New energy battery fault diagnosis method based on artificial intelligence

Abstract

The invention relates to the technical field of battery fault diagnosis, and discloses a new energy battery fault diagnosis method based on artificial intelligence, and the method comprises the steps: obtaining multi-mode spatio-temporal data of a new energy battery; identifying a plurality of internal physical parameters of the battery by adopting a physical information neural network; and constructing a dynamically evolved state-of-health space manifold based on the physical parameters, and determining the abnormal degree of the battery by calculating the distance from the current parameter point to the manifold. Wherein the geometric attribute of the manifold can dynamically adjust the training process of the physical information neural network, prognosis can be carried out on a fault evolution path based on causal inference, and an excitation signal can be actively generated under specific conditions to carry out closed-loop exploration. According to the invention, through deep fusion of a physical model and data intelligence, accurate diagnosis, traceability and prognosis of early weak faults can be realized without depending on fault samples, and the interpretability, accuracy and perspectiveness of a diagnosis system are significantly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of battery fault diagnosis, and specifically to a new energy battery fault diagnosis method based on artificial intelligence. Background Art

[0002] With the rapid development of industries like new energy vehicles and energy storage power stations, the safety and reliability of lithium-ion batteries, as core energy storage units, have become a key bottleneck restricting the industry's development. Over the long term, battery systems are subject to the combined influence of multiple factors, including electrical, thermal, mechanical, and chemical factors, inevitably leading to performance degradation and even various types of failures. Promptly and accurately diagnosing early, latent faults and predicting their future evolution are core technologies for ensuring the safe operation of battery systems, enabling predictive maintenance, and extending their service life.

[0003] Currently, fault diagnosis techniques for new energy batteries fall into two main streams. One approach is model-based. This approach establishes an electrochemical model or equivalent circuit model of the battery, compares the model output with actual measurements, and uses the residuals to identify faults. These methods offer excellent interpretability, and their diagnostic results are directly linked to the physical and chemical processes within the battery. However, due to the extreme complexity of battery internal reactions, any model is necessarily a simplification of the actual process. This simplification limits the model's accuracy when dealing with variable operating conditions and individual battery differences. Particularly during battery aging, the rigidity of model parameters makes it difficult to track dynamic changes in the battery, thus compromising diagnostic accuracy.

[0004] Another category is the data-driven diagnostic methods that have emerged in recent years. These methods utilize machine learning or deep learning algorithms to directly learn the nonlinear mapping relationship between sensor signals and fault states from massive amounts of historical operating data. Their powerful data fitting capabilities demonstrate high diagnostic accuracy under specific conditions. However, the successful application of these methods often implies a strict prerequisite: a large training dataset containing various types of fault samples covering the entire life cycle. In actual engineering, serious battery failures are inherently low-probability events, and obtaining sufficient and diverse fault sample data is costly and impractical. This strong reliance on scarce fault samples severely limits the generalization and reliability of data-driven methods in practical applications.

[0005] More importantly, most current mainstream data-driven models exist as "black boxes." While they can determine whether something is "normal" or "abnormal," their internal decision-making logic is opaque to users, making it impossible to illuminate the physical root cause of the fault. This significantly undermines engineers' trust in diagnostic results and hinders subsequent fault tracing and repair decision-making.

[0006] Furthermore, existing diagnostic technologies mostly remain at the "post-facto" level, meaning they can only make judgments after fault characteristics become apparent. They generally lack the "prognostic" capability to predict fault development trends. Furthermore, effectively distinguishing between slow performance degradation caused by normal aging and subtle parameter anomalies caused by early fault initiation remains a major unresolved challenge in this field. This directly leads to a high risk of false positives or missed positives in actual operation of existing systems.

[0007] Therefore, this field urgently needs a new battery fault diagnosis technology that can get rid of the dependence on fault samples, enhance model interpretability, accurately distinguish between aging and early failures, and have the ability to predict fault evolution in a forward-looking manner. Summary of the Invention

[0008] The technical problem to be solved by the present invention is that the existing artificial intelligence-based battery fault diagnosis method usually relies on a large number of fault samples for training, the model has poor interpretability, it is difficult to detect early weak faults, and it is impossible to effectively predict the future evolution trend of the fault.

[0009] To solve the above technical problems, the present invention provides a new energy battery fault diagnosis method and system based on artificial intelligence. This solution constructs a parameter identification model with deep integration of physical information and a dynamically evolving health state space. It can not only accurately diagnose faults, but also trace the source of faults, predict their evolution, and conduct closed-loop active exploration, thereby significantly improving the accuracy, explainability, and foresight of diagnosis.

[0010] A first aspect of the present invention provides a new energy battery fault diagnosis method based on artificial intelligence, the method comprising: Obtain multimodal spatiotemporal data of new energy batteries; Based on the multimodal spatiotemporal data, a physical information neural network is used to identify multiple physical parameters of the new energy battery by minimizing a composite loss function including data fidelity loss and physical law loss; Based on the multiple physical parameters, a health state space manifold is constructed, and the abnormality degree of the new energy battery is determined by calculating the distance from a current parameter point to the health state space manifold.

[0011] In some optional embodiments, to achieve adaptability and robustness in the identification process, the physical law loss in the composite loss function is accompanied by an adaptive regularization weight. The value of the adaptive regularization weight is dynamically adjusted based on the geometric properties of the current parameter point on the health state space manifold.

[0012] Specifically, the composite loss function can be expressed as: in, is the composite loss function, For data fidelity loss, Loss due to physical laws, is the fixed weight of the data fidelity loss. is the adaptive regularization weight, which is a function that depends on the current parameter point and health state space manifold function of the relationship between them.

[0013] When the current parameter point In the health state space manifold When the curvature or density of the The value of is increased to strengthen the physical law constraints; on the contrary, when it is in a high curvature or sparse area, The value of was adjusted down to improve the fit to the actual measured data.

[0014] In some optional embodiments, in order to effectively distinguish sudden failures from natural battery aging, the health state space manifold is a dynamically evolving manifold. The method further models the continuous transformation of the health state space manifold caused by battery aging based on Lie group theory to update the health state space manifold. Specifically, at any time The health state space manifold The initial health state space manifold Through a transformation operator that describes the aging process Evolved to: ; In this way, the diagnostic benchmark can be dynamically adjusted over the battery life cycle.

[0015] In some optional embodiments, in order to predict the fault evolution trend, the method further includes constructing a causal evolution map, and when the abnormality exceeds a preset fault threshold, a fault evolution path of the new energy battery is predicted based on the causal evolution map. Specifically, the process of constructing the causal evolution map includes: using the time series of multiple physical parameters identified by the physical information neural network, learning the causal relationship and time lag relationship between the multiple physical parameters through a causal inference algorithm to generate the causal evolution map. When an initial fault is detected, the parameter deviation corresponding to the fault can be used as input. To predict its subsequent impact.

[0016] In some optional embodiments, to form a closed-loop diagnosis, the method further includes generating an active diagnostic excitation signal under a preset excitation trigger condition and applying it to the new energy battery. Accordingly, the step of acquiring multimodal spatiotemporal data of the new energy battery includes acquiring response data of the new energy battery under the action of the active diagnostic excitation signal. The preset excitation trigger condition may include a model uncertainty represented by the loss of physical laws exceeding a first threshold, or an abnormality suspicion represented by the abnormality degree exceeding a second threshold.

[0017] Furthermore, in order to make the active diagnosis excitation signal have the best information detection efficiency, its generation process includes calculating by optimizing an information gain function. The optimization problem can be expressed as: st in, is the active diagnostic stimulus signal to be optimized, is an information gain function, which aims to maximize the reduction in model uncertainty and the change in abnormality level expected to be brought about by the active diagnosis excitation signal. is the equivalent internal resistance of the battery pack, is the maximum excitation energy allowed.

[0018] In some optional implementations, to improve the sensitivity of anomaly detection, the distance is the geodesic distance from the current parameter point to the health state space manifold.

[0019] A second aspect of the present invention provides a new energy battery fault diagnosis system based on artificial intelligence, the system comprising: A data acquisition module for acquiring multimodal spatiotemporal data of new energy batteries; a parameter identification module for identifying multiple physical parameters of the new energy battery based on the multimodal spatiotemporal data using a physical information neural network by minimizing a composite loss function including data fidelity loss and physical law loss; A state diagnosis module is used to construct a health state space manifold based on the multiple physical parameters, and determine the abnormality degree of the new energy battery by calculating the distance from a current parameter point to the health state space manifold.

[0020] The technical solution provided by this invention embeds a physical model into the loss function of a neural network, enabling the model to learn the inherent physical laws of the battery from readily available normal operating data, eliminating its reliance on scarce fault samples. By constructing a dynamically evolving health state space manifold, it can effectively distinguish between faults and aging. Furthermore, by introducing causal inference and active excitation, the system is upgraded from a passive diagnostic tool to a closed-loop intelligent system with both prognostic and active exploration capabilities, significantly enhancing the depth and foresight of diagnosis.

[0021] The present invention provides a new energy battery fault diagnosis method based on artificial intelligence. It has the following beneficial effects: 1. This invention utilizes a physical information neural network, incorporating the battery's electrochemical and thermodynamic models as physical law losses into the neural network training process. This allows the model to learn the battery's inherent physical laws from massive amounts of normal operating data and identify parameters with clear physical meaning. This significantly reduces reliance on scarce and costly fault sample data. Furthermore, the output physical parameters (such as internal resistance and diffusion coefficient) significantly enhance the interpretability of fault diagnosis results compared to the abstract features of traditional black-box models.

[0022] 2. This invention achieves sensitive detection of early-stage, subtle faults by constructing a healthy state space manifold in a high-dimensional physical parameter space and transforming the fault diagnosis problem into calculating the geometric distance from the current parameter point to this manifold. A small physical parameter change caused by an early-stage fault, easily obscured by noise in the raw sensor signal, manifests itself in geometric space as a measurable deviation from the healthy manifold, thus achieving a shift from post-event alarm to pre-event early warning.

[0023] 3. This invention uses Lie group theory to model the dynamic evolution of the health state space manifold due to aging, enabling the diagnostic "health baseline" to adaptively adjust over the battery's lifecycle. This method effectively distinguishes between rapid parameter deviations caused by sudden failures and slow, collective parameter drift caused by normal aging, significantly reducing the false alarm rate over the battery's lifecycle and improving the long-term accuracy and reliability of diagnosis.

[0024] 4. This invention utilizes a causal inference algorithm to construct a causal evolution graph using the time series of identified physical parameters, achieving a functional leap from "static diagnosis" to "dynamic prognosis." This method not only diagnoses the current fault but also, based on the learned causal chain, predicts the fault's future development trends and potential chain reactions (such as heat spread risks), providing valuable lead time for preventive maintenance and proactive safety control.

[0025] 5. This invention utilizes a closed-loop active diagnostic incentive mechanism to enable the system to proactively generate and apply optimal incentive signals for active exploration when faced with model uncertainty or suspected fault signals. This shift from passive observation to active interaction enables the system to specifically acquire the most informative response data, quickly eliminating diagnostic ambiguity and confirming fault status, thereby improving the ultimate confidence and efficiency of diagnosis under complex operating conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 is a flow chart of the method of the present invention; Figure 2 This is a system structure diagram of the present invention.

[0027] Among them, 101 is a data acquisition module; 102 is a parameter identification module; 103 is a status diagnosis module; 104 is a fault prognosis module; and 105 is an active excitation module. DETAILED DESCRIPTION

[0028] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the present specification. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0029] Example: Refer to the attached Figure 1 This figure is a schematic diagram of the system architecture of an artificial intelligence-based new energy battery fault diagnosis method provided by one embodiment of the present invention. This invention provides an artificial intelligence-based new energy battery fault diagnosis method. This method achieves accurate diagnosis and forward-looking prognosis of battery health status through a closed-loop system consisting of multiple functional modules.

[0030] Refer to the attached Figure 1 , the new energy battery fault diagnosis method based on artificial intelligence can include the following steps: S1. Data acquisition and fusion: Acquire multimodal spatiotemporal data of the new energy battery under various working conditions, and structure the data into a high-dimensional data tensor.

[0031] S2. Physical parameter identification: Based on the high-dimensional data tensor, a set of key physical parameters that can characterize the internal state of the battery and change dynamically with time and space are inverted through a physical information neural network with built-in physical law constraints.

[0032] S3. Health status diagnosis: In the high-dimensional physical parameter space, a health status space manifold is constructed that can dynamically evolve with battery aging. The degree of battery abnormality is accurately quantified by calculating the geometric distance from the currently identified physical parameter point to the manifold.

[0033] S4. Fault tracing and prognosis: After diagnosing an anomaly, a pre-built causal evolution map is used to trace the root cause of the fault and predict the future development path and potential risks of the fault.

[0034] S5. Closed-loop active excitation: Under specific conditions, such as when there is large uncertainty in model identification or when early signs of faults are detected, the system proactively designs and requests the application of an optimal diagnostic excitation signal to conduct targeted information collection, thereby starting a new round of diagnosis and learning cycles.

[0035] In a specific embodiment, the system architecture provided by the present invention includes five core functional modules that work together: a data acquisition module 101, a parameter identification module 102, a status diagnosis module 103, a fault prognosis module 104, and an active stimulation module 105. These five modules together form an intelligent closed loop from passive monitoring to active detection.

[0036] The system workflow begins with the data acquisition module 101. This module is responsible for continuously collecting multi-source heterogeneous sensor data streams from the battery management system (BMS), including but not limited to the voltage of each battery cell, bus current, and temperature readings distributed at different physical locations within the battery pack. In order to fully preserve the inherent spatiotemporal coupling of these data, this module aligns and fuses the collected data to construct a high-dimensional spatiotemporal data tensor. The mathematical expression of this tensor is: ; in, is the number of sampling points of the time series, and Represent the horizontal and vertical indexes of the battery cell in the physical array of the battery pack, respectively, and It represents the number of types of physical quantities collected at each point in time and space. This structured data tensor provides a complete information view for subsequent in-depth analysis.

[0037] Subsequently, the structured data tensor The data is then transmitted to the parameter identification module 102. This module, the core engine of the system, does not directly establish an end-to-end mapping from raw data to fault labels. Instead, it performs a more fundamental task: inverting the battery's internal physical parameters, which cannot be directly measured. This module identifies the battery's internal state through a physical information neural network that deeply integrates the fundamental laws of battery electrochemistry and thermodynamics. Its output is a series of parameters with clear physical meaning that change dynamically over time and space, such as ohmic internal resistance, charge transfer resistance, and solid-phase lithium ion diffusion coefficient. This process lays a solid physical foundation for subsequent precise diagnosis and interpretable analysis.

[0038] The physical parameter vector output by the parameter identification module 102 is input to the state diagnosis module 103. In the high-dimensional physical parameter space, this module constructs a low-dimensional, smooth health state space manifold based on massive health operation data. This manifold constitutes a "health benchmark" that characterizes all normal operating states of the battery system. Unlike the static benchmark, this manifold can model and evolve the systematic and slow "drift" caused by normal battery aging based on Lie group theory, thereby dynamically adjusting the diagnostic benchmark. The diagnostic process is achieved by calculating the geodesic distance from the parameter point of the current battery cell to the dynamic health manifold. The distance value directly and sensitively quantifies the degree to which the battery cell deviates from the healthy state.

[0039] When the status diagnosis module 103 detects an anomaly (i.e., the distance exceeds a threshold), the fault prognosis module 104 is activated. This module first uses the "causal evolution graph" previously learned through a causal inference algorithm to trace the root cause of the current anomaly. Furthermore, the module uses the currently detected fault signature as the initial disturbance and performs forward reasoning on the causal graph to predict the fault's most likely development path, spread rate, and potential chain reactions over a period of time, achieving a functional transition from diagnosis to prognosis.

[0040] The ingenuity of this invention lies in its closed-loop nature, which is achieved by the active excitation module 105. This module continuously monitors the internal state of the system, particularly the model uncertainty (i.e., the magnitude of the loss of physical laws) of the parameter identification module 102 and the suspiciousness of anomalies detected by the state diagnosis module 103. Once the preset conditions are triggered, the module will immediately initiate an optimization process to calculate a diagnostic excitation signal that maximizes information gain while minimizing energy injection. This signal is applied to the battery through the BMS, and the data acquisition module 101 synchronously collects its response data, thus opening a new, more targeted, and high-precision diagnostic cycle. This closed-loop design enables the system to have a scientist-like ability to "propose a hypothesis, design an experiment, and verify the hypothesis," greatly enhancing the intelligence and reliability of the diagnosis.

[0041] In a specific embodiment of the present invention, each core functional module constituting the system closed loop is described in detail.

[0042] The core of the parameter identification module 102 is a physical information neural network (PINN). The neural network is denoted as , with a set of trainable parameters Its design idea is to combine the powerful nonlinear fitting ability of neural networks with the clear first principles of physics. The input of the network is the normalized space-time coordinates ,in represents time, and Represents the two-dimensional physical position of the battery cell in the battery pack. The output of the network is a physical parameter vector that can fully characterize the internal state of the battery cell at that time and space point. .

[0043] The physical parameter vector Specifically, it includes a series of core parameters that are highly sensitive to the electrochemical health status of the battery but cannot be directly measured by external sensors. Its composition can be expressed as: ; in, The ohmic internal resistance represents the total impedance of electronic and ionic conduction inside the battery cell; is the charge transfer resistance, which reflects the kinetics of the electrochemical reaction at the electrode / electrolyte interface; is the double layer capacitance; is the solid-phase lithium ion diffusion coefficient, which characterizes the ability of lithium ions to migrate inside the electrode active material; is the equivalent thermal conductivity at that location.

[0044] In order to train the physical information neural network , this invention constructs a special composite loss function This loss function does not simply pursue data fitting accuracy, but achieves a deep integration and balance between data drive and physical laws through a sophisticated weighted sum. The mathematical expression is: ; In this formula, represents the data fidelity loss, Represents the loss of physical laws, is the fixed weight of the data fidelity loss, and is a key adaptive regularization weight.

[0045] Data fidelity loss This ensures that the physical parameters identified by the network can accurately explain the external observed phenomena. Specifically, the system converts the physical parameter vector output by the neural network into , substitute into a set of partial differential equations (PDEs) describing the electrochemical-thermodynamic behavior of the battery, denoted as By solving this set of equations, we can forward calculate the value of The predicted state (such as voltage, temperature, etc.) on That is all At each measurement point, the predicted value is consistent with the actual measurement value of the BMS. The mean square error between .

[0046] Loss of physical laws The law of physics itself is transformed into a differentiable and powerful regularization constraint, forcing the output of the neural network to be physically self-consistent. The loss term is calculated at a series of collocation points randomly selected in the training domain. At these points, no real measurement data is required, but the physical parameters of the network output are Directly substitute into the partial differential equations , calculate the mean square value of its residual. An ideal solution that fully complies with physical laws should have a zero residual.

[0047] A core technical feature of the present invention is the adaptive regularization weight The weight is not a fixed hyperparameter that needs to be manually tuned, but is fed back in real time by the state diagnosis module 103 and is the current parameter point. In the health state space manifold The geometric information on it is dynamically determined. Its adjustment logic is: when the parameter point Operating in the core region of the manifold (which typically has high data density and low local curvature, representing stable and well-understood healthy behaviors), the system will increase , thereby strengthening the constraints of physical laws; on the contrary, when the parameter point Drifting to the edge of the manifold or to sparse regions (which are often a precursor to failure or aging), the system will reduce The value of allows the model to rely more on current real data for learning to capture new behaviors that may be beyond the scope of description of the basic physical model.

[0048] In the state diagnosis module 103, a manifold learning algorithm, such as UMAP (Uniform Manifold Approximation and Projection), is first used to process the large number of physical parameter point clouds generated by the parameter identification module 102 when the system is running healthy. This process projects the high-dimensional parameter space into a low-dimensional embedding space that can reveal the inherent geometric structure of the data, thereby constructing the initial healthy state space manifold. . Subsequently, any cell in The degree of abnormality at the moment , is precisely defined as its parameter point To the current health manifold The geodesic distance : ; In order to fundamentally solve the confusion problem between failure and normal aging, the present invention uses the health state space manifold Dynamic evolution modeling. The system does not regard the health baseline as static, but continuously monitors the slow, consistent collective drift of all healthy cell parameters due to aging. This drift is modeled as a continuous transformation flow described by Lie group theory acting on the manifold. By learning the generators of the transformation flow (i.e., the “direction” and “rate” of aging), the system can predict and update the position of the health manifold at any time. Therefore, the health benchmark at the current moment is From the initial manifold Evolved from: ; This dynamic baseline enables diagnostics to be performed throughout the battery's lifecycle with a high degree of accuracy.

[0049] When the anomaly is confirmed, the fault prognosis module 104 is activated. The core of this module is a pre-built causal evolution graph. . The graph is learned by applying a causal inference algorithm (such as a PC stabilization algorithm or a vector autoregression model) to the parameter time series with clear physical meaning output by the parameter identification module 102. The nodes in the graph represent different physical parameters of different battery cells, while the directed edges represent the causal relationship between them with time delays. Using this graph, the module can trace the root cause of the fault by tracing back the upstream path of the abnormal node, and can also propagate the characteristics of the current fault as the initial disturbance on the graph to proactively predict the future evolution path of the fault and potential chain reactions.

[0050] Finally, the active excitation module 105 forms an intelligent closed loop of the system. The excitation trigger decision of this module is composed of a comprehensive model uncertainty (Loss due to physical laws Direct quantification) and abnormal suspicion (By degree of abnormality The function of the time rate of change of Once triggered, the module will immediately solve an optimization problem to generate the optimal diagnostic excitation current signal The goal of this optimization problem is to meet strict energy constraints. Under the premise of maximizing an information gain function : st ; The information gain function JJ is defined as the weighted sum of the expected reduction in model uncertainty and the amplification effect of abnormal features, and its expression is: ; in and is the weight coefficient. In this way, the system can proactively obtain the key information required for decision-making in the most economical and efficient way.

[0051] To further illustrate the collaborative working methods between the various technical modules of the present invention and the resulting technical advantages, a specific implementation process example will be used below. This example uses the scenario of an early, hidden internal micro-short circuit failure in a single cell of a new energy battery pack as a detailed description of the complete workflow of the present invention, from normal monitoring to final prognosis.

[0052] In the initial stage of this embodiment, the entire battery system is in a healthy operating state. The data acquisition module 101 continuously collects data such as voltage, current and temperature of each battery cell from the BMS and constructs it into a spatiotemporal data tensor. The parameter identification module 102 receives the data tensor, and its internal physical information neural network Continuously monitor the internal physical parameters of each battery cell At this point, all the battery cell parameter points are stably distributed in the health state space manifold maintained by the state diagnosis module 103 and dynamically evolving with aging. Therefore, the calculated geodesic distance of each parameter point to the manifold is Both are far less than the preset fault threshold Based on this, the system determines that the entire battery pack is in a healthy state.

[0053] As the operating time accumulates, it is assumed that a cell in the battery pack (denoted as Cell A) begins to experience an early internal micro-short circuit. This fault initially manifests as a weak, intermittent increase in the self-discharge rate. The fluctuations it produces in the external voltage and temperature signals are extremely small and are usually drowned out by normal measurement noise and operating condition fluctuations. Traditional threshold-based diagnostic methods are difficult to detect. However, the parameter identification module 102 of the present invention can capture this internal change. This micro-short circuit phenomenon will be reflected in the electrochemical model as an ohmic internal resistance and charge transfer resistance Therefore, Identified parameter points of Cell A Begins to gradually but surely deviate from the healthy state space manifold main area.

[0054] The state diagnosis module 103 calculates the arrive The geodesic distance Although at the initial stage of the fault, the distance value has not yet reached the fault threshold that triggers the alarm , but the system detects that the distance value shows a monotonically increasing trend. The active excitation module 105 makes a decision based on this trend. , when it exceeds the preset second threshold, the system determines that there is a potential anomaly that is worthy of active exploration.

[0055] At this time, the active excitation module 105 is activated and immediately starts to solve the optimal diagnostic excitation signal. This module generates a specific small current excitation signal by solving the above optimization problem. The waveform and frequency of the signal are specially designed to maximize the system response that can distinguish between ohmic internal resistance and charge transfer resistance, that is, the information gain function The items related to these two parameters are given higher weights. It is then sent to the BMS and applied to the battery system.

[0056] The data acquisition module 101 collects data synchronously The response data of the entire battery pack under stimulation. Since the stimulation signal is designed specifically, the weak abnormal features contained in the response data of Cell A are significantly amplified. When this more informative data is sent to the parameter identification module 102 again, Able to identify CellA with higher confidence and Abnormal deviation of parameters. As a result, the newly calculated parameter points A significant "jump" occurs in manifold space, to the healthy manifold The geodesic distance The failure threshold is decisively exceeded at this moment .

[0057] Once the fault is confirmed, the system immediately enters the diagnosis and prognosis stage. The status diagnosis module 103 analyzes the parameter points The direction vector deviates from the manifold and is found to be and The two parameter dimensions have the largest components, thus diagnosing the fault type as an internal problem related to increased internal resistance. At the same time, the fault prognosis module 104 is activated and the diagnosis result is input as the initial event into the pre-built causal evolution graph. middle.

[0058] By reasoning on the graph, the system can generate a series of forward-looking prognostic conclusions. For example, based on the learned causal chain, the graph predicts: "The continued increase in the internal resistance of Cell A will The Joule heating effect causes an observable rise in the local temperature within 1 hour; this temperature rise will further increase in After 1 hour, the electrolyte side reaction of its adjacent cell Cell B is accelerated by heat conduction, thereby causing the charge transfer resistance of Cell B to Start to rise".

[0059] At this point, through a complete closed-loop process, the present invention not only successfully discovered and confirmed an early hidden fault that was difficult to detect using traditional methods, but also accurately diagnosed its physical root cause and provided a quantitative prediction of the future evolution path of the fault and potential chain reactions, thereby providing strong technical support for achieving accurate predictive maintenance and proactive safety management.

[0060] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A new energy battery fault diagnosis method based on artificial intelligence, characterized in that: The following steps are involved: S1. Acquire multimodal spatiotemporal data of the new energy battery; S2. Based on the multimodal spatiotemporal data, a physical information neural network is used to identify multiple physical parameters of the new energy battery by minimizing a composite loss function including data fidelity loss and physical law loss; S3. Construct a health state space manifold based on the multiple physical parameters, and determine the abnormality degree of the new energy battery by calculating the distance from a current parameter point to the health state space manifold.

2. The artificial intelligence-based new energy battery fault diagnosis method according to claim 1, characterized in that: The physical law loss in the composite loss function carries an adaptive regularization weight; the value of the adaptive regularization weight is dynamically adjusted according to the geometric properties of the current parameter point on the health state space manifold.

3. The artificial intelligence-based new energy battery fault diagnosis method according to claim 1, characterized in that: The health state space manifold is a dynamically evolving manifold; the method further includes: modeling the continuous transformation of the health state space manifold caused by battery aging based on Lie group theory to update the health state space manifold.

4. The artificial intelligence-based new energy battery fault diagnosis method according to claim 1, characterized in that: The method further comprises: Construct a causal evolution map; Furthermore, when the abnormality level exceeds a preset fault threshold, a fault evolution path of the new energy battery is predicted based on the causal evolution graph.

5. The artificial intelligence-based new energy battery fault diagnosis method according to claim 4 is characterized in that: The step of constructing a causal evolution map includes: using the time series of multiple physical parameters identified by the physical information neural network, and learning the causal relationship and time lag relationship between the multiple physical parameters through a causal inference algorithm to generate the causal evolution map.

6. The artificial intelligence-based new energy battery fault diagnosis method according to claim 1, characterized in that: The method further comprises: Under a preset excitation trigger condition, generating an active diagnostic excitation signal and applying it to the new energy battery; Correspondingly, the step of acquiring multimodal spatiotemporal data of the new energy battery includes acquiring response data of the new energy battery under the action of the active diagnosis excitation signal.

7. The artificial intelligence-based new energy battery fault diagnosis method according to claim 6, characterized in that: The preset excitation triggering condition includes at least one of the following: A model uncertainty characterized by the loss of physical laws exceeds a first threshold; Alternatively, an abnormality suspicion level represented by the abnormality level exceeds a second threshold.

8. The artificial intelligence-based new energy battery fault diagnosis method according to claim 6, characterized in that: The step of generating an active diagnostic excitation signal includes calculating the active diagnostic excitation signal by optimizing an information gain function; wherein the information gain function is intended to maximize the reduction in model uncertainty and the change in abnormality level expected to be brought about by the active diagnostic excitation signal.

9. The artificial intelligence-based new energy battery fault diagnosis method according to claim 1, characterized in that: The distance is the geodesic distance from the current parameter point to the health state space manifold.

10. The artificial intelligence-based new energy battery fault diagnosis system according to claim 1, used to execute the method according to any one of claims 1 to 9, characterized in that: include: A data acquisition module for acquiring multimodal spatiotemporal data of new energy batteries; a parameter identification module for identifying multiple physical parameters of the new energy battery based on the multimodal spatiotemporal data using a physical information neural network by minimizing a composite loss function including data fidelity loss and physical law loss; A state diagnosis module is used to construct a health state space manifold based on the multiple physical parameters, and determine the abnormality degree of the new energy battery by calculating the distance from a current parameter point to the health state space manifold.

Images