Energy storage device damage diagnosis and interaction system based on penetration vision and large model

Abstract

The application provides a penetration vision and large model-based energy storage device damage diagnosis and interaction system, relates to the technical field of energy storage device damage diagnosis, and comprises a penetration type 3D perception module, a periodic topology analysis module, a feature mapping and retrieval module, and a large model reasoning and interaction module, which scans and reconstructs the energy storage device; a space-time decoupling periodic Transformer network is constructed, a periodic mask matrix is introduced to force the network to focus on the repetition law of the internal structure of the battery, internal micro-deformation and structural damage are identified by calculating the topological consistency without the need for a large number of negative sample training, feature projection technology is used to map the visual defect features into text embedding, a large language model is combined with retrieval enhancement generation technology to generate a diagnosis report containing physical cause analysis and maintenance suggestions, and interactive question and answer in natural language is supported for the user. The application can solve the pain points of the prior art, such as difficult quantification of three-dimensional deformation, difficult small sample training and lack of intelligent decision-making ability.

Description

Technical Field

[0001] This invention relates to the field of energy storage device damage diagnosis technology, specifically to an energy storage device damage diagnosis and interactive system based on penetrating vision and a large model. Background Technology

[0002] Energy storage devices, such as hydrogen fuel cell stacks and power battery modules, have intricate, periodic stacked structures inside. Currently, non-destructive testing of internal damage in such devices mainly relies on penetrating vision technologies such as industrial CT and X-rays, which perform visual analysis on acquired slice images or identify defects based on traditional machine learning models.

[0003] However, existing methods mostly focus on detecting apparent anomalies, with weak connections between their diagnostic logic and expert knowledge or physical mechanisms, and they struggle to handle the complex periodic topological relationships inherent in sequential slices. Specifically, firstly, conventional visual algorithms or neural networks are unable to effectively model and utilize the inherent periodic structural features within energy storage devices, resulting in insensitivity to potential damage caused by stacking mismatch and periodic deformation. Secondly, detection results are typically isolated defect markers, lacking deep causal analysis and natural language interaction capabilities that integrate multimodal knowledge (such as technical specifications and historical cases), leading to poor interpretability of diagnostic reports and an inability to effectively respond to further inquiries from maintenance personnel.

[0004] Therefore, there is an urgent need for an intelligent diagnostic system that can deeply integrate penetrating vision, periodic structure analysis, and large model reasoning. Summary of the Invention

[0005] In order to solve at least some of the technical problems in the related technologies, the present invention provides a damage diagnosis and interaction system for energy storage devices based on penetrating vision and large models.

[0006] To achieve the above objectives, the technical solution adopted by the present invention includes:

[0007] A damage diagnosis and interactive system for energy storage devices based on penetrating vision and large models includes:

[0008] The penetrating 3D sensing module is configured to scan the target energy storage device using industrial CT or X-rays, reconstruct three-dimensional voxel data characterizing its internal physical structure, and slice the three-dimensional voxel data along the stacking direction of the energy storage device to generate an input sequence tensor.

[0009] A periodic topology analysis module is connected to the penetrating 3D perception module and includes a spatiotemporally decoupled periodic Transformer network. The spatiotemporally decoupled periodic Transformer network is configured to extract features from the input sequence tensor, identify structural anomalies by calculating the spatial features of each slice unit in the sequence and the periodic topological consistency between each slice unit, and output a structured defect feature vector.

[0010] The feature mapping and retrieval module is connected to the periodic topology analysis module and includes a feature projection unit and a knowledge retrieval unit. The feature projection unit is used to map the structured defect feature vector into text embedding features, and the knowledge retrieval unit is used to retrieve relevant technical specifications or historical failure cases from a pre-set industrial knowledge base as knowledge context based on the defect feature vector.

[0011] The large model reasoning and interaction module is connected to the periodic topology analysis module and includes a large language model and a user interface. The large language model is configured to receive the text embedding features and the knowledge context, generate a diagnostic report containing physical cause analysis, and respond to the user's natural language questions through the user interface.

[0012] Optionally, the spatiotemporally decoupled periodic Transformer network includes cascaded intra-slice spatial attention units and inter-slice periodic attention units; the intra-slice spatial attention units are configured to compute the attention of local geometric features within a single slice; the inter-slice periodic attention units are configured to compute long-distance dependencies between slice sequences along the stacking direction and introduce a periodic mask matrix when computing attention weights.

[0013] Optionally, a periodic mask matrix is ​​introduced in the inter-slice periodic attention unit. , its first element The definition satisfies the following rules:

[0014] When the index values ​​of two slices differ When less than or equal to the preset local neighborhood threshold k, To focus on the continuity of local neighborhoods;

[0015] when Furthermore, the index value difference satisfies the periodicity condition. hour, To focus on global cycle consistency, where T is the design cycle parameter of the energy storage device and λ is the scaling factor;

[0016] otherwise, To shield against irrelevant noise interference.

[0017] Optionally, the periodic topology analysis module uses a structural consistency loss function. For anomaly detection, the loss function is defined as follows: ,in, For the first The original features of each slice The reconstructed features are predicted based on the spatiotemporally decoupled periodic Transformer network, where ∇ is the gradient operator used to calculate the rate of change of the features in space. These are the weighting coefficients used to balance the reconstruction error term and the gradient consistency term;

[0018] when When the value exceeds a preset threshold, it is determined that there is structural damage at the corresponding slice location.

[0019] Optionally, the gradient consistency term Specifically, this is achieved by calculating the absolute value of the horizontal gradient difference. The absolute value of the gradient difference in the vertical direction The sum of these components is used to enhance the detection sensitivity for edge deformation and microstructural damage.

[0020] Optionally, the feature projection unit maps high-dimensional visual defect feature vectors to text embedding features aligned with the semantic space of a large language model using a learnable projection matrix. The mapping process is expressed as follows: ;in, For the generated text embedding features, For structured defect feature vectors, and For a trainable parameter matrix, Let Gaussian error be the activation function of the linear unit. Presentation layer normalization operation.

[0021] Optionally, the retrieval operation performed by the knowledge retrieval unit is based on similarity matching between the structured defect feature vector and the vectorized representation of the entries in the industrial knowledge base, to retrieve the most relevant technical specifications or historical failure cases, which constitute the knowledge context C provided to the large language model.

[0022] Optionally, the process of generating a diagnostic report using the large language model is modeled as follows: given visual features... Given the knowledge context C, find the optimal diagnostic text sequence. The maximum a posteriori probability problem, i.e.: ,in, It is a sequence of diagnostic terms; and a cluster search algorithm is used to efficiently search the solution space to generate a professional diagnostic report containing physical causes, consequences assessment and maintenance recommendations.

[0023] Optionally, the energy storage device is a hydrogen fuel cell stack, and the periodic topology analysis module is specifically configured to identify bipolar plate flow channel deformation, membrane electrode misalignment, or abnormal sealing ring compression.

[0024] Optionally, the energy storage device is a power battery module or battery pack, and the periodic topology analysis module is specifically configured to identify cell casing extrusion deformation, internal electrode group twisting, or busbar connection breakage.

[0025] Beneficial effects:

[0026] 1. Through the above technical solution, firstly, the present invention can achieve a complete closed loop from non-invasive internal sensing to intelligent diagnosis and interaction. Specifically, the penetrating 3D sensing module of the present invention utilizes industrial CT or X-rays for scanning and 3D reconstruction, which can acquire complete 3D data of the internal structure of the energy storage device in a non-destructive manner, overcoming the limitations of external visual inspection. Simultaneously, the input sequence tensor output by this module can provide a standardized data foundation for subsequent analysis.

[0027] Secondly, this invention provides structured input for deep diagnostics through an intermediate transformation between feature extraction and anomaly identification. Specifically, the periodic topology analysis module receives the aforementioned sequence tensor and processes it using its internally contained spatiotemporally decoupled periodic Transformer network. This transforms and refines the raw, high-dimensional, and intuitive visual slice data into a highly generalized structured defect feature vector. In this transformation process, the system does not simply compare the original images but performs deep representation learning and feature encoding to identify structural anomalies. This vector is a key intermediate product connecting front-end perception and back-end intelligent diagnosis.

[0028] Third, this invention enables a leap from machine vision features to understandable and interactive professional knowledge. Specifically, for intelligent diagnosis and interaction, firstly, through the feature mapping and retrieval module, the system proactively associates and integrates defect features (structured defect feature vectors) from the machine vision field with industrial knowledge (technical specifications, historical cases) from the text field to generate a knowledge context, thereby achieving knowledge enhancement across modal information. Simultaneously, through the large-model reasoning and interaction module, the system can utilize a large language model to integrate visual feature cues and knowledge context, ultimately generating a diagnostic report containing physical cause analysis, which can then interact with the user through a natural language interface. This effectively achieves a significant improvement in the interpretability and professionalism of diagnostic results, as well as a fundamental improvement in system usability.

[0029] Overall, this invention constructs an automated and intelligent damage diagnosis system through a specific modular architecture of "penetrating 3D perception → periodic topological analysis → intelligent diagnosis and interaction." This system can progressively transform non-invasively acquired internal visual information into machine-understandable defect features, then integrate domain knowledge to ultimately generate in-depth diagnostic reports that are understandable and interactive for human experts. In other words, this invention solves the problems of disconnect between detection and interpretation, and the lack of knowledge integration and natural interaction in the diagnostic process in traditional methods, achieving intelligent and integrated detection and diagnosis processes.

[0030] 2. Other beneficial effects or advantages of the present invention will be described in detail in the specific embodiments. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] in:

[0033] Figure 1 This is a schematic diagram of the overall architecture and data flow of an energy storage device damage diagnosis and interaction system based on penetrating vision and large model, provided by an exemplary embodiment of the present invention.

[0034] Figure 2 This is a schematic diagram of the structure of a spatiotemporally decoupled periodic Transformer network (SP-Former) provided in an exemplary embodiment of the present invention;

[0035] Figure 3 This is a visualization heatmap of a periodic mask matrix provided in an exemplary embodiment of the present invention;

[0036] Figure 4 This is a timing diagram of an intelligent diagnosis and interaction process provided by an exemplary embodiment of the present invention.

[0037] Explanation of markings in the diagram:

[0038] 101-Penetrating 3D Perception Module; 102-Periodic Topology Analysis Module; 103-Feature Mapping and Retrieval Module; 104-Large Model Inference and Interaction Module; 1-Energy Storage Device (Test Item, such as Battery Pack or Stack); 2-Penetrating X-ray Imaging Device (such as Industrial CT Scanner); 3-Three-Dimensional Voxel Data; 4-Spatiotemporally Decoupled Periodic Transformer Network (SP-Former); 5-Slice Sequence; 6-Structured Defect Feature Vector ( ); 7-Text embedding features; 8-Industry knowledge base; 9-Large Language Model (LLM); 10-User interface. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0040] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0041] Furthermore, the terms "comprising" and "having," and any variations thereof, used in the description of this invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include other steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices. It should also be noted that in embodiments of this invention, the words "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in embodiments of this invention should not be construed as preferred or advantageous over other embodiments or designs. Specifically, the use of the words "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0042] To facilitate a clearer and more accurate understanding of the technical solutions of this invention by those skilled in the art, the existing related technologies and their technical problems will be described in more detail below.

[0043] With the rapid development of the new energy industry, the manufacturing quality and safety of electrochemical energy storage devices (such as fuel cells and lithium batteries) have attracted much attention. These devices are usually assembled from hundreds of individual cells or dozens of cells through stacking, series and parallel connection, and the integrity of their internal structure directly determines the performance and lifespan of the product.

[0044] Currently, existing technologies for internal testing of packaged energy storage devices mainly suffer from the following defects and shortcomings:

[0045] One drawback is that traditional detection methods cannot quantify minute three-dimensional deformations. Specifically, current technologies mostly use 2D X-ray projection imaging, but this leads to image overlap between objects, making it difficult to distinguish minute deformations in the depth direction. For example, micron-level collapse of bipolar channels in fuel cells, or changes in interlayer spacing after compression of lithium battery cells, are easily obscured or ignored in 2D images. Although industrial CT can acquire three-dimensional data, there is currently a lack of efficient automated analysis algorithms for the complex layered structures inside batteries. It largely relies on manual layer-by-layer image reading, which is extremely inefficient and highly subjective.

[0046] Second, deep learning models face the challenges of a "cold start" and a lack of samples. Specifically, existing mainstream machine vision algorithms (such as 3D-CNN) are supervised learning methods, requiring a large number of defect samples for training. However, on mature battery production lines, severe internal structural damage (such as internal fractures, deep foreign objects, and severe misalignments) is an extremely low-probability event, making it difficult to collect tens of thousands of negative samples. This results in traditional AI models being unable to work effectively in the early stages of deployment on new production lines (the cold start phase), leading to a high false negative rate.

[0047] Thirdly, there is a lack of intelligent diagnostic and engineering decision-making capabilities. Specifically, existing testing equipment typically only outputs OK / NG labels or the geometric coordinates of defects. The system cannot combine complex engineering design specifications, assembly tolerance tables, or historical failure cases to perform physical consequence analysis on detected defects. For example, the algorithm detects that "a gap has decreased by 0.5mm," but it cannot answer whether "this change will lead to airtightness failure" or "whether immediate rework is necessary." Ultimately, a senior engineer must intervene to make the judgment, thus failing to achieve true unmanned intelligent closed-loop operation.

[0048] Therefore, there is an urgent need for an intelligent system that can accurately identify three-dimensional structural anomalies without requiring a large number of negative samples and can provide physical diagnostic advice like a human expert.

[0049] In view of this, the present invention provides a novel solution: a damage diagnosis and interactive system for energy storage devices based on penetrating vision and a large model. The technical concept of this invention lies in the systematic integration of the three-dimensional visual representation of the internal structure of the energy storage device, its inherent periodic topological prior knowledge, and a multimodal industrial knowledge base to construct a closed-loop intelligent system of "visual perception - structural understanding - knowledge-enhanced diagnosis." Specifically, firstly, a spatiotemporally decoupled periodic Transformer network is proposed. By introducing a periodic mask matrix and structural consistency loss, the model can actively learn and infer normal periodic structural patterns from serialized tomographic images, thereby keenly capturing subtle geometric deviations caused by stacking anomalies. Furthermore, the identified structural defect features are bridged to the semantic space of a large language model through a learnable mapping, and relevant normative and case knowledge is dynamically retrieved as context, ultimately driving the large model to generate a professional diagnostic report with physical causal explanations and natural language interaction capabilities.

[0050] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.

[0051] like Figures 1 to 4 As shown, this invention provides a damage diagnosis and interactive system for energy storage devices based on penetrating vision and a large model, including a penetrating 3D perception module, a periodic topology analysis module, a feature mapping and retrieval module, and a large model reasoning and interaction module. Among them,

[0052] The penetrating 3D sensing module is configured to scan the target energy storage device using industrial CT or X-rays, reconstruct three-dimensional voxel data characterizing its internal physical structure, and slice the three-dimensional voxel data along the stacking direction of the energy storage device to generate an input sequence tensor.

[0053] The periodic topology analysis module is connected to the penetrating 3D perception module and includes a spatiotemporally decoupled periodic Transformer network. The spatiotemporally decoupled periodic Transformer network is configured to extract features from the input sequence tensor, identify structural anomalies by calculating the spatial features of each slice unit in the sequence and the periodic topological consistency between each slice unit, and output a structured defect feature vector.

[0054] The feature mapping and retrieval module is connected to the periodic topology analysis module and includes a feature projection unit and a knowledge retrieval unit. The feature projection unit is used to map the structured defect feature vector into text embedding features, and the knowledge retrieval unit is used to retrieve relevant technical specifications or historical failure cases from a pre-set industrial knowledge base as knowledge context based on the defect feature vector.

[0055] The large model reasoning and interaction module connects to the periodic topology analysis module and includes a large language model and a user interface. The large language model is configured to receive text embedding features and knowledge context, generate a diagnostic report containing physical cause analysis, and respond to users' natural language questions through the user interface.

[0056] Through the above technical solution, firstly, the present invention can achieve a complete closed loop from non-invasive internal sensing to intelligent diagnosis and interaction. Specifically, the penetrating 3D sensing module of the present invention utilizes industrial CT or X-rays for scanning and 3D reconstruction, which can acquire complete 3D data of the internal structure of the energy storage device in a non-destructive manner, overcoming the limitations of external visual inspection. Simultaneously, the input sequence tensor output by this module can provide a standardized data foundation for subsequent analysis.

[0057] Secondly, this invention provides structured input for deep diagnostics through an intermediate transformation between feature extraction and anomaly identification. Specifically, the periodic topology analysis module receives the aforementioned sequence tensor and processes it using its internally contained spatiotemporally decoupled periodic Transformer network. This transforms and refines the raw, high-dimensional, and intuitive visual slice data into a highly generalized structured defect feature vector. In this transformation process, the system does not simply compare the original images but performs deep representation learning and feature encoding to identify structural anomalies. This vector is a key intermediate product connecting front-end perception and back-end intelligent diagnosis.

[0058] Third, this invention enables a leap from machine vision features to understandable and interactive professional knowledge. Specifically, for intelligent diagnosis and interaction, firstly, through the feature mapping and retrieval module, the system proactively associates and integrates defect features (structured defect feature vectors) from the machine vision field with industrial knowledge (technical specifications, historical cases) from the text field to generate a knowledge context, thereby achieving knowledge enhancement across modal information. Simultaneously, through the large-model reasoning and interaction module, the system can utilize a large language model to integrate visual feature cues and knowledge context, ultimately generating a diagnostic report containing physical cause analysis, which can then interact with the user through a natural language interface. This effectively achieves a significant improvement in the interpretability and professionalism of diagnostic results, as well as a fundamental improvement in system usability.

[0059] Overall, this invention constructs an automated and intelligent damage diagnosis system through a specific modular architecture of "penetrating 3D perception → periodic topological analysis → intelligent diagnosis and interaction." This system can progressively transform non-invasively acquired internal visual information into machine-understandable defect features, then integrate domain knowledge to ultimately generate in-depth diagnostic reports that are understandable and interactive for human experts. In other words, this invention solves the problems of disconnect between detection and interpretation, and the lack of knowledge integration and natural interaction in the diagnostic process in traditional methods, achieving intelligent and integrated detection and diagnosis processes.

[0060] like Figure 1 As shown, the system of the present invention is divided into four cascaded modules according to data processing logic, and its working principle is as follows:

[0061] For the penetrating 3D sensing module 101, this module is responsible for the physical acquisition of data. For example... Figure 1 As shown, a non-destructive scan of the energy storage device 1 on the conveyor belt is performed using a transmission X-ray imaging device 2 (e.g., a cone-beam CT system). This process acquires X-ray attenuation information inside the energy storage device and reconstructs it into digitized three-dimensional voxel data 3, thereby enabling visualization of the internal structure without disassembling the battery.

[0062] For the periodic topology analysis module 102, this module is responsible for extracting visual features. The system first discretizes the three-dimensional voxel data into a slice sequence 5 along the stacking direction. Subsequently, these slices are input into a spatiotemporally decoupled periodic Transformer network 4. This network uses a neural network architecture to perform deep analysis on the slice sequence, accurately identify structural anomalies by calculating the topological consistency between slices, and encode them into high-dimensional numerical vectors.

[0063] For the feature mapping and retrieval module 103, this module is responsible for connecting visual perception and cognitive reasoning. As shown by the arrow in the figure, the structured defect feature vector 6 output by the previous stage (represented as...) First, the features are transformed into text embedding features 7 that are semantically aligned with the language model through a feature projection unit. Simultaneously, the system retrieves relevant technical documents or historical cases from a pre-built industrial knowledge base 8 based on the defect features to construct a knowledge context.

[0064] For the large model reasoning and interaction module 104, this module is responsible for intelligent decision-making and human-computer interaction. The large language model 9 (displayed as a mechanical brain icon) receives text embedding features and knowledge context, and performs logical reasoning and causal analysis. The final diagnosis results, maintenance suggestions, and damage heatmap are presented to the operator through the user interface 10, and support users to conduct multi-round question-and-answer interactions with the system through natural language.

[0065] In one embodiment of the present invention, the spatiotemporally decoupled periodic Transformer network of the present invention may include intra-slice spatial attention units and inter-slice periodic attention units connected in series; the intra-slice spatial attention units are configured to calculate the attention of local geometric features within a single slice; the inter-slice periodic attention units are configured to calculate the long-distance dependencies between slice sequences along the stacking direction and introduce a periodic mask matrix when calculating attention weights.

[0066] In this embodiment, the spatiotemporally decoupled periodic Transformer network of the present invention includes a spatial attention unit within a slice and a periodic attention unit between slices connected in series. This enables the hierarchical and decoupled analysis of the internal structure of the energy storage device, allowing the model to efficiently capture the fine geometric features (spatial attention) within a single slice and the long-term dependencies (periodic attention) along the stacking direction that reflect periodic patterns, thereby improving the targeting and efficiency of feature extraction.

[0067] like Figure 2 As shown, the network processing flow of this invention can adopt a bottom-up hierarchical structure, and the specific working principle and steps are as follows:

[0068] S201: Input serialization (bottom).

[0069] Figure 2 The bottom of the image shows the battery slice sequence. The system first discretizes the three-dimensional voxel data reconstructed from CT scans and slices it along the physical stacking direction (Z-axis) of the energy storage device (such as a fuel cell stack or power battery module), generating an input tensor containing N consecutive slices, labeled "slice 1" to "slice N". Each slice represents the two-dimensional physical cross-sectional information of a single battery cell or cell unit.

[0070] S202: Spatial feature coding (first layer).

[0071] The data then enters the "intra-slice spatial attention" processing layer. As shown by the grid icons connected above each slice in the figure, this module processes each slice independently through a self-attention mechanism without inter-layer information exchange. This step aims to capture the microscopic geometric features within a single slice (such as the edge sharpness of the bipolar channel and the thickness distribution of the membrane electrode), mapping the high-dimensional image voxels into compact spatial feature vectors.

[0072] S203: Periodic sequence aggregation (intermediate layer).

[0073] The extracted spatial feature vectors are input into the core "inter-slice periodic attention" module (the large box in the middle of the figure). This module uses the Transformer architecture to perform temporal modeling of the slice sequence. The connection method between nodes in the figure intuitively illustrates the core innovation of this invention:

[0074] Local continuity connections (solid lines): Solid line connections between adjacent nodes indicate that the network focuses on the continuity of the local neighborhood, which is used to detect abrupt changes or breaks between adjacent slices;

[0075] Periodic skip connections (dashed lines): The arched dashed lines in the diagram, labeled "periodic skip connections," vividly illustrate how the network forces the current slice to establish long-distance dependencies with isomorphic slices spanning different periods, based on the inherent periodicity parameter T of the battery design. This connection mechanism allows the network to "skip" intermediate units and directly compare different battery units in the same period phase, thus keenly capturing subtle deviations from periodic patterns even without negative samples.

[0076] S204: Damage Feature Map (top).

[0077] After deep fusion and comparison of the aforementioned spatiotemporal features, the network finally outputs the "abnormal heatmap" at the top. The dark or bright spots shown in the image are the "damage areas." These areas correspond to the spatial locations where the structural consistency loss exceeds a preset threshold, intuitively representing the specific physical coordinates and degree of damage caused by flow channel collapse, component misalignment, or foreign object compression inside the battery, providing accurate feature inputs for subsequent large-scale model diagnosis.

[0078] In one embodiment of the present invention, a periodic mask matrix is ​​introduced in the inter-slice periodic attention unit of the present invention. , its first element The definition satisfies the following rules:

[0079] When the index values ​​of two slices differ When less than or equal to the preset local neighborhood threshold k, To focus on the continuity of local neighborhoods;

[0080] when Furthermore, the index value difference satisfies the periodicity condition. hour, To focus on global cycle consistency, where T is the design cycle parameter of the energy storage device and λ is the scaling factor;

[0081] otherwise, To shield against irrelevant noise interference.

[0082] In this embodiment, by precisely defining the periodic mask matrix using mathematical rules, strong structural prior knowledge about the periodicity of the energy storage device can be injected into the model. This matrix can intelligently guide the model's attention, not only maintaining local continuity and enhancing consistency at periodic nodes, but also shielding irrelevant noise, thereby effectively improving the model's sensitivity to and recognition accuracy of periodic topological anomalies.

[0083] It should also be noted that for the introduced periodic mask matrix In other words, its calculation formula satisfies:

[0084]

[0085] In the formula, Let be the output feature tensor of the inter-slice periodic attention unit, and Softmax be the normalized exponential function. , , These are query, key, and value matrices, respectively. This is the dimension scaling factor.

[0086] like Figure 3 As shown, the periodic mask matrix is ​​a sparse matrix, where the horizontal and vertical axes represent the indices of the input slice sequence. The striped regions (including thick and thin stripes) represent high attention weights, and the blank areas represent regions masked by the mask. The specific distribution characteristics are as follows:

[0087] Main diagonal region: The continuous thick stripes distributed along the diagonal in the figure correspond to areas with small slice index differences. This indicates that the network is configured to strongly focus on the local neighborhood of the current slice to capture the physical continuity features of the battery assembly in the stacking direction.

[0088] Off-diagonal periodic strips: These are several thin strips located on either side of the main diagonal and parallel to it. The spacing between these strips corresponds to the design period parameter T of the energy storage device. This indicates that the network, through a masking mechanism, forces long-distance dependencies across periods, allowing the current slice to be directly compared with other slices in the same period phase, thereby achieving accurate determination of the overall structural periodic consistency.

[0089] In one embodiment of the present invention, the gradient consistency term of the present invention Specifically, this is achieved by calculating the absolute value of the horizontal gradient difference. The absolute value of the gradient difference in the vertical direction The sum of these components is used to enhance the detection sensitivity for edge deformation and microstructural damage.

[0090] In this way, by specifying the gradient consistency term (decomposing it into horizontal and vertical directions), the ability to detect edge contours and micro-geometric distortions can be effectively enhanced. This allows the system to not only focus on overall feature differences, but also to sensitively capture subtle local deformations that characterize the initiation of damage, thereby increasing the probability of detecting early damage and micro-defects.

[0091] In one embodiment of the present invention, the periodic topology analysis module of the present invention uses a structural consistency loss function. For anomaly detection, the loss function is defined as follows: ,in, For the first The original features of each slice The reconstructed features are predicted based on a spatiotemporally decoupled periodic Transformer network, where ∇ is the gradient operator used to calculate the rate of change of the features in space. These are the weighting coefficients used to balance the reconstruction error term and the gradient consistency term; when When the value exceeds a preset threshold, it is determined that there is structural damage at the corresponding slice location.

[0092] This embodiment provides a specific and quantifiable automatic anomaly detection mechanism. Specifically, by calculating the structural consistency loss between the reconstructed features and the original features and comparing it with a threshold, the system can objectively and automatically locate damaged slices, thereby reducing subjective judgment and effectively improving the automation and reliability of damage detection.

[0093] In one embodiment of the present invention, the feature projection unit maps high-dimensional visual defect feature vectors to text embedding features aligned with the semantic space of a large language model using a learnable projection matrix. The mapping process is expressed as follows: ;in, For the generated text embedding features, For structured defect feature vectors, and For a trainable parameter matrix, Let Gaussian error be the activation function of the linear unit. Presentation layer normalization operation.

[0094] In this way, the specific learnable transformation process of mapping visual defect feature vectors to text embedding space can effectively ensure that the semantic alignment from the visual domain to the linguistic domain is obtained through data-driven optimization rather than a fixed mapping, thereby improving the accuracy and adaptability of the generated prompts to large models and laying a key foundation for generating high-quality diagnostic reports.

[0095] It should be noted that the feature projection unit uses a learnable linear projection matrix. , to transform the structured defect feature vector Converted into text embedding features aligned with the dimensions of the large language model. The calculation formula is as follows: ;in, This is the bias vector.

[0096] In one embodiment of the present invention, the retrieval operation performed by the knowledge retrieval unit of the present invention is based on similarity matching between the structured defect feature vector and the vectorized representation of the entries in the industrial knowledge base, and retrieves the most relevant technical specifications or historical failure cases to constitute the knowledge context C provided to the large language model.

[0097] In this way, by clearly defining knowledge retrieval as a mechanism based on vector similarity matching, the system can quickly and accurately retrieve the most relevant technical information and historical cases from the vast industrial knowledge base, providing high-quality and highly relevant contextual information for the large model, thereby effectively improving the accuracy and professionalism of subsequent diagnostic reasoning.

[0098] In one embodiment of the present invention, the process of generating a diagnostic report using the large language model of the present invention is modeled as follows: given visual features... Given the knowledge context C, find the optimal diagnostic text sequence. The maximum a posteriori probability problem, i.e. ,in, It is a sequence of diagnostic terms; and a cluster search algorithm is used to efficiently search the solution space to generate a professional diagnostic report containing physical causes, consequences assessment and maintenance recommendations.

[0099] In this implementation, the diagnostic report generation process is formalized as a probabilistic optimization problem and solved using a cluster search. This ensures that the diagnostic text output by the large language model is the optimal or near-optimal sequence given all input information, thereby guaranteeing the logic, professionalism, and completeness of the generated report and avoiding arbitrary or incoherent text generation.

[0100] For example, in one embodiment of the present invention, intelligent diagnosis and interaction can be configured to perform the following steps:

[0101] Receive users' natural language queries;

[0102] The user's question text, the diagnostic report, and historical failure cases retrieved from the industrial knowledge base are used as context inputs into the large language model;

[0103] The reasoning capabilities of large language models are used to generate maintenance recommendations or life predictions for the structural damage.

[0104] Furthermore, such as Figure 4 As shown, it illustrates the complete data processing flow from the "visual perception space" to the "cognitive decision-making space," specifically including the following five cascading steps:

[0105] Step S401: Visual feature extraction.

[0106] The system first receives a structured defect feature vector output by the periodic topology analysis module. This vector highly compresses the three-dimensional geometric information of the damage (such as deformation, location index, and topology deviation value), serving as the raw input for subsequent diagnosis.

[0107] Step S402: Feature mapping and projection.

[0108] By utilizing feature projection units and a pre-trained projection matrix, numerical visual defect features are mapped across modalities to text embedding features aligned with the semantic space dimension of a large language model, thereby achieving the conversion from visual modality to language modality.

[0109] Step S403: Knowledge base retrieval.

[0110] By employing Retrieval Enhanced Generation (RAG) technology, the system automatically retrieves the most relevant technical document fragments (such as assembly tolerance specifications and historical FMEA cases) from a pre-built industrial knowledge base based on the type and location characteristics of defects, and constructs a knowledge context with engineering basis.

[0111] Step S404: Large model reasoning and diagnosis.

[0112] Large Language Models (LLMs) receive combined inputs (including text embedding features and knowledge context) and perform logical reasoning based on deep neural networks. The model compares the "current deformation facts" with the "retrieved design standards" to analyze the physical causes of the damage and the potential risk level.

[0113] Step S405: Interactive report generation.

[0114] The system ultimately outputs a natural language diagnostic report containing specific disposal suggestions (such as "suggest scrapping" or "seal ring needs to be replaced"), and enters an interactive standby state to respond in real time to further technical inquiries from users regarding the diagnostic results.

[0115] In one embodiment of the present invention, the energy storage device of the present invention is a hydrogen fuel cell stack, and the periodic topology analysis module is specifically configured to identify bipolar plate flow channel deformation, membrane electrode misalignment, or abnormal sealing ring compression.

[0116] In this embodiment, the application scenario is clearly defined as a hydrogen fuel cell stack, and the system is specifically designed to identify its unique and critical damage types (such as bipolar plate flow channel deformation). This demonstrates the practical value and technical relevance of this invention in a specific high-end industrial field.

[0117] In one embodiment of the present invention, the energy storage device of the present invention is a power battery module or battery pack, and the periodic topology analysis module is specifically configured to identify cell casing extrusion deformation, internal electrode group twisting or busbar connection breakage.

[0118] In this embodiment, the application scenario is clearly defined as a power battery module or battery pack, and its identifiable typical damage modes (such as cell casing compression) are identified. This demonstrates the direct application value and market prospects of this invention in key industries such as new energy vehicles and large-scale energy storage.

[0119] In the above embodiments, it should be noted that,

[0120] First, regarding the construction of the spatiotemporally decoupled periodic Transformer network (SP-Former).

[0121] To handle the massive amounts of three-dimensional data from energy storage devices and accurately capture periodic anomalies, this invention constructs an SP-Former network:

[0122] Step A (Input Tensorization): Slice the voxel data V reconstructed from CT along the stacking direction to obtain the input sequence. Where S is the number of battery cells.

[0123] Step B (Inter-Slice Periodic Attention Calculation):

[0124] This invention innovatively designs a periodic mask matrix. This is used to force the network to pay attention to the physical cycle patterns of the battery.

[0125] The formula for calculating the attention score matrix A is as follows:

[0126]

[0127] In the formula, the mask matrix The Each element is defined as:

[0128]

[0129] For the local context window size (e.g., take...) This refers to the three cells at the front and rear. T is the inherent cycle of the battery design (for example, if each cell in a fuel cell is one cycle, then T=1; if it is a dual-cell cycle, then T=2). A scaling factor for periodically monitored weights.

[0130] In this way, using this formula, the network can predict the current [number]th [cell] based on the surrounding normal battery cells. The "ideal state" of a battery.

[0131] Step C (Damage determination based on gradient sensitivity):

[0132] To accurately detect micron-level structural extrusion (such as channel wall bending), this invention introduces a gradient consistency loss:

[0133]

[0134] This formula uses the Sobel operator or first-order difference to calculate the gradient of the slice in the horizontal (h) and vertical (w) directions. Compared to simple pixel error, gradient loss is several orders of magnitude more sensitive to edge deformation.

[0135] Second, regarding visual-linguistic feature alignment and diagnostic reasoning.

[0136] Step D (Feature Projection):

[0137] To enable the Large Language Model (LLM) to understand 3D visual features, this system utilizes feature projection units for modality transformation.

[0138] Let the damage feature vector output by the visual network be... The word embedding dimension of the large model is .

[0139] The projection is calculated as follows:

[0140]

[0141] Here, a two-layer perceptron (MLP) structure is used as the projector. and The training parameters ensure that visual features are smoothly mapped to the semantic space.

[0142] Step E (Bayesian Diagnostic Inference):

[0143] The process of generating diagnostic reports using a large model is mathematically modeled as maximizing the posterior probability. Given visual features... Based on the retrieved knowledge context C (such as the "Assembly Tolerance Table"), a diagnostic term sequence Y is generated:

[0144]

[0145] The system uses the Beam Search algorithm to search for the optimal text sequence, thereby outputting a professional report such as "Severe compression detected, tolerance standard violated, scrapping recommended".

[0146] In summary, compared with the prior art, the present invention has the following significant advantages:

[0147] First, it can solve the problem of training with small samples. The spatiotemporally decoupled periodic network proposed in this invention uses the repetitive structure of the battery itself as a "natural reference," which can detect anomalies through consistency analysis without a large number of defective samples, greatly reducing the threshold for system deployment.

[0148] Secondly, it enables the transition from qualitative detection to quantitative diagnosis. By introducing large model and knowledge retrieval technology (RAG), this invention not only identifies "where it is broken," but also infers "the consequences of the failure" and "how to repair" based on design specifications, bridging the gap between testing equipment and maintenance decisions.

[0149] Third, high-precision three-dimensional deformation sensing. Based on voxel-level three-dimensional analysis, this invention can accurately capture microstructural damage such as interlayer spacing changes and flow channel collapse that cannot be identified by traditional 2D X-rays, significantly improving the safety of battery products.

[0150] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A damage diagnosis and interactive system for energy storage devices based on penetrating vision and large models, characterized in that, include: The penetrating 3D sensing module is configured to scan the target energy storage device using industrial CT or X-rays, reconstruct three-dimensional voxel data characterizing its internal physical structure, and slice the three-dimensional voxel data along the stacking direction of the energy storage device to generate an input sequence tensor. A periodic topology analysis module is connected to the penetrating 3D perception module and includes a spatiotemporally decoupled periodic Transformer network. The spatiotemporally decoupled periodic Transformer network is configured to extract features from the input sequence tensor, identify structural anomalies by calculating the spatial features of each slice unit in the sequence and the periodic topological consistency between each slice unit, and output a structured defect feature vector. The feature mapping and retrieval module is connected to the periodic topology analysis module and includes a feature projection unit and a knowledge retrieval unit. The feature projection unit is used to map the structured defect feature vector into text embedding features, and the knowledge retrieval unit is used to retrieve relevant technical specifications or historical failure cases from a pre-set industrial knowledge base as knowledge context based on the defect feature vector. The large model reasoning and interaction module is connected to the periodic topology analysis module and includes a large language model and a user interface. The large language model is configured to receive the text embedding features and the knowledge context, generate a diagnostic report containing physical cause analysis, and respond to the user's natural language questions through the user interface.

2. The energy storage device damage diagnosis and interactive system based on penetration vision and large model as described in claim 1, characterized in that, The spatiotemporally decoupled periodic Transformer network includes cascaded intra-slice spatial attention units and inter-slice periodic attention units; the intra-slice spatial attention units are configured to compute the attention to local geometric features within a single slice; The inter-slice periodic attention unit is configured to calculate long-distance dependencies between slice sequences along the stacking direction and introduces a periodic mask matrix when calculating attention weights.

3. The energy storage device damage diagnosis and interactive system based on penetration vision and large model according to claim 2, characterized in that, The periodic mask matrix introduced in the inter-slice periodic attention unit , its first element The definition satisfies the following rules: When the index values ​​of two slices differ When less than or equal to the preset local neighborhood threshold k, To focus on the continuity of local neighborhoods; when Furthermore, the index value difference satisfies the periodicity condition. hour, To focus on global cycle consistency, where T is the design cycle parameter of the energy storage device and λ is the scaling factor; otherwise, To shield against irrelevant noise interference.

4. The energy storage device damage diagnosis and interactive system based on penetration vision and large model according to claim 3, characterized in that, The periodic topology analysis module uses a structural consistency loss function. For anomaly detection, the loss function is defined as follows: ,in, For the first The original features of each slice The reconstructed features are predicted based on the spatiotemporally decoupled periodic Transformer network, where ∇ is the gradient operator used to calculate the rate of change of the features in space. These are the weighting coefficients used to balance the reconstruction error term and the gradient consistency term; when When the value exceeds a preset threshold, it is determined that there is structural damage at the corresponding slice location.

5. The energy storage device damage diagnosis and interactive system based on penetration vision and large model according to claim 4, characterized in that, The gradient consistency term Specifically, this is achieved by calculating the absolute value of the horizontal gradient difference. The absolute value of the gradient difference in the vertical direction The sum of these components is used to enhance the detection sensitivity for edge deformation and microstructural damage.

6. The energy storage device damage diagnosis and interactive system based on penetration vision and large model according to claim 1, characterized in that, The feature projection unit maps high-dimensional visual defect feature vectors to text embedding features aligned with the semantic space of a large language model using a learnable projection matrix. The mapping process is represented as follows: ;in, For the generated text embedding features, For structured defect feature vectors, and For a trainable parameter matrix, Let Gaussian error be the activation function of the linear unit. Presentation layer normalization operation.

7. The energy storage device damage diagnosis and interactive system based on penetration vision and large model according to claim 1, characterized in that, The retrieval operation performed by the knowledge retrieval unit is based on similarity matching between the structured defect feature vector and the vectorized representation of the entries in the industrial knowledge base, retrieving the most relevant technical specifications or historical failure cases, which constitute the knowledge context C provided to the large language model.

8. The energy storage device damage diagnosis and interactive system based on penetration vision and large model according to claim 1, characterized in that, The process of generating diagnostic reports using the large language model is modeled as follows: given visual features... Given the knowledge context C, find the optimal diagnostic text sequence. The maximum a posteriori probability problem, i.e.: ,in, It is a sequence of diagnostic terms; and a cluster search algorithm is used to efficiently search the solution space to generate a professional diagnostic report containing physical causes, consequences assessment and maintenance recommendations.

9. The damage diagnosis and interactive system for energy storage devices based on penetration vision and large model according to any one of claims 1-8, characterized in that, The energy storage device is a hydrogen fuel cell stack, and the periodic topology analysis module is specifically configured to identify bipolar plate flow channel deformation, membrane electrode misalignment, or abnormal sealing ring compression.

10. The damage diagnosis and interactive system for energy storage devices based on penetration vision and large model according to any one of claims 1-8, characterized in that, The energy storage device is a power battery module or battery pack, and the periodic topology analysis module is specifically configured to identify cell casing extrusion deformation, internal electrode group twisting, or busbar connection breakage.

Images