Method and system for on-line detection of multiple parameters of power battery during charging process

Abstract

The application discloses a charging process power battery multi-parameter online detection method and system, relates to the technical field of battery life detection, and comprises the following steps: constructing a charging space-time topology graph in a charging process, collecting multi-modal sensing data, extracting three types of internal features of a battery through an impedance decoupling network containing multiple encoders and decoupling branches; constructing a migration sample in combination with similar working conditions of neighbor vehicles, obtaining a health degree feature through a space-time graph neural network, and finally outputting a power battery health degree detection result by a decoder, which is suitable for a complex charging scene, realizes non-invasive online detection, improves the precision and reliability of power battery health degree detection, and can accurately identify the attenuation mechanism, thereby providing support for battery life cycle management.

Description

Technical Field

[0001] This invention relates to the field of battery life testing technology, and in particular to a method and system for online testing of multiple parameters of a power battery during the charging process. Background Technology

[0002] Current methods for testing power batteries during charging mostly only collect conventional electrical parameters such as voltage, current, and temperature, calculating battery health through coulomb integration or equivalent circuit models. Some solutions introduce impedance detection methods, but these are mostly offline and invasive, making them unsuitable for online charging scenarios. At the data processing level, existing technologies typically rely solely on historical data from individual vehicles for analysis, failing to utilize the spatiotemporal correlation formed by charging pile clusters and charging time sequences. At the model level, most methods use a single network for simple fitting, failing to achieve multi-modal signal fusion and attenuation mode classification, making it difficult to comprehensively reflect the internal aging state of the battery.

[0003] Existing detection methods cannot cope with power disturbances and multi-vehicle concurrent operation in V2G scenarios, and it is difficult to collect multi-dimensional features that reflect the internal conduction, liquid phase transport and interface reaction of the battery. They lack modeling of the spatiotemporal correlation of charging and cannot use data from vehicles under similar operating conditions to improve detection accuracy. At the same time, they lack an integrated network architecture that can decouple the internal aging characteristics of the battery, integrate spatiotemporal information and simultaneously output health, remaining life and degradation type, resulting in one-sided online detection results, insufficient accuracy and inability to identify degradation mechanisms.

[0004] Therefore, how to achieve high-precision health detection of power batteries in complex charging scenarios has become an urgent problem to be solved. Summary of the Invention

[0005] The main purpose of this application is to provide a method and system for online detection of multiple parameters of power batteries during the charging process, aiming to solve the technical problem of how to achieve high-precision health detection of power batteries in complex charging scenarios.

[0006] To achieve the above objectives, this application proposes a method for online detection of multiple parameters of a power battery during the charging process, comprising: During the charging process, a spatiotemporal topology network for charging is constructed with charging piles as spatial nodes and charging periods as time scales, resulting in a spatiotemporal topology map. Collect multimodal sensor data of the target vehicle; The multimodal sensing data is input to the impedance decoupling network for feature extraction to obtain the battery's internal conduction characteristics, liquid transport characteristics, and interface reaction characteristics. Based on the neighboring vehicle nodes in the spatiotemporal topology graph, historical attenuation trajectories similar to the operating conditions of the target vehicle are obtained to obtain a transfer learning sample set. The battery's internal conduction characteristics, liquid transport characteristics, interface reaction characteristics, and transfer learning sample set are input into a spatiotemporal graph neural network for processing to obtain health characteristics; Based on the health characteristics, the current health value, the predicted remaining cycle life value, and the degradation type label are output through the decoder network to obtain the power battery health detection result.

[0007] In one embodiment, the step of constructing a spatiotemporal topology network for charging, with charging piles as spatial nodes and charging periods as time scales, to obtain a spatiotemporal topology map during the charging process includes: Obtain the geographical coordinates and grid access capacity of all charging piles within the target area to obtain the spatial attribute set of the charging piles; Based on historical charging records, the frequency of vehicle access and power load curves of each charging pile within a preset time period are statistically analyzed to obtain the charging pile time attribute set. Using charging piles as graph nodes and the connection relationships of charging piles with geographical distance less than a preset distance threshold and the same power grid phase as graph edges, a spatial adjacency matrix is ​​constructed to obtain a spatial graph structure; Using charging vehicles at the same charging pile at different times as time-series nodes and vehicles with the same battery model and overlapping state of charge intervals as time-series edges, a time correlation matrix is ​​constructed to obtain a time graph structure. The spatial graph structure and the temporal graph structure are fused through graph product operation to obtain a spatiotemporal topological graph.

[0008] In one embodiment, the step of collecting multimodal sensing data of the target vehicle includes: During the constant current charging phase, a charging micro-pulse excitation with a preset amplitude and preset duration is inserted, and the voltage transient response curves before and after the pulse are collected to obtain the charging micro-response signal. A temperature distribution sequence is obtained by collecting the two-dimensional thermal field distribution during the charging process using temperature sensors arranged on the surface of the battery module at a preset sampling frequency. Acoustic sensors attached to the battery casing are used to collect acoustic vibration signals generated by electrolyte vaporization and ion migration during charging, and the vibration signals are obtained by spectrum analysis. The charging micro-response signal, the temperature distribution sequence, and the vibration signal are aligned according to timestamps and enhanced with intermodal cross-correlation to obtain multimodal sensing data.

[0009] In one embodiment, the step of extracting features from the multimodal sensing data input impedance decoupling network to obtain battery internal conduction features, liquid transport features, and interface reaction features includes: An impedance decoupling network is constructed, wherein the impedance decoupling network includes a voltage response encoder, a thermal field encoder, an acoustic encoder, a feature fusion layer, and an impedance decoupling head. The voltage response encoder includes a first convolutional layer, a bidirectional long short-term memory layer, and a first fully connected layer. The thermal field encoder includes a second convolutional layer, a spatial attention layer, and a channel attention layer. The acoustic encoder includes a third convolutional layer, a Mel frequency cepstral coefficient extraction layer, and a second fully connected layer. The impedance decoupling head consists of three decoupling branches in parallel, and each decoupling branch consists of a third fully connected layer, a fourth convolutional layer, and a fourth fully connected layer. The charging micro-response signal in the multimodal sensing data is input to the voltage response encoder for time-frequency feature extraction to obtain the voltage time-frequency features; The temperature distribution sequence in the multimodal sensing data is input into the thermal field encoder to locate the thermal anomaly region and extract its features, thereby obtaining the spatial features of the thermal field. The vibration signal in the multimodal sensing data is input into an acoustic encoder to extract acoustic features, thereby obtaining acoustic frequency domain features; The voltage time-frequency features, the thermal field spatial features, and the acoustic frequency domain features are input into the feature fusion layer and subjected to cross-modal attention weighting to obtain a fusion vector. The fusion vector input impedance decoupling head is output separately to obtain the battery's internal conduction characteristics, liquid transport characteristics, and interface reaction characteristics.

[0010] In one embodiment, the step of obtaining a transfer learning sample set by acquiring historical decay trajectories similar to the target vehicle's operating conditions based on neighboring vehicle nodes in the spatiotemporal topology graph includes: In the spatiotemporal topology graph, with the charging pile node where the target vehicle is located as the center, the neighboring vehicle nodes within a preset number of hops are queried to obtain a set of candidate neighbor nodes. Based on the historical charging records of each vehicle in the candidate neighbor node set, the similarity of the charging power curve, the similarity of the ambient temperature distribution, and the similarity of the initial state of charge with the target vehicle are calculated to obtain a multi-dimensional operating condition similarity vector. Based on the multi-dimensional working condition similarity vector, neighbor vehicles with similarity greater than a preset similarity threshold are selected to obtain a set of highly similar neighbor vehicles; Obtain the historical health detection records and corresponding loop count sequences of the highly similar neighbor vehicle set, construct the decay trajectory matrix, and obtain the transfer learning sample set.

[0011] In one embodiment, the step of inputting the battery's internal conduction characteristics, the liquid transport characteristics, the interface reaction characteristics, and the transfer learning sample set into a spatiotemporal graph neural network for processing to obtain health characteristics includes: Construct a spatiotemporal graph neural network, wherein the spatiotemporal graph neural network includes a spatial flow branch, a temporal flow branch, a cross-flow fusion module, a temporal pooling layer, and a fifth fully connected layer; In the spatial flow branch, the battery internal conduction characteristics, the liquid transport characteristics, and the interface reaction characteristics are used as the initial node characteristics. The spatial dependency relationship between neighboring nodes is learned through the graph attention layer to obtain the spatial aggregation characteristics. The graph attention layer includes a first query matrix, a first key matrix, a first value matrix, and a multi-head attention mechanism. In the temporal branch, the historical decay trajectory in the transfer learning sample set is used as the temporal input. The temporal dependency of decay evolution is learned through a temporal convolutional network and a gated recurrent unit to obtain temporal evolution features. The temporal convolutional network includes causal convolutional layers and dilated convolutional layers, and the gated recurrent unit includes an update gate and a reset gate. In the cross-stream fusion module, the interaction weights between the spatial aggregation features and the temporal evolution features are calculated through a cross-modal attention mechanism to obtain cross-stream fusion features, wherein the cross-modal attention mechanism includes a second query matrix, a second key matrix, and a second value matrix; The cross-stream fusion features are input into the temporal pooling layer and the fifth fully connected layer for dimensionality compression to obtain the health features.

[0012] In one embodiment, the step of obtaining the power battery health detection result by outputting the current health value, the predicted remaining cycle life value, and the degradation type label through a decoder network based on the health characteristics includes: Construct a decoder network, wherein the decoder network includes a health branch, a lifetime prediction branch, and a decay mode classification branch. The health branch includes a sixth fully connected layer, a first deconvolutional layer, and a seventh fully connected layer. The lifetime prediction branch includes an eighth fully connected layer, a second bidirectional long short-term memory layer, and a ninth fully connected layer. The decay mode classification branch includes a tenth fully connected layer, a fourth convolutional layer, and an eleventh fully connected layer. The health feature is mapped to the health value space through the health branch to obtain the current health evaluation value; The health characteristics are output in a sequence generation manner through the lifespan prediction branch to obtain the remaining cycle life prediction value; The health characteristics are mapped to the decay mode category space through the decay mode classification branch to obtain decay mode classification labels, wherein the decay mode classification labels include normal decay mode, lithium deposition decay mode, electrolyte drying decay mode and active material loss decay mode. The power battery health test result is obtained based on the current health evaluation value, the remaining cycle life prediction value, and the degradation mode classification label.

[0013] Furthermore, to achieve the above objectives, this application also proposes an online multi-parameter detection system for a power battery during the charging process, the online multi-parameter detection system for a power battery during the charging process comprising: The spatiotemporal topology construction module is used to construct a spatiotemporal topology network for charging, with charging piles as spatial nodes and charging periods as time scales, during the charging process, and obtain a spatiotemporal topology map. The multimodal acquisition module is used to acquire multimodal sensing data of the target vehicle based on the spatiotemporal topology map; The impedance decoupling module is used to input the multimodal sensing data into the impedance decoupling network for feature extraction, thereby obtaining the battery's internal conduction characteristics, liquid transport characteristics, and interface reaction characteristics. The transfer learning module is used to obtain historical decay trajectories similar to the operating conditions of the target vehicle based on neighbor vehicle nodes in the spatiotemporal topology graph, and to obtain a transfer learning sample set. The spatiotemporal graph neural network module is used to input the battery's internal conduction features, liquid transport features, interface reaction features, and transfer learning sample set into the spatiotemporal graph neural network for processing to obtain health features; The results module is used to output the current health value, the predicted remaining cycle life value, and the degradation type label through the decoder network based on the health characteristics, so as to obtain the power battery health detection results.

[0014] Furthermore, the system also includes: the impedance decoupling module, which is further used to construct an impedance decoupling network, wherein the impedance decoupling network includes a voltage response encoder, a thermal field encoder, an acoustic encoder, a feature fusion layer, and an impedance decoupling head. The voltage response encoder includes a first convolutional layer, a bidirectional long short-term memory layer, and a first fully connected layer. The thermal field encoder includes a second convolutional layer, a spatial attention layer, and a channel attention layer. The acoustic encoder includes a third convolutional layer, a Mel-frequency cepstral coefficient extraction layer, and a second fully connected layer. The impedance decoupling head consists of three parallel decoupling branches, each of which consists of a third fully connected layer, a fourth convolutional layer, and a fourth fully connected layer. For the multimodal... The charging micro-response signal from the sensing data is input to a voltage response encoder for time-frequency feature extraction to obtain voltage time-frequency features; the temperature distribution sequence from the multimodal sensing data is input to a thermal field encoder for thermal anomaly region localization and feature extraction to obtain thermal field spatial features; the vibration signal from the multimodal sensing data is input to an acoustic encoder for acoustic signature feature extraction to obtain acoustic frequency domain features; the voltage time-frequency features, the thermal field spatial features, and the acoustic frequency domain features are input to a feature fusion layer for cross-modal attention weighting to obtain a fusion vector; the fusion vector is input to an impedance decoupling head for output to obtain battery internal conduction features, liquid transport features, and interface reaction features.

[0015] Furthermore, the system also includes: the spatiotemporal graph neural network module, which is further used to construct the spatiotemporal graph neural network, wherein the spatiotemporal graph neural network includes a spatial flow branch, a temporal flow branch, a cross-flow fusion module, a temporal pooling layer, and a fifth fully connected layer; in the spatial flow branch, the battery internal conduction features, the liquid transport features, and the interface reaction features are used as initial node features, and the spatial dependencies between neighboring nodes are learned through a graph attention layer to obtain spatial aggregation features, wherein the graph attention layer includes a first query matrix, a first key matrix, a first value matrix, and a multi-head attention mechanism; in the temporal flow branch, the transfer learning sample set is used as the initial node features, and the spatial dependencies between neighboring nodes are learned through a graph attention layer to obtain spatial aggregation features, wherein the graph attention layer includes a first query matrix, a first key matrix, a first value matrix, and a multi-head attention mechanism; Historical decay trajectories are used as temporal inputs. Temporal evolution features are obtained by learning the temporal dependencies of decay evolution through a temporal convolutional network and a gated recurrent unit. The temporal convolutional network includes causal convolutional layers and dilated convolutional layers, and the gated recurrent unit includes update gates and reset gates. In the cross-stream fusion module, the interaction weights between the spatial aggregation features and the temporal evolution features are calculated using a cross-modal attention mechanism to obtain cross-stream fused features. This cross-modal attention mechanism includes a second query matrix, a second key matrix, and a second value matrix. The cross-stream fused features are then input into a temporal pooling layer and a fifth fully connected layer for dimensionality compression to obtain health features.

[0016] This application constructs a charging spatiotemporal topology map during the charging process, collects multimodal sensor data, and extracts three types of internal features of the battery through an impedance decoupling network containing multiple encoders and decoupling branches. It also constructs transfer samples based on similar operating conditions of neighboring vehicles, obtains health features through a spatiotemporal graph neural network, and finally outputs the power battery health detection results by the decoder. This method is suitable for complex charging scenarios, realizes non-invasive online detection, improves the accuracy and reliability of power battery health detection, can accurately identify degradation mechanisms, and provides support for battery life cycle management. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the first embodiment of the online multi-parameter detection method for power batteries during the charging process of this application. Figure 2 This is a flowchart illustrating the second embodiment of the online multi-parameter detection method for power batteries during the charging process of this application. Figure 3 This is a schematic diagram of the module structure of the online multi-parameter detection system for the power battery during the charging process of this application. Figure 4 This is a schematic diagram of the equipment structure of the hardware operating environment involved in the online detection method for multiple parameters of a power battery during the charging process in the embodiments of this application.

[0019] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0020] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0021] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0022] Current methods for testing power batteries during charging mostly only collect conventional electrical parameters such as voltage, current, and temperature, calculating battery health through coulomb integration or equivalent circuit models. Some solutions introduce impedance detection methods, but these are mostly offline and invasive, making them unsuitable for online charging scenarios. At the data processing level, existing technologies typically rely solely on historical data from individual vehicles for analysis, failing to utilize the spatiotemporal correlation formed by charging pile clusters and charging time sequences. At the model level, most methods use a single network for simple fitting, failing to achieve multi-modal signal fusion and attenuation mode classification, making it difficult to comprehensively reflect the internal aging state of the battery.

[0023] Existing detection methods are unable to handle power disturbances and multi-vehicle concurrent operation in V2G scenarios, making it difficult to collect multi-dimensional features reflecting internal battery conduction, liquid phase transport, and interface reactions. They lack modeling of the spatiotemporal correlation of charging, making it impossible to improve detection accuracy using data from vehicles under similar conditions. Furthermore, they lack an integrated network architecture capable of decoupling internal battery aging characteristics, fusing spatiotemporal information, and simultaneously outputting health status, remaining lifespan, and degradation type, resulting in incomplete, inaccurate, and unidentifiable degradation mechanisms in online detection results. Therefore, achieving high-precision health detection of power batteries in complex charging scenarios is an urgent problem to be solved.

[0024] Based on the above, this application also provides a method for online detection of multiple parameters of a power battery during the charging process, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the online multi-parameter detection method for power batteries during the charging process of this application.

[0025] In this embodiment, the online detection method for multiple parameters of the power battery during the charging process includes steps S10 to S60: Step S10: During the charging process, a charging spatiotemporal topology network is constructed with charging piles as spatial nodes and charging periods as time scales to obtain a spatiotemporal topology map.

[0026] It's important to note that a charging pile is a hardware device that provides power to electrical equipment. This device transmits power through circuit connections and can stably output charging voltage and current adapted to the equipment. The charging period is the time interval during which charging occurs. This interval is divided based on the time nodes when the equipment starts and ends charging, accurately defining the time range of the charging activity. The spatiotemporal topology network is a relational network that integrates spatial location and temporal scales. This network combines spatial nodes and temporal scales to form a network architecture with corresponding relationships. The spatiotemporal topology network can clearly present the correspondence between the spatial location of the charging pile and the charging period, providing a basic framework for the subsequent sorting and analysis of charging-related data.

[0027] Further, step S10 includes: obtaining the geographical coordinates and grid access capacity of all charging piles in the target area to obtain a spatial attribute set of charging piles; statistically analyzing the vehicle access frequency and power load curve of each charging pile within a preset time period based on historical charging records to obtain a temporal attribute set of charging piles; constructing a spatial adjacency matrix with charging piles as graph nodes and charging piles with geographical distance less than a preset distance threshold and the same grid phase as graph edges to obtain a spatial graph structure; constructing a temporal correlation matrix with charging vehicles of the same charging pile at different time periods as temporal nodes and vehicles with the same battery model and overlapping state of charge intervals as temporal edges to obtain a temporal graph structure; and fusing the spatial graph structure and the temporal graph structure through graph product operation to obtain a spatiotemporal topology graph.

[0028] It's important to understand that geographic coordinates are numerical parameters used to identify the physical location of a charging pile. These parameters typically consist of longitude and latitude, accurately determining the charging pile's specific spatial location within a target area. Grid access capacity is the maximum power capacity a charging pile can obtain from the grid. This value represents the upper limit of power a charging pile can stably obtain from the grid, determining the scale of charging operations it can support. Historical charging records are data retained from past charging operations, recording key information such as the time, object, and power of the charging activity, accurately reflecting the charging pile's actual operating status. Grid phase is the phase state of alternating current during grid transmission. This state is determined by the grid's power supply characteristics. Same grid phase indicates that the charging pile is connected to the same type of phase-matched grid power supply line; only charging piles with consistent phases can establish an effective spatial connection.

[0029] Specifically, firstly, the geographical coordinates and grid access capacity of all charging piles within the target area are obtained to form a spatial attribute set for the charging piles. Then, based on historical charging records, the vehicle access frequency and power load curves of each charging pile within a preset time period (e.g., 24 hours) are statistically analyzed to obtain a temporal attribute set for the charging piles. This establishes a unified description of the static physical attributes and dynamic operational attributes of the charging piles, providing a node feature foundation for subsequent graph structure construction. For example, a charging pile in a commercial area can be marked as "(longitude 114.3, latitude 30.5, capacity 120 kW)" and associated with its high-frequency, high-load operation characteristics during the morning peak. Secondly, a spatial adjacency matrix is ​​constructed using charging piles as graph nodes and charging pile connections with geographical distances less than a preset distance threshold (e.g., 500 meters) and the same grid phase as graph edges to obtain a spatial graph structure. This associates charging piles that are physically adjacent and have synchronized grid connections, capturing the grid load coupling effect and voltage fluctuation correlation of multiple vehicles charging within the same power supply area. This ensures that neighboring nodes have a shareable grid environment consistency. For example, three adjacent charging piles under the same substation can be bidirectionally connected to ensure they exhibit similar voltage drop characteristics during grid disturbances. Then, a time graph structure is constructed by using charging vehicles at the same charging pile at different times as time-series nodes and charging vehicles with the same battery model and overlapping state-of-charge intervals as time-series edges. This structure associates vehicles with similar operating conditions at the same physical location at different times, establishing a transmission path from historical decay trajectories to the current detection time. This allows the detection results of similar operating conditions in the past to be used for transfer learning of the current vehicle. For example, a vehicle that charged from 20% to 80% at this charging pile this Wednesday can be connected with a historical vehicle that charged at the same pile, with the same battery model, and in the same interval last Wednesday, using its known health decay data. Finally, the spatial graph structure and the time graph structure are fused through graph product operations to obtain a spatiotemporal topology graph. The charging pile (spatial dimension) and the charging vehicle (time / behavioral dimension) are regarded as two orthogonal coordinate axes describing the same physical process—energy replenishment, just as locating a point on a map requires longitude and latitude. Secondly, the new spatiotemporal nodes obtained through graph product operations (such as Cartesian product) physically represent a spatiotemporally coupled charging feature observation unit. They no longer refer to a single hardware or mobile entity, but rather represent an evolution of a specific battery state under a specific power environment. This enables a topological representation of multi-vehicle charging scenarios in the same area at the same time, generating a composite graph structure that simultaneously includes "neighboring charging pile relationships" and "historical-current vehicle relationships." This allows the currently detected vehicle to obtain real-time grid state information from its spatial neighbors and historical attenuation references from its temporal neighbors, providing a structured sample retrieval and feature propagation foundation for subsequent transfer learning based on neighboring vehicles.

[0030] Step S20: Collect multimodal sensing data of the target vehicle.

[0031] It should be noted that step S20 includes: inserting a charging micro-pulse excitation with a preset amplitude and preset duration during the constant current charging stage, and acquiring the voltage transient response curves before and after the pulse to obtain the charging micro-response signal; acquiring the two-dimensional thermal field distribution during the charging process through a temperature sensor arranged on the surface of the battery module at a preset sampling frequency to obtain the temperature distribution sequence; acquiring the acoustic vibration signal generated by electrolyte vaporization and ion migration during the charging process through an acoustic sensor attached to the battery casing, and performing spectrum analysis to obtain the vibration signal; aligning the charging micro-response signal, temperature distribution sequence, and vibration signal according to the timestamp and performing intermodal cross-correlation enhancement to obtain multimodal sensing data.

[0032] Specifically, firstly, when the battery enters the constant current charging stable period, the charging controller actively injects a very short (e.g., 50ms-200ms) current step pulse with a constant amplitude into the battery. A high-frequency sampling chip (sampling rate no less than 1kHz) records the dynamic evolution waveform of the battery terminal voltage in real time at the moment of pulse triggering, during the pulse duration, and after the pulse is removed. This is done to extract the ohmic internal resistance through the "instantaneous voltage jump" and capture the dynamic characteristics of electrochemical polarization and concentration polarization through the subsequent "exponential recovery curve," obtaining a micro-response signal reflecting the battery's internal impedance characteristics. Secondly, high-precision temperature sensors (such as thermistors) deployed on the surface of the battery module capture temperature data from different parts of the battery at fixed time intervals and map it onto a two-dimensional coordinate plane to form a continuous thermal imaging sequence. This is done to accurately identify whether there are non-uniform heat generation phenomena caused by excessive local current density or micro-short circuits inside the battery by observing the rate of heat generation and diffusion.

[0033] Then, a piezoelectric acoustic emission (AE) sensor attached to the hard surface of the battery is activated to monitor in real time the mechanical vibration waves excited by ion shuttles, solid electrolyte interphase (SEI) repair, and electrolyte decomposition during the charging and discharging chemical reactions inside the battery. These vibration waves are generated by tiny bubbles and are immediately transformed into power spectral density maps containing different frequency components using Fast Fourier Transform (FFT). This is done to help identify internal physical changes that are imperceptible to the naked eye and voltmeters, such as lithium plating tendency or electrolyte consumption, through acoustic signature features. Finally, a global timestamp is applied to the voltage waveform, temperature thermal field sequence, and vibration spectrum using a unified clock synchronization protocol. Cross-correlation is used to find correlation features between different modal data, and weighted enhancement processing is performed to remove common-mode noise (such as environmental vibration interference). This is done to transform isolated physical quantities into structured multimodal feature vectors, providing data input with strong causal relationships for subsequent impedance feature decoupling in deep neural networks.

[0034] Step S30: Input the multimodal sensing data into the impedance decoupling network for feature extraction to obtain the battery's internal conduction features, liquid transport features, and interface reaction features.

[0035] It should be noted that the impedance decoupling network includes a voltage response encoder, a thermal field encoder, an acoustic encoder, a feature fusion layer, and an impedance decoupling head. The voltage response encoder includes a first convolutional layer, a bidirectional long short-term memory layer, and a first fully connected layer. The thermal field encoder includes a second convolutional layer, a spatial attention layer, and a channel attention layer. The acoustic encoder includes a third convolutional layer, a Mel frequency cepstral coefficient extraction layer, and a second fully connected layer. The impedance decoupling head consists of three decoupling branches connected in parallel. Each decoupling branch consists of a third fully connected layer, a fourth convolutional layer, and a fourth fully connected layer.

[0036] Specifically, the charging micro-response signal, temperature distribution sequence, and vibration signal are simultaneously input into the impedance decoupling network in parallel. The system utilizes internal voltage response encoders, thermal field encoders, and acoustic encoders to specifically extract the temporal fluctuations of the charging micro-response signal, the spatial distribution of the temperature distribution sequence, and the frequency domain features of the vibration signal, respectively. This is done to preserve the unique physical sensitivity of different modes at the low-level data stage. Secondly, the intermediate feature vectors output from the three encoders are converged into a feature fusion layer. This layer automatically identifies the correlation and redundancy between different signals through a cross-modal attention mechanism. For example, it automatically increases the weight of thermal field features under high-temperature conditions and increases the weight of voltage features during the pulse phase. This synthesizes fragmented sensor evidence into a comprehensive feature vector representing the real-time overall picture of the battery. Finally, the fused comprehensive feature vector is input into the impedance decoupling head, which consists of three independent branches. Using a parallel fully connected + convolutional structure, the mixed features are forcibly decomposed and mapped to three physical dimensions, similar to spectral color separation: internal conduction features representing the aging of the metal current collector, interfacial reaction features representing the chemical reaction state of the electrode surface, and liquid transport features representing the electrolyte ion migration capability. This approach, through the rigid constraints of the model structure, achieves an essential transformation from "black box characteristics" to "electrochemical mechanism characteristics," providing a scientific characteristic foundation for subsequent accurate assessment of health.

[0037] Step S40: Based on the neighboring vehicle nodes in the spatiotemporal topology graph, obtain the historical decay trajectory similar to the target vehicle's operating conditions to obtain the transfer learning sample set.

[0038] It should be noted that step S40 includes: querying neighboring vehicle nodes within a preset hop count range in the spatiotemporal topology map, centered on the charging pile node where the target vehicle is located, to obtain a candidate neighbor node set; calculating the similarity of the charging power curve, ambient temperature distribution, and initial state of charge with the target vehicle based on the historical charging records of each vehicle in the candidate neighbor node set, to obtain a multi-dimensional operating condition similarity vector; filtering neighboring vehicles with similarity greater than a preset similarity threshold based on the multi-dimensional operating condition similarity vector, to obtain a highly similar neighbor vehicle set; obtaining the historical health detection records and corresponding loop count sequences of the highly similar neighbor vehicle set, constructing an attenuation trajectory matrix, and obtaining a transfer learning sample set.

[0039] Specifically, firstly, in the spatiotemporal topology graph, with the charging pile node where the target vehicle is located as the center, neighboring vehicle nodes within a preset number of hops (e.g., 2 hops) are queried to obtain a candidate neighbor node set. This set includes not only directly connected 1-hop neighbors (vehicles at the same charging pile during the same time or adjacent charging piles at the same time), but also 2-hop neighbors (vehicles at adjacent charging piles at different times). This expands the candidate sample size and solves the problem of insufficient similar vehicles in sparse charging scenarios. For example, when a vehicle is charging at a charging station on the edge of a city, there are only 2 vehicles within a 1-hop range. After expanding to 2 hops, 12 vehicles from 3 other charging stations within a 3-kilometer radius are included, increasing the candidate set from 2 to 14. Secondly, for the similarity of charging power curves, a dynamic time warping algorithm is used to calculate the optimal matching distance between two power-time series. For example, the Euclidean distance is calculated after nonlinearly aligning the three-stage curve of the target vehicle ("30 kW constant power - power reduction - trickle charging") with the curve of the neighboring vehicle. For the similarity of ambient temperature distribution, temperature points throughout the charging process are extracted to form a temperature distribution vector, and the Wasserstein distance is calculated to capture the overall distribution difference rather than single-point deviation. For example, the difference between a constant temperature of 25 degrees Celsius throughout the process and a temperature change from 20 degrees Celsius in the early stage to 30 degrees Celsius in the later stage is distinguished. For the similarity of the initial state of charge, the absolute reciprocal of the state of charge difference is directly calculated. For example, the similarity between the target vehicle starting to charge at 15% and the neighboring vehicle starting to charge at 12% or 18% is 0.95 and 0.83, respectively. Finally, the three normalized similarities are weighted and concatenated into a 3D operating condition similarity vector to comprehensively characterize the electrical, thermodynamic and initial condition consistency of the charging operating conditions. Then, based on the multi-dimensional operating condition similarity vector, neighbor vehicles with similarity greater than a preset similarity threshold (e.g., 0.85) are selected to obtain a set of highly similar neighbor vehicles. A weighted voting mechanism is adopted to require that at least two of the three dimensions of similarity are higher than the threshold and the weighted sum is higher than the threshold. This avoids the situation where a single dimension is abnormally high in similarity but the overall operating conditions are mismatched. For example, if a neighbor vehicle has a power curve similarity of 0.92 but a temperature distribution similarity of only 0.61 and an initial state of charge similarity of 0.88, and the weighted sum of 0.82 is lower than the threshold, it will be excluded. This ensures that the selected neighbors and the target vehicle have truly transferable operating condition comparability.Finally, historical health detection records and corresponding cycle count sequences of highly similar neighbor vehicles are obtained to construct a decay trajectory matrix and obtain a transfer learning sample set. For each highly similar neighbor vehicle, its health detection values ​​and corresponding cumulative cycle counts during the past 20 charging cycles are extracted to form a two-dimensional time series of "cycle count - health". For example, the data of a certain neighbor vehicle is [(100 times, 98.5%), (150 times, 97.2%), ..., (500 times, 89.3%)]. The sequences of all neighbors are aligned and interpolated to a unified grid according to the cycle count and stacked to form a decay trajectory matrix. The rows of the matrix represent different neighbor vehicles, the columns represent the standardized cycle count intervals, and the matrix elements are the corresponding health values. In this way, the historical decay experience of the group is structured into a transfer learning sample set that can be used for comparison with the current state of the target vehicle. This allows new vehicle models or new user vehicles to use the decay patterns of similar working conditions to predict health and check for anomalies.

[0040] Step S50: Input the battery's internal conduction characteristics, liquid transport characteristics, interface reaction characteristics, and transfer learning sample set into the spatiotemporal graph neural network for processing to obtain health characteristics.

[0041] It should be noted that step S50 includes: constructing a spatiotemporal graph neural network, wherein the spatiotemporal graph neural network includes a spatial flow branch, a temporal flow branch, a cross-flow fusion module, a temporal pooling layer, and a fifth fully connected layer; in the spatial flow branch, the battery internal conduction features, liquid transport features, and interface reaction features are used as the initial features of the nodes, and the spatial dependencies between neighboring nodes are learned through a graph attention layer to obtain spatial aggregation features, wherein the graph attention layer includes a first query matrix, a first key matrix, a first value matrix, and a multi-head attention mechanism; in the temporal flow branch, the historical decay in the transfer learning sample set is used as the initial features of the nodes. The trajectory, as a temporal input, learns the temporal dependency of decay evolution through a temporal convolutional network and a gated recurrent unit to obtain temporal evolution features. The temporal convolutional network includes causal convolutional layers and dilated convolutional layers, and the gated recurrent unit includes update gates and reset gates. In the cross-stream fusion module, the interaction weights between spatial aggregation features and temporal evolution features are calculated through a cross-modal attention mechanism to obtain cross-stream fusion features. The cross-modal attention mechanism includes a second query matrix, a second key matrix, and a second value matrix. The cross-stream fusion features are then input into a temporal pooling layer and a fifth fully connected layer for dimensionality compression to obtain health features.

[0042] Specifically, a heterogeneous dual-flow network is first initialized. The 32-dimensional internal battery conduction features, 32-dimensional liquid transport features, and 32-dimensional interface reaction features are concatenated into a 96-dimensional initial feature vector for each node. The target node features are projected into a query vector using a first query matrix (96×64 dimensional). All neighboring node features are projected into key and value vectors using a first key matrix (96×64 dimensional) and a first value matrix (96×64 dimensional). The dot product of the query vector and the key vector is calculated and normalized using Softmax to obtain the attention weights. For example, the attention weight between the target node and a highly similar neighbor is 0.25, while it is only 0.03 with a less similar neighbor. A multi-head attention mechanism concatenates the outputs of eight independent attention heads into a 512-dimensional spatial aggregated feature, enabling the target vehicle to absorb the impedance feature differences of neighboring vehicles under similar operating conditions and correct for its own detection uncertainties in the spatial flow branch. Then, historical decay trajectories from the transfer learning sample set are synchronously input. The "cycle count - health" sequence of each neighboring vehicle in the decay trajectory matrix is ​​used as input. A causal convolutional layer ensures that only historical information is used to predict future states. An expanded convolutional layer (with expansion factors increasing exponentially by 1, 2, 4, ..., 2) broadens the temporal receptive field, capturing long-term decay trends rather than local fluctuations, outputting 128-dimensional temporal convolutional features. These features are then processed by update gates (controlling the retention rate of historical information) and reset gates (controlling the forgetting rate) in the gated recurrent unit. For example, when a neighbor's recent health drops sharply, the reset gate reduces its historical weight, and the update gate increases the influence of recent samples, outputting 128-dimensional temporal evolution features. This allows the extraction of common patterns and individual differences in decay dynamics from the group's historical trajectories. This approach aims to extract universal patterns of battery decay from the long-term experience of "predecessor vehicles" to obtain temporal evolution features. Then, the spatial aggregation features are projected into query vectors through the second query matrix (512×128 dimensions), and the temporal evolution features are projected into key vectors and value vectors through the second key matrix (128×128 dimensions) and the second value matrix (128×128 dimensions). The correlation weight between a certain impedance feature pattern in the spatial dimension and a certain decay stage in the temporal dimension is calculated. For example, when the interface reaction impedance in the spatial aggregation features increases significantly, the attention weight of the corresponding lithium deposition decay stage in the temporal evolution features is automatically enhanced, and a 256-dimensional cross-current fusion feature is output. This achieves the cross-dimensional correlation between "current impedance state - historical decay trajectory", so that the health assessment has both micro-mechanism interpretability and macro-trend predictability.Finally, the cross-stream fusion features are input into the temporal pooling layer, and the importance weights of each time step are calculated and weighted summed using the attention pooling mechanism. Then, the fifth fully connected layer (256×128 dimensions) is used for nonlinear dimensional compression to obtain 128-dimensional health features. This compresses the rich information of the spatiotemporal dual streams into a compact health representation vector, providing a unified feature basis for the subsequent decoder output of health values, lifetime prediction, and decay mode classification.

[0043] Step S60: Based on the health characteristics, the current health value, the predicted remaining cycle life value, and the degradation type label are output through the decoder network to obtain the power battery health detection result.

[0044] It should be noted that step S60 includes: First, constructing a decoder network. It should be noted that the decoder network includes a health branch, a lifetime prediction branch, and a decay mode classification branch. The health branch includes a sixth fully connected layer, a first deconvolutional layer, and a seventh fully connected layer. The lifetime prediction branch includes an eighth fully connected layer, a second bidirectional long short-term memory layer, and a ninth fully connected layer. The decay mode classification branch includes a tenth fully connected layer, a fourth convolutional layer, and an eleventh fully connected layer.

[0045] Next, the health features are mapped to the health value space through the health branch to obtain the current health evaluation value. Specifically, a health score branch is first constructed, comprising a sixth fully connected layer (128×64 dimensional), a first deconvolutional layer (64×32 dimensional), and a seventh fully connected layer (32×1 dimensional). This gradually decodes high-dimensional abstract features into scalar health score values. Then, the 128-dimensional health score features are input into the sixth fully connected layer for initial dimensionality reduction. A ReLU activation function is used to introduce non-linearity, outputting a 64-dimensional intermediate representation. This 64-dimensional intermediate representation is then input into the first deconvolutional layer for feature upsampling and detail restoration. Batch normalization is used to stabilize the training process, outputting a 32-dimensional refined feature. Finally, the 32-dimensional refined feature is input into the seventh fully connected layer. A Sigmoid activation function constrains the output to the 0-1 range, and multiplying by 100 yields the current health score (in percentage form). For example, an output of 0.876 maps to 87.6%. This achieves end-to-end decoding from microscopic impedance features and spatiotemporal correlation learning to macroscopic health score values, giving the evaluation results clear physical meaning and engineering usability.

[0046] Next, the health features are output as a sequence through the lifespan prediction branch to obtain the predicted remaining cycle lifespan. Specifically, the lifespan prediction branch is first constructed, containing an eighth fully connected layer (128×64-dimensional), a second bidirectional long short-term memory layer (64×32-dimensional hidden units), and a ninth fully connected layer (32×1-dimensional). The health features are used as the initial state input, and a sequence of future cycle counts is generated point-by-point through autoregression. Then, the 128-dimensional health features are input to the eighth fully connected layer for dimensionality adaptation, outputting a 64-dimensional state initialization vector. This sets the initial hidden state and cell state of the long short-term memory network, allowing the network to remember the decay stage information implied by the current health level. The initialized state is then input to the second bidirectional long short-term memory layer, which uses forward units to capture the natural decay trend of health and backward units to verify the consistency of historical trajectories. A 32-dimensional hidden state is output at each time step and mapped to the predicted health value through the ninth fully connected layer. Then, a sequence is generated... The column generation method concatenates the predicted health value of the current step with the hidden state and feeds it back to the next step input. This process iterates until the predicted health value drops to a preset lifespan termination threshold (80%). The number of iterations required is the remaining number of cycles. For example, starting from the current health value of 87.6%, if step 1 predicts 86.2%, step 2 predicts 84.8%, ..., step 150 predicts 79.8%, then the output remaining cycle lifespan prediction value is 150 cycles. This achieves a complete trajectory prediction from the current state to the end of lifespan, rather than a single-point estimation, giving the prediction results process interpretability and a basis for uncertainty quantification. Finally, the predicted complete decay trajectory is compared and verified with the historical neighbor trajectories. When the slope of the predicted trajectory deviates significantly from the group's experience, a confidence downgrade prompt is triggered. This utilizes the group's prior knowledge to constrain the rationality of individual predictions and avoids extreme abnormal predictions from misleading operation and maintenance decisions.

[0047] Then, the health characteristics are mapped to the degradation mode category space through the degradation mode classification branch to obtain degradation mode classification labels. It should be noted that the degradation mode classification labels include normal degradation mode, lithium deposition degradation mode, electrolyte drying degradation mode, and active material loss degradation mode. Normal degradation mode refers to the natural aging degradation type of the power battery during normal charging and discharging due to conventional chemical and physical changes, without abnormal side reactions, and the degradation rate conforms to normal usage patterns. Lithium deposition degradation mode is a type of degradation caused by abnormal deposition of lithium metal on the electrode surface during charging, leading to a decrease in battery performance; it belongs to non-natural aging degradation caused by abnormal side reactions. Electrolyte drying degradation mode is a type of degradation caused by the gradual consumption of electrolyte, reduced water content, or decreased fluidity, resulting in decreased ion transport efficiency and thus battery performance degradation; it belongs to aging caused by internal dielectric loss. Active material loss degradation mode is a type of degradation caused by the shedding, dissolution, or structural failure of active materials in the power battery electrodes, leading to a reduction in the effective substances participating in electrochemical reactions and thus a decrease in battery capacity and power. Specifically, firstly, an attenuation mode classification branch is constructed, comprising a tenth fully connected layer (128×64 dimensional), a third one-dimensional convolutional layer (64×32 dimensional, kernel length 3), and an eleventh fully connected layer (32×4 dimensional), thereby decoding the health features into probability distributions of four attenuation modes. Then, the 128-dimensional health features are input into the tenth fully connected layer for preliminary transformation, outputting a 64-dimensional intermediate classification representation. Next, the 64-dimensional intermediate representation is input into the third one-dimensional convolutional layer, where three convolutional kernels of length 3 slide along the feature dimensions to capture local combinations of different impedance features in the health features, such as simultaneously detecting the local feature combination of "high interface reaction impedance + low liquid transport impedance," outputting a 32-dimensional convolutional classification feature. Finally, the 32-dimensional convolutional classification feature is input into the eleventh fully connected layer, through a soft... The max activation function outputs a 4-dimensional probability vector, corresponding to the confidence levels of the normal decay mode, lithium deposition decay mode, electrolyte drying decay mode, and active material loss decay mode, respectively. For example, an output of [0.15, 0.62, 0.18, 0.05] indicates a lithium deposition decay mode with a confidence level of 0.62. Finally, the classification confidence level is calculated based on the entropy value of the probability vector. When the difference between the highest probability and the second highest probability is less than a preset confidence threshold (e.g., 0.3), it is marked as a "mixed decay mode" and triggers multimodal data verification. For example, a fuzzy distribution of [0.35, 0.30, 0.28, 0.07] will prompt the acquisition of higher resolution micro-pulse response data. This enables refined diagnosis and uncertainty perception of decay modes, providing a clear mechanistic direction for subsequent optimization of differentiated charging and discharging strategies.

[0048] Finally, the health status test results of the power battery are obtained based on the current health evaluation value, the predicted remaining cycle life value, and the degradation mode classification label. Specifically, the system first encapsulates the current health assessment value, remaining cycle life prediction value, and degradation mode classification labels in a structured manner to form a three-dimensional detection result that includes numerical evaluation, trend prediction, and mechanism diagnosis. Then, it retrieves the corresponding charge and discharge strategy optimization suggestion library based on the degradation mode classification labels. For example, for lithium deposition degradation mode, it generates targeted operation and maintenance guidance such as "suggesting reducing the charging cut-off voltage to 4.1 volts, limiting the charging rate to 0.5 degrees Celsius, and increasing the resting interval to 30 minutes". Next, it compares the health assessment value with a preset health threshold (e.g., 80%). When the assessment value is lower than the threshold, a yellow warning is triggered. When it is lower than the life end threshold (e.g., 70%), a red warning is triggered, realizing graded alarms. Finally, it integrates the detection results, trend prediction curves, degradation mode visualization heatmaps, and strategy optimization suggestions to generate a standardized detection report. This report is pushed to users through the vehicle terminal and simultaneously uploaded to the cloud-based battery asset management system, supporting V2G charge and discharge scheduling optimization, battery residual value assessment, and tiered utilization decisions, forming a complete closed loop from online detection to operation and maintenance actions.

[0049] This embodiment constructs a charging spatiotemporal topology map during the charging process, collects multimodal sensor data, and extracts three types of internal features of the battery through an impedance decoupling network containing multiple encoders and decoupling branches. It constructs transfer samples based on similar operating conditions of neighboring vehicles, obtains health features through a spatiotemporal graph neural network, and finally outputs the power battery health detection results by the decoder. It is adapted to complex charging scenarios, realizes non-invasive online detection, improves the accuracy and reliability of power battery health detection, can accurately identify the degradation mechanism, and provides support for the full life cycle management of batteries.

[0050] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in Embodiment 1 above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 2 The online multi-parameter detection method for power batteries during the charging process, step S30, further includes steps S201 to S206: Step S201: Construct an impedance decoupling network.

[0051] It should be noted that the impedance decoupling network includes a voltage response encoder, a thermal field encoder, an acoustic encoder, a feature fusion layer, and an impedance decoupling head. Specifically, the voltage response encoder includes a first convolutional layer, a bidirectional long short-term memory layer, and a first fully connected layer. The first convolutional layer uses 32 one-dimensional convolutional kernels of length 3 sliding along the time dimension with a stride of 1 to extract local gradient changes in the voltage transient response curve, such as capturing the slope of the pulse rising edge and the curvature of the falling edge, and outputting a 128-dimensional time-frequency feature map. The bidirectional long short-term memory layer contains 64 forward units and 64 backward units. The forward units process the relaxation process of the voltage from the pulse start point to the steady state in forward time order, and the backward units capture the initial polarization features from the steady state back to the pulse start point in reverse time order, such as distinguishing the voltage recovery modes of fast double-layer charging (millisecond level) and slow diffusion control (second level), and outputting a 128-dimensional bidirectional hidden state. The first fully connected layer compresses the 128-dimensional hidden state into a 64-dimensional voltage hidden vector, thereby extracting the kinetic features related to solid-phase lithium-ion diffusion from the macroscopic voltage response. The thermal field encoder consists of a second convolutional layer, a spatial attention layer, and a channel attention layer. The second convolutional layer uses 16 3×3 two-dimensional convolutional kernels to scan the thermal field matrix of 16 measurement points, extracting local temperature gradients and heat flow directions. For example, it identifies the heat conduction path from the tab hotspot to the center of the cell and outputs a 32-dimensional thermal field feature map. The spatial attention layer calculates the deviation weight of each measurement point from the global average temperature. For example, it increases the weight of the tab region with a temperature 8 degrees Celsius higher than the average to 0.35 and decreases it to 0.05 in the low-temperature edge region, achieving adaptive focusing on the thermal anomaly region. The channel attention layer performs weighted filtering of feature channels with different temperature ranges. For example, it enhances the response of high-temperature channels to the risk of thermal runaway and outputs a 64-dimensional thermal field latent vector, thereby extracting the heat transfer characteristics related to the liquid electrolyte transport resistance from the thermal field spatial distribution. The acoustic encoder comprises a third convolutional layer, a Mel-frequency cepstral coefficient extraction layer, and a second fully connected layer. The third convolutional layer uses 64 one-dimensional convolutional kernels of length 5 to extract the local harmonic structure of the acoustic spectrum, such as capturing the frequency band energy distribution of the 3200 Hz characteristic peak, and outputting a 256-dimensional acoustic feature map. The Mel-frequency cepstral coefficient extraction layer maps the spectrum to 40 Mel filter banks, simulating the low sensitivity of the human ear to high-frequency signals. After logarithmic compression, it calculates the discrete cosine transform, retaining the first 13 coefficients and their first-order differences, for example, characterizing the dynamic changes in the amplitude attenuation of characteristic peaks and the generation of new peaks. The second fully connected layer outputs a 64-dimensional acoustic latent vector, thereby extracting the acoustic features related to the precipitation of interface side reaction gases from the acoustic spectrum. The impedance decoupling head consists of three parallel decoupling branches, each composed of a third fully connected layer, a fourth convolutional layer, and a fourth fully connected layer.The third fully connected layer maps the 192-dimensional fused latent vector to a 96-dimensional branch-specific space. The fourth convolutional layer extracts scale-specific features with different configurations. The first branch uses 8 convolutional kernels of length 2 to extract low-frequency trend features to characterize solid-phase diffusion impedance. The second branch uses 16 convolutional kernels of length 3 to extract mid-frequency fluctuation features to characterize liquid-phase transport impedance. The third branch uses 8 convolutional kernels of length 2 to extract high-frequency transient features to characterize interface reaction impedance. The fourth fully connected layer outputs 32-dimensional internal conduction features, 32-dimensional liquid transport features, and 32-dimensional interface reaction features, respectively. This decouples the macroscopic multimodal response into three interpretable microscopic electrochemical process features, providing well-defined input variables for subsequent degradation mode diagnosis.

[0052] Step S202: Extract time-frequency features from the charging micro-response signal input voltage response encoder in the multimodal sensing data to obtain voltage time-frequency features.

[0053] Specifically, the charging micro-response signal is first digitized at a preset sampling frequency (10 kHz) to obtain a voltage sequence containing 4000 sampling points, including 200 milliseconds before and after the pulse, thus fully capturing the dynamic process of voltage from steady-state rise, pulse sustaining, to relaxation recovery. Then, the voltage sequence is input into the first convolutional layer, where 32 one-dimensional convolutional kernels of length 3 and stride 1 slide along the time dimension. Each kernel learns a local voltage change pattern; for example, the first kernel detects the steep slope of the pulse rising edge, the 15th kernel identifies the exponential decay curvature in the early relaxation phase, and the 32nd kernel captures the small overshoot after steady-state recovery. A feature map with 32 channels × 3998 time steps is output to extract multi-scale local features of the voltage transient response. Next, the convolutional feature map is input into a bidirectional long short-term memory layer, and the forward long... The short-term memory unit processes the voltage from the pulse start point to the pulse end point in forward time sequence, capturing the temporal evolution of voltage from rapid polarization to slow diffusion control. The backward long short-term memory unit traces back from the steady-state end point to the pulse start point in reverse time sequence, identifying the initial voltage drop depth and recovery speed. The outputs of 64 hidden units in each direction are concatenated into a 128-dimensional temporal encoding vector to model the long-term dependence of voltage response and bidirectional contextual information. Finally, the 128-dimensional temporal encoding vector is input into the first fully connected layer, and a 64-dimensional voltage time-frequency feature is output through a linear transformation of the weight matrix (128×64-dimensional) and the introduction of nonlinearity through the ReLU activation function. This compresses the high-dimensional temporal information into a compact feature representation, while retaining the key kinetic fingerprint of the internal electrochemical polarization process of the battery under pulse excitation, providing microscopic mechanism clues in the voltage dimension for subsequent impedance decoupling.

[0054] Step S203: Input the temperature distribution sequence in the multimodal sensing data into the thermal field encoder to locate the thermal anomaly region and extract its features, thereby obtaining the spatial features of the thermal field.

[0055] It should be noted that, firstly, the temperature distribution sequence is expanded along the charging process time axis to obtain a two-dimensional thermal field tensor (time step × measurement point number) containing 16 distributed measurement points, each sampled at a frequency of 1 Hz. For example, a 100-minute charging process forms a 6000×16 thermal field matrix, which fully describes the spatiotemporal evolution of the battery module surface temperature. Then, the thermal field matrix is ​​input into the second convolutional layer, using 16 3×3 two-dimensional convolutional kernels sliding along both the time and measurement point dimensions with a stride of 1. Each convolutional kernel learns a local thermal pattern, for example, the first... The first convolutional kernel identifies the continuous heating trend at a single measurement point, the 8th convolutional kernel captures the thermal conduction gradient between adjacent measurement points, and the 16th convolutional kernel detects synchronous thermal jumps at multiple measurement points, outputting a 16-channel × 5998 time-step × 14-point thermal field feature map to extract the local spatiotemporal correlation features of the thermal field distribution. Then, the thermal field feature map is input into a spatial attention layer, and the deviation of each measurement point from the global average temperature is calculated as the initial weight. For example, if the temperature at the electrode measurement point is 42 degrees Celsius, which deviates from the global average of 35 degrees Celsius by 7 degrees Celsius, the initial weight is set to 0.25. When the temperature at the edge measurement point deviates by 2 degrees Celsius from 33 degrees Celsius, the weight is reduced to 0.08. Then, Softmax normalization is applied to make the sum of the weights of all measurement points equal to 1, achieving adaptive focusing and localization of the thermal anomaly area. Simultaneously, the thermal field feature map is input into the channel attention layer. For each feature channel, a channel descriptor after global average pooling is calculated. Through two fully connected layers (16×4D and 4×16D), the nonlinear interaction between channels is learned, such as enhancing the channel weights sensitive to the high-temperature range and suppressing the background channels at normal temperature. The output is a 16-dimensional channel weight vector, which is then sequentially applied to the feature map. The features are multiplied by the channel attention; then the spatial attention-weighted features are added element-wise and fused with the channel attention-weighted features to output an enhanced thermal field feature of 5998 time steps × 14 measurement points × 16 channels; finally, the enhanced thermal field feature is globally averaged along the time and measurement point dimensions, compressed into a 16-dimensional statistical feature, and then input into the second fully connected layer (16 × 64 dimensions) for mapping and expansion, outputting a 64-dimensional thermal field spatial feature, thereby capturing the location, intensity and evolution pattern of the thermal anomaly region, providing clues to the liquid phase transport resistance in the thermal dimension for subsequent impedance decoupling.

[0056] Step S204: The vibration signal in the multimodal sensing data is input into the acoustic encoder to extract acoustic features and obtain acoustic frequency domain features.

[0057] It should be noted that, firstly, the vibration signal is digitized at a preset sampling frequency (20 kHz) to obtain an acoustic time series containing the entire charging process. For example, a 100-minute charging process generates a long sequence of 120 million sampling points. This sequence is then divided into overlapping frames using a sliding time window (window length 1024 points, step size 512 points). Each frame is weighted by a Hanning window and then subjected to a Fast Fourier Transform to obtain a 513-dimensional spectral vector. This converts the time-domain vibration signal into a frequency-domain representation, revealing the acoustic characteristic frequency components induced by the electrochemical reaction. Then, the frequency... The spectral sequence is input into the third convolutional layer, where 64 one-dimensional convolutional kernels of length 5 and stride 1 slide along the frequency dimension. Each kernel learns a local spectral pattern; for example, the first kernel detects low-frequency mechanical vibrations in the 100-300 Hz range, the 32nd kernel identifies characteristic peak energy accumulations near 3200 Hz, and the 64th kernel captures high-frequency noise floor above 5000 Hz. The output is a 64-channel × 509-frequency-step convolutional feature map, thus extracting the multi-scale local structure of the acoustic spectrum. The convolutional feature map is then input... The Mel frequency cepstral coefficient extraction layer uses 40 triangular Mel filter banks to cover the range from 20 Hz to 8000 Hz, simulating the nonlinear perception characteristics of the human ear that is sensitive to low frequencies and insensitive to high frequencies. The spectral energy of each filter bank is weighted and summed, and the natural logarithm is taken to compress the dynamic range. Then, a discrete cosine transform is performed to decorrelate the features, retaining the first 13 Mel frequency cepstral coefficients and their first-order differences (26 dimensions in total). For example, the second-order coefficients characterize the overall spectral tilt, the eighth-order coefficients reflect the mid-frequency peak complexity, and the first-order difference captures characteristic peaks. The acoustic temporal features are output as the drift velocity changes over time, resulting in a 26-dimensional × time frame. These 26-dimensional acoustic temporal features are then input into a second fully connected layer (26 × 64-dimensional). Through linear transformation of the weight matrix and the introduction of nonlinear saturation characteristics via the Tanh activation function, a 64-dimensional acoustic frequency domain feature is output. This compresses the acoustic vibration signal into a compact acoustic fingerprint representation while retaining the frequency fingerprints of microscopic electrochemical processes such as electrolyte vaporization bubble rupture and lithium ion insertion / extraction lattice vibrations. This provides acoustic-dimensional clues to the interface reaction activity for subsequent impedance decoupling.

[0058] Step S205: Input the voltage time-frequency features, thermal field spatial features, and acoustic frequency domain features into the feature fusion layer and perform cross-modal attention weighting to obtain the fusion vector.

[0059] It should be noted that the 64-dimensional voltage time-frequency features, 64-dimensional thermal field spatial features, and 64-dimensional acoustic frequency domain features are projected onto a unified feature space through independent fully connected layers (64×64-dimensional) to obtain the projected voltage features, thermal field features, and acoustic features, thereby eliminating the dimensional differences and semantic gaps between the three modal features. Then, the pairwise similarity matrices between the three modal features are calculated. For example, the dot product similarity between voltage and thermal field features is 0.78, voltage and acoustic is 0.45, and thermal field and acoustic is 0.62. Softmax normalization is used to convert the similarity into cross-modal attention weights to measure the intrinsic correlation strength between signals from different physical fields. Then, using voltage features as queries and thermal field features as keys and values, voltage-thermal field interaction features are calculated through scaled dot product attention. For example, when voltage diffusion features and thermal resistance features are highly correlated, the attention weight is increased to 0.6 and the pairs are aggregated. Similarly, based on the thermal field information, voltage-acoustic interaction features and thermal field-acoustic interaction features are calculated to obtain three sets of 64-dimensional cross-modal interaction vectors. Then, the three sets of interaction vectors are residually connected with the original three-modal features. The retention ratio of the new fused information and the original information is controlled by a gating mechanism (Sigmoid gating weight 0.7), and a 192-dimensional primary fused vector is output. Finally, the 192-dimensional primary fused vector is input into the fully connected layer (192×192-dimensional) at the end of the fusion layer. Through the layer normalization stabilization training process, and then the nonlinearity is introduced by the ReLU activation function, a 192-dimensional fused vector is output. This achieves complementary enhancement and noise suppression of the electro-thermal-acoustic three-modal signals, so that the dynamic information of voltage response, the transmission resistance information of thermal field distribution, and the interface activity information of acoustic spectrum are mutually calibrated and verified in a unified latent space, providing a complete and low-redundancy fusion representation for subsequent impedance decoupling.

[0060] Step S206: Output the fused vector input impedance decoupling head to obtain the battery's internal conduction characteristics, liquid transport characteristics, and interface reaction characteristics.

[0061] It should be noted that the 192-dimensional fused vector is first processed in parallel through three decoupled branches. Each decoupled branch consists of a third fully connected layer, a fourth convolutional layer, and a fourth fully connected layer. This allows for the learning of specific feature transformations for three different scales of electrochemical processes: solid-phase diffusion, liquid-phase transport, and interfacial reactions. In the battery internal conduction feature branch, the third fully connected layer (192×96-dimensional) maps the fused vector to a 96-dimensional intermediate space, and the fourth convolutional layer uses eight one-dimensional convolutional kernels of length 2 with a stride of 1 to extract low-frequency trend features. For example, to capture the gradient component in the fusion vector that characterizes the slow diffusion of lithium ions within the active material particles, an 8-channel × 95-dimensional feature map is output. The fourth fully connected layer (760 × 32-dimensional) flattens and compresses the feature map into a 32-dimensional internal battery conduction feature, thus characterizing the solid-phase lithium ion diffusion resistance. In the liquid transport feature branch, the third fully connected layer (192 × 96-dimensional) is also mapped to 96 dimensions. The fourth convolutional layer uses 16 one-dimensional convolutional kernels of length 3 to extract mid-frequency fluctuation features, such as capturing the electrolyte in the porous electrode in the fusion vector. The oscillating components representing the impeded ion migration within the pores are output as a 16-channel × 94-dimensional feature map. The fourth fully connected layer (1504 × 32-dimensional) is compressed into a 32-dimensional liquid transport feature map to characterize the resistance to ion transport in the liquid phase. In the interfacial reaction feature branch, the third fully connected layer (192 × 96-dimensional) is mapped to 96 dimensions. The fourth convolutional layer uses eight 1-dimensional convolutional kernels of length 2 to extract high-frequency transient features, such as capturing the fast transient components representing the charge transfer reaction kinetics and solid electrolyte interfacial membrane impedance in the fusion vector, outputting an 8-channel × 95-dimensional feature map. The fourth fully connected layer (760×32D) is compressed into 32-dimensional interface reaction features to characterize the activity of the electrochemical reaction interface. Finally, the 32-dimensional features output from the three branches are spliced ​​into a 96-dimensional impedance feature vector, or a single branch feature can be independently called according to the needs of downstream tasks. This decouples the macroscopic multimodal response into three interpretable microscopic electrochemical process features, enabling subsequent health assessments to distinguish different decay mechanisms such as lithium deposition (dominated by interface reaction), electrolyte drying (dominated by liquid transport), and loss of active material (dominated by internal conduction).

[0062] This embodiment constructs an impedance decoupling network containing multiple encoders, a feature fusion layer, and three parallel decoupling branches. Multimodal sensing data is then input separately to the corresponding encoders to extract voltage time-frequency, thermal field spatial, and acoustic frequency domain features. The feature fusion layer performs cross-modal weighting to obtain a fused vector, which is then input to the impedance decoupling head to decouple and output three types of internal battery features. By specifically adapting the network architecture to the multimodal data feature extraction requirements, cross-modal fusion uncovers feature correlations, and precise decoupling achieves effective feature separation, providing high-quality core battery features for subsequent detection.

[0063] Based on the first embodiment of this application, this application also provides an online detection system for multiple parameters of a power battery during the charging process. Please refer to [link / reference]. Figure 3 The system includes: The spatiotemporal topology construction module 10 is used to construct a charging spatiotemporal topology network with charging piles as spatial nodes and charging periods as time scales during the charging process, thereby obtaining a spatiotemporal topology map.

[0064] The multimodal acquisition module 20 is used to acquire multimodal sensing data of the target vehicle based on the spatiotemporal topology map.

[0065] Impedance decoupling module 30 is used to input multimodal sensing data into impedance decoupling network for feature extraction, thereby obtaining battery internal conduction characteristics, liquid transport characteristics, and interface reaction characteristics.

[0066] The transfer learning module 40 is used to obtain historical decay trajectories similar to the target vehicle's operating conditions based on neighboring vehicle nodes in the spatiotemporal topology graph, and to obtain a transfer learning sample set.

[0067] The spatiotemporal graph neural network module 50 is used to input the battery's internal conduction features, liquid transport features, interface reaction features, and transfer learning sample set into the spatiotemporal graph neural network for processing to obtain health features.

[0068] The result module 60 is used to output the current health value, the predicted remaining cycle life value, and the degradation type label through the decoder network based on the health characteristics, so as to obtain the power battery health detection result.

[0069] The online multi-parameter detection system for power batteries during the charging process provided in this application, employing the online multi-parameter detection method for power batteries during the charging process described in the above embodiments, can solve the technical problem of how to achieve high-precision health detection of power batteries in complex charging scenarios. Compared with the prior art, the beneficial effects of the online multi-parameter detection system for power batteries during the charging process provided in this application are the same as those of the online multi-parameter detection method for power batteries during the charging process provided in the above embodiments, and other technical features of the online multi-parameter detection system for power batteries during the charging process are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0070] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0071] This application provides an online multi-parameter detection device for power batteries during the charging process. The online multi-parameter detection device for power batteries during the charging process includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the online multi-parameter detection method for power batteries during the charging process described in Embodiment 1 above.

[0072] The following is for reference. Figure 4 This document illustrates a structural schematic diagram of a power battery multi-parameter online detection device suitable for implementing embodiments of this application. The power battery multi-parameter online detection device in the charging process embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (Personal Digital Assistants), PMPs (Portable Media Players), and in-vehicle terminals (e.g., in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. Figure 4 The online multi-parameter detection device for the charging process of the power battery shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0073] like Figure 4As shown, the online multi-parameter detection device for the charging process of a power battery may include a processing unit 1001 (e.g., a central processing unit, a graphics processor, etc.), which can perform various appropriate actions and processes according to the program stored in the read-only memory (ROM) 1002 or the program loaded from the storage device 1003 into the random access memory (RAM) 1004. The RAM 1004 also stores various programs and data required for the operation of the online multi-parameter detection device for the charging process of a power battery. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the online multi-parameter monitoring device for the charging process to wirelessly or wiredly communicate with other devices to exchange data. Although various online multi-parameter monitoring devices for the charging process are shown in the figures, it should be understood that it is not required to implement or possess all of them. More or fewer may be implemented alternatively.

[0074] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0075] The online multi-parameter detection device for power batteries during the charging process provided in this application, employing the online multi-parameter detection method for power batteries during the charging process described in the above embodiments, can solve the technical problem of how to achieve high-precision health detection of power batteries in complex charging scenarios. Compared with the prior art, the beneficial effects of the online multi-parameter detection device for power batteries during the charging process provided in this application are the same as those of the online multi-parameter detection method for power batteries during the charging process provided in the above embodiments, and other technical features in this online multi-parameter detection device for power batteries during the charging process are the same as those disclosed in the method of the previous embodiment, and will not be repeated here.

[0076] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0077] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0078] This application provides a computer-readable medium having computer-readable program instructions (i.e., a computer program) stored thereon, which are used to execute the online multi-parameter detection method for power batteries during the charging process in the above embodiments.

[0079] The computer-readable medium provided in this application may be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor devices, or any combination thereof. More specific examples of computer-readable media may include, but are not limited to: electrical connections with one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable medium may be any tangible medium containing or storing a program that can be executed by instructions, used by a device, or used in conjunction with it. The program code contained on the computer-readable medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0080] The aforementioned computer-readable medium may be included in the online multi-parameter detection device for power batteries during the charging process; or it may exist independently and not be assembled into the online multi-parameter detection device for power batteries during the charging process.

[0081] The aforementioned computer-readable medium carries one or more programs that, when executed by the online multi-parameter detection device for the charging process power battery, enable the device to write computer program code for performing the operations of this application in one or more programming languages ​​or a combination thereof. These programming languages ​​include object-oriented programming languages—such as Java, Smalltalk, and C++—and conventional procedural programming languages—such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0082] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and computer program products according to various embodiments of this application. In this regard, all blocks in the flowcharts or block diagrams may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that all blocks in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using dedicated hardware-based implementations that perform the specified functions or operations, or using a combination of dedicated hardware and computer instructions.

[0083] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0084] The readable medium provided in this application is a computer-readable medium, which stores computer-readable program instructions (i.e., a computer program) for executing the above-described online multi-parameter detection method for power batteries during the charging process. This solves the technical problem of how to achieve high-precision health detection of power batteries in complex charging scenarios. Compared with the prior art, the beneficial effects of the computer-readable medium provided in this application are the same as those of the online multi-parameter detection method for power batteries during the charging process provided in the above embodiments, and will not be repeated here.

[0085] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the online multi-parameter detection method for power batteries during the charging process as described above.

[0086] The computer program product provided in this application can solve the technical problem of how to achieve high-precision health detection of power batteries in complex charging scenarios. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the online multi-parameter detection method for power batteries during the charging process provided in the above embodiments, and will not be repeated here.

[0087] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for online detection of multiple parameters of a power battery during the charging process, characterized in that, The method includes: During the charging process, a spatiotemporal topology network for charging is constructed with charging piles as spatial nodes and charging periods as time scales, resulting in a spatiotemporal topology map. Collect multimodal sensor data of the target vehicle; The multimodal sensing data is input to the impedance decoupling network for feature extraction to obtain the battery's internal conduction characteristics, liquid transport characteristics, and interface reaction characteristics. Based on the neighboring vehicle nodes in the spatiotemporal topology graph, historical attenuation trajectories similar to the operating conditions of the target vehicle are obtained to obtain a transfer learning sample set. The battery's internal conduction characteristics, liquid transport characteristics, interface reaction characteristics, and transfer learning sample set are input into a spatiotemporal graph neural network for processing to obtain health characteristics; Based on the health characteristics, the current health value, the predicted remaining cycle life value, and the degradation type label are output through the decoder network to obtain the power battery health detection result.

2. The method as described in claim 1, characterized in that, The step of constructing a spatiotemporal topology network for charging, with charging piles as spatial nodes and charging periods as time scales, to obtain a spatiotemporal topology map during the charging process includes: Obtain the geographical coordinates and grid access capacity of all charging piles within the target area to obtain the spatial attribute set of the charging piles; Based on historical charging records, the frequency of vehicle access and power load curves of each charging pile within a preset time period are statistically analyzed to obtain the charging pile time attribute set. Using charging piles as graph nodes and the connection relationships of charging piles with geographical distance less than a preset distance threshold and the same power grid phase as graph edges, a spatial adjacency matrix is ​​constructed to obtain a spatial graph structure; Using charging vehicles at the same charging pile at different times as time-series nodes and vehicles with the same battery model and overlapping state of charge intervals as time-series edges, a time correlation matrix is ​​constructed to obtain a time graph structure. The spatial graph structure and the temporal graph structure are fused through graph product operation to obtain a spatiotemporal topological graph.

3. The method as described in claim 1, characterized in that, The steps for collecting multimodal sensing data of the target vehicle include: During the constant current charging phase, a charging micro-pulse excitation with a preset amplitude and preset duration is inserted, and the voltage transient response curves before and after the pulse are collected to obtain the charging micro-response signal. A temperature distribution sequence is obtained by collecting the two-dimensional thermal field distribution during the charging process using temperature sensors arranged on the surface of the battery module at a preset sampling frequency. Acoustic sensors attached to the battery casing are used to collect acoustic vibration signals generated by electrolyte vaporization and ion migration during charging, and the vibration signals are obtained by spectrum analysis. The charging micro-response signal, the temperature distribution sequence, and the vibration signal are aligned according to timestamps and enhanced with intermodal cross-correlation to obtain multimodal sensing data.

4. The method as described in claim 1, characterized in that, The step of extracting features from the multimodal sensing data input impedance decoupling network to obtain the battery's internal conduction features, liquid transport features, and interface reaction features includes: An impedance decoupling network is constructed, wherein the impedance decoupling network includes a voltage response encoder, a thermal field encoder, an acoustic encoder, a feature fusion layer, and an impedance decoupling head. The voltage response encoder includes a first convolutional layer, a bidirectional long short-term memory layer, and a first fully connected layer. The thermal field encoder includes a second convolutional layer, a spatial attention layer, and a channel attention layer. The acoustic encoder includes a third convolutional layer, a Mel frequency cepstral coefficient extraction layer, and a second fully connected layer. The impedance decoupling head consists of three decoupling branches in parallel, and each decoupling branch consists of a third fully connected layer, a fourth convolutional layer, and a fourth fully connected layer. The charging micro-response signal in the multimodal sensing data is input to the voltage response encoder for time-frequency feature extraction to obtain the voltage time-frequency features; The temperature distribution sequence in the multimodal sensing data is input into the thermal field encoder to locate the thermal anomaly region and extract its features, thereby obtaining the spatial features of the thermal field. The vibration signal in the multimodal sensing data is input into an acoustic encoder to extract acoustic features, thereby obtaining acoustic frequency domain features; The voltage time-frequency features, the thermal field spatial features, and the acoustic frequency domain features are input into the feature fusion layer and subjected to cross-modal attention weighting to obtain a fusion vector. The fusion vector input impedance decoupling head is output separately to obtain the battery's internal conduction characteristics, liquid transport characteristics, and interface reaction characteristics.

5. The method as described in claim 1, characterized in that, The step of obtaining a transfer learning sample set by acquiring historical attenuation trajectories similar to the target vehicle's operating conditions based on neighboring vehicle nodes in the spatiotemporal topology graph includes: In the spatiotemporal topology graph, with the charging pile node where the target vehicle is located as the center, the neighboring vehicle nodes within a preset number of hops are queried to obtain a set of candidate neighbor nodes. Based on the historical charging records of each vehicle in the candidate neighbor node set, the similarity of the charging power curve, the similarity of the ambient temperature distribution, and the similarity of the initial state of charge with the target vehicle are calculated to obtain a multi-dimensional operating condition similarity vector. Based on the multi-dimensional working condition similarity vector, neighbor vehicles with similarity greater than a preset similarity threshold are selected to obtain a set of highly similar neighbor vehicles; Obtain the historical health detection records and corresponding loop count sequences of the highly similar neighbor vehicle set, construct the decay trajectory matrix, and obtain the transfer learning sample set.

6. The method as described in claim 1, characterized in that, The step of inputting the battery's internal conduction characteristics, liquid transport characteristics, interface reaction characteristics, and transfer learning sample set into a spatiotemporal graph neural network for processing to obtain health characteristics includes: Construct a spatiotemporal graph neural network, wherein the spatiotemporal graph neural network includes a spatial flow branch, a temporal flow branch, a cross-flow fusion module, a temporal pooling layer, and a fifth fully connected layer; In the spatial flow branch, the battery internal conduction characteristics, the liquid transport characteristics, and the interface reaction characteristics are used as the initial node characteristics. The spatial dependency relationship between neighboring nodes is learned through the graph attention layer to obtain the spatial aggregation characteristics. The graph attention layer includes a first query matrix, a first key matrix, a first value matrix, and a multi-head attention mechanism. In the temporal branch, the historical decay trajectory in the transfer learning sample set is used as the temporal input. The temporal dependency of decay evolution is learned through a temporal convolutional network and a gated recurrent unit to obtain temporal evolution features. The temporal convolutional network includes causal convolutional layers and dilated convolutional layers, and the gated recurrent unit includes an update gate and a reset gate. In the cross-stream fusion module, the interaction weights between the spatial aggregation features and the temporal evolution features are calculated through a cross-modal attention mechanism to obtain cross-stream fusion features, wherein the cross-modal attention mechanism includes a second query matrix, a second key matrix, and a second value matrix; The cross-stream fusion features are input into the temporal pooling layer and the fifth fully connected layer for dimensionality compression to obtain the health features.

7. The method as described in claim 1, characterized in that, The step of obtaining the power battery health detection result by outputting the current health value, the predicted remaining cycle life value, and the degradation type label through the decoder network based on the health characteristics includes: Construct a decoder network, wherein the decoder network includes a health branch, a lifetime prediction branch, and a decay mode classification branch. The health branch includes a sixth fully connected layer, a first deconvolutional layer, and a seventh fully connected layer. The lifetime prediction branch includes an eighth fully connected layer, a second bidirectional long short-term memory layer, and a ninth fully connected layer. The decay mode classification branch includes a tenth fully connected layer, a fourth convolutional layer, and an eleventh fully connected layer. The health feature is mapped to the health value space through the health branch to obtain the current health evaluation value; The health characteristics are output in a sequence generation manner through the lifespan prediction branch to obtain the remaining cycle life prediction value; The health characteristics are mapped to the decay mode category space through the decay mode classification branch to obtain decay mode classification labels, wherein the decay mode classification labels include normal decay mode, lithium deposition decay mode, electrolyte drying decay mode and active material loss decay mode. The power battery health test result is obtained based on the current health evaluation value, the remaining cycle life prediction value, and the degradation mode classification label.

8. A multi-parameter online detection system for a power battery during the charging process, characterized in that, The system includes: The spatiotemporal topology construction module is used to construct a spatiotemporal topology network for charging, with charging piles as spatial nodes and charging periods as time scales, during the charging process, and obtain a spatiotemporal topology map. The multimodal acquisition module is used to acquire multimodal sensing data of the target vehicle based on the spatiotemporal topology map; The impedance decoupling module is used to input the multimodal sensing data into the impedance decoupling network for feature extraction, thereby obtaining the battery's internal conduction characteristics, liquid transport characteristics, and interface reaction characteristics. The transfer learning module is used to obtain historical decay trajectories similar to the operating conditions of the target vehicle based on neighbor vehicle nodes in the spatiotemporal topology graph, and to obtain a transfer learning sample set. The spatiotemporal graph neural network module is used to input the battery's internal conduction features, liquid transport features, interface reaction features, and transfer learning sample set into the spatiotemporal graph neural network for processing to obtain health features; The results module is used to output the current health value, the predicted remaining cycle life value, and the degradation type label through the decoder network based on the health characteristics, so as to obtain the power battery health detection results.

9. The system as described in claim 8, characterized in that, The system also includes: The impedance decoupling module is further used to construct an impedance decoupling network, wherein the impedance decoupling network includes a voltage response encoder, a thermal field encoder, an acoustic encoder, a feature fusion layer, and an impedance decoupling head. The voltage response encoder includes a first convolutional layer, a bidirectional long short-term memory layer, and a first fully connected layer. The thermal field encoder includes a second convolutional layer, a spatial attention layer, and a channel attention layer. The acoustic encoder includes a third convolutional layer, a Mel-frequency cepstral coefficient extraction layer, and a second fully connected layer. The impedance decoupling head consists of three parallel decoupling branches, each of which consists of a third fully connected layer, a fourth convolutional layer, and a fourth fully connected layer. For the multimodal sensing data... The charging micro-response signal is input to the voltage response encoder for time-frequency feature extraction to obtain voltage time-frequency features; the temperature distribution sequence in the multimodal sensing data is input to the thermal field encoder for thermal anomaly region localization and feature extraction to obtain thermal field spatial features; the vibration signal in the multimodal sensing data is input to the acoustic encoder for acoustic signature feature extraction to obtain acoustic frequency domain features; the voltage time-frequency features, the thermal field spatial features, and the acoustic frequency domain features are input to the feature fusion layer for cross-modal attention weighting to obtain a fusion vector; the fusion vector is input to the impedance decoupling head for output to obtain battery internal conduction features, liquid transport features, and interface reaction features.

10. The system as described in claim 8, characterized in that, The system also includes: The spatiotemporal graph neural network module is further used to construct a spatiotemporal graph neural network, wherein the spatiotemporal graph neural network includes a spatial flow branch, a temporal flow branch, a cross-flow fusion module, a temporal pooling layer, and a fifth fully connected layer; in the spatial flow branch, the battery internal conduction features, the liquid transport features, and the interface reaction features are used as initial node features, and the spatial dependencies between neighboring nodes are learned through a graph attention layer to obtain spatial aggregation features, wherein the graph attention layer includes a first query matrix, a first key matrix, a first value matrix, and a multi-head attention mechanism; in the temporal flow branch, the historical decay trajectories in the transfer learning sample set are used as initial node features. The trace is used as a temporal input. The temporal dependency of decay evolution is learned through a temporal convolutional network and a gated recurrent unit to obtain temporal evolution features. The temporal convolutional network includes causal convolutional layers and dilated convolutional layers, and the gated recurrent unit includes update gates and reset gates. In the cross-stream fusion module, the interaction weights between the spatial aggregation features and the temporal evolution features are calculated through a cross-modal attention mechanism to obtain cross-stream fusion features. The cross-modal attention mechanism includes a second query matrix, a second key matrix, and a second value matrix. The cross-stream fusion features are input into a temporal pooling layer and a fifth fully connected layer for dimensionality compression to obtain health features.

Images