influence relationship of the voltage changing the electrode polarization state on the electrochemical impedance spectrum is represented by the Butler-Volmer equation to obtain the graph topology features.

5. The method for evaluating the health state and predicting the life of the EMU battery according to claim 4, characterized in that, The formula corresponding to the influence relationship is: Among them, the boundary conditions are: Wherein, and are the electronic current of the external circuit of the battery and the ionic current of the internal circuit of the battery, respectively; and are the electronic potential of the external circuit of the battery and the ionic potential of the internal circuit of the battery, respectively; and are the electrical conductivities of the electrode solid skeleton and the electrolyte, respectively; is the working current output by the electrode; is the electrode thickness; is the product of the specific surface area of the porous electrode and the exchange current density; is the transfer coefficient; is the number of electrons transferred by the electrode reaction; is a constant; is the corresponding position of the electrode.

6. A method for evaluating the health state and predicting the remaining useful life of the batteries of multiple unit trains according to claim 1, characterized in that, After establishing the corresponding graph neural network model, it further includes: Based on the node features and adjacency relationships of the graph structure, a spatio-temporal attention module is embedded between each graph neural network layer. The spatio-temporal attention module includes: For any node, according to the feature change trend of its adjacent nodes in the time series and the spatial topological connection strength, a time attention weight matrix and a spatial attention weight matrix are respectively generated; the time attention weight matrix and the spatial attention weight matrix are coupled to obtain a spatio-temporal fusion attention coefficient; based on the spatio-temporal fusion attention coefficient, the features of the adjacent nodes are weighted and averaged to update the feature representation of the current node.

7. A method for evaluating the health state and predicting the remaining useful life of a battery in a multiple unit train according to claim 1, wherein The capacity decay inflection point is obtained by the knee algorithm; the knee algorithm includes: Extracting the time series data of the capacity retention rate during the full life cycle of the battery from the database, and performing smoothing processing on the time series data to obtain a denoised aging trajectory curve; Connecting the starting point and the ending point of the aging trajectory curve to generate a reference straight line, calculating the perpendicular distance from each sampling point on the curve to the reference straight line to form a distance sequence; Traversing the distance sequence, and selecting the point with the maximum perpendicular distance as the capacity decay inflection point.

8. A method for evaluating the health state and predicting the remaining useful life of a battery of a multiple unit train according to claim 1, characterized in that, After obtaining the prediction result, it further includes: Using temperature as the acceleration stress and the thickening of the negative electrode SEI film as the main attenuation mechanism, the capacity decay law is described by the Arrhenius model to obtain the target model; According to the relative capacity decay amount of the battery at different cycle numbers, the capacity value of the battery after different cycle numbers is calculated using the target model.

9. The method for evaluating the health state and predicting the remaining useful life of the EMU battery according to claim 8, wherein The formula corresponding to the target model is: Among them, is the relative capacity attenuation after n cycles; is a constant greater than 0; is the activation energy; is the gas constant; is the absolute temperature; is the number of cycles; is the exponent; is the capacity value estimated by the model after n cycles of the battery; is the initial capacity of the battery.

10. A system for evaluating the health status and predicting the remaining useful life of the batteries of multiple unit trains, characterized in that, Including: A data establishment module for collecting floating charge voltage fluctuation data and EMU battery temperature data based on the EMU running state at a preset speed within a preset time under the floating charge state, and establishing a database; An aging mechanism information acquisition module for monitoring the target parameters of the battery through the data corresponding to the charge and discharge cycles in the database. The target parameters include the capacity retention rate, the active lithium stock, and the phase change amount of the positive electrode material; establishing an association mapping relationship between the target parameters and the loss of internal chemical active substances, the deterioration of the electrode structure, and the interfacial side reactions of the battery to obtain aging mechanism information; A feature data extraction module for extracting feature data at different time levels, including extracting the frequency of rapid acceleration or rapid deceleration events during EMU operation at a second-level time granularity, aligning data with different sampling frequencies using the dynamic time warping algorithm and identifying the driving mode of the EMU; extracting the temperature data of the EMU battery at a minute-level time granularity; extracting the voltage signal and current signal of the EMU at an hour-level time granularity; Performing frequency domain analysis on the voltage signal and current signal to obtain the electrochemical impedance spectrum; A neural network establishment module, which is used to take the aging mechanism information as a graph structure, where nodes in the graph structure represent electrochemical impedance spectra or characteristic data corresponding to different time levels, and edges in the graph structure represent the correlation relationships between characteristic data at different time levels; extract graph topological features reflecting the battery aging state based on the graph structure, and establish a corresponding graph neural network model; A result module, which is used to take the characteristic data as input, and the remaining life and capacity fade inflection point of the battery sample as output, train the graph neural network model to obtain a prediction model; input new voltage data, current data, temperature data and electrochemical impedance spectra into the prediction model to obtain a prediction result.

Images