Tire failure detection method, device, electronic equipment and computer readable medium

Abstract

The application discloses a tire fault detection method and device, electronic equipment and computer readable medium, relates to the technical field of computers, and the method comprises the following steps: in response to triggering a tire detection task, acquiring multi-modal detection data for a target tire, determining the type of the multi-modal detection data and the detection task; according to the type and the detection task, the multi-modal detection data is allocated to the corresponding fault detection agent, the fault detection agent carries out fault detection on the multi-modal detection data allocated based on the corresponding detection task, and a fault detection result is obtained, the fault detection agent comprises an appearance defect detection agent, an internal structure detection agent and an operating state detection agent; the fault detection result is summarized and input to a decision and report generation agent, the decision and report generation agent carries out comprehensive decision based on the summarized fault detection result, and a tire fault detection result is obtained. The tire fault detection cost is effectively reduced, and the tire fault detection efficiency and accuracy are improved.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a tire fault detection method, apparatus, electronic device, and computer-readable medium. Background Technology

[0002] Currently, tire fault detection primarily relies on manual visual inspection at the production stage. This method is hampered by reliance on quality inspectors' visual checks, high staff turnover due to harsh workshop environments, subjective inspection standards, rising labor costs, and lagging information technology. Automated inspection based on traditional machine vision, on the vehicle side, focuses on single-dimensional condition detection. For example, it might analyze radial acceleration signals to determine wear or use sound signals to identify foreign objects in the tire. These methods suffer from fragmented detection dimensions, lack of data integration, and limited data dimensionality, making it difficult to comprehensively assess tire health. Tire fault detection is costly, inefficient, and lacks accuracy. Summary of the Invention

[0003] In view of this, embodiments of this application provide a tire fault detection method, apparatus, electronic device, and computer-readable medium, which can solve the problems of high cost, low efficiency, and low accuracy in existing tire fault detection methods.

[0004] To achieve the above objectives, according to one aspect of the embodiments of this application, a tire fault detection method is provided, comprising: In response to triggering a tire inspection task, acquire multimodal inspection data for the target tire, and determine the type of multimodal inspection data and the inspection task; Based on the type and detection task, a corresponding fault detection agent is assigned to the multimodal detection data. The fault detection agent performs fault detection on the assigned multimodal detection data based on the corresponding detection task to obtain the fault detection result. The fault detection agent includes an appearance defect detection agent, an internal structure detection agent, and an operating status detection agent. The fault detection results are summarized and input into the decision and report generation agent. The decision and report generation agent makes a comprehensive decision based on the summarized fault detection results to obtain the tire fault detection results.

[0005] Optionally, acquire multimodal detection data for the target tire, determine the type of multimodal detection data and the detection task, including: The system calls upon data acquisition and standardized intelligent agents to aggregate multimodal detection data for the target tire, including image data, X-ray data, and time-series signals. The multimodal detection data is standardized to obtain standard data packets, and the standard data packets are labeled with the corresponding type and detection task.

[0006] Optionally, a corresponding fault detection agent is assigned to the multimodal detection data according to the type and detection task, including: A standard data package, labeled as image data and as surface defect detection task, is assigned to the appearance defect detection agent. Assign standard data packets, labeled as X-ray data and with the detection task labeled as detecting internal defects, to the internal structure detection agent. Standard data packets, labeled as time-series signals and as service status detection tasks, are assigned to the operational status detection agent.

[0007] Optionally, the fault detection agent performs fault detection on the assigned multimodal detection data based on the corresponding detection task to obtain fault detection results, including: The appearance defect detection agent extracts multi-level features of the assigned standard data packets according to the detection task corresponding to the assigned standard data packets; For each level of feature in the multi-level feature set, the feature distance field between the level feature and the nearest neighbor normal feature prototype in the multi-level normal feature prototype library is calculated. The multi-level normal feature prototype is obtained by clustering the multi-level features of multiple good product images of new specification tires in the feature space. The feature distance field is input into the anomaly amplification network, which outputs an anomaly heatmap and defect classification confidence based on the input feature distance field.

[0008] Optionally, the defect classification confidence level is obtained based on the following method: The anomaly amplification network calculates the mean and standard deviation of the heatmap corresponding to the anomaly heatmap; Adaptive threshold segmentation of abnormal heatmaps is performed based on the mean and standard deviation of the heatmaps to obtain candidate defect regions. For each candidate defect region, extract the abnormal intensity statistics, geometric features, and texture features; Defect types are obtained by feature pattern matching based on abnormal intensity statistics, geometric features, and texture features, and the defect classification confidence score corresponding to each defect type is calculated.

[0009] Optionally, the fault detection agent performs fault detection on the assigned multimodal detection data based on the corresponding detection task to obtain fault detection results, including: The running status detection agent maps the multimodal signals in the allocated standard data packets onto a virtual, equally spaced time axis to obtain aligned multimodal signals that retain the timing relationship. For each modal signal in the aligned multimodal signal that preserves the temporal relationship, high-level spatiotemporal features corresponding to the modal signal are extracted, and modal key vectors representing the global state of the modal signal are generated based on the high-level spatiotemporal features. For each detection task corresponding to the standard data packet allocated to the operational status detection agent, determine the modality importance query vector corresponding to the detection task, and determine the attention weight combination for the detection task based on the modality importance query vector and each modality key vector; Based on the combination of each modality key vector and attention weights, weighted attention fusion is performed to generate a fusion feature vector for the detection task; The lightweight task heads for all detection tasks are computed in parallel based on their respective fused feature vectors during a single forward propagation, and each task outputs its own running status detection results.

[0010] Optionally, before obtaining the fault detection results, the method further includes: In response to a request to add a new detection task, a new detection task is added, and a modal importance query vector and a lightweight task header are configured for the new detection task.

[0011] In addition, this application also provides a tire fault detection device, including: The data acquisition unit is configured to acquire multimodal detection data for the target tire in response to triggering a tire detection task, and to determine the type of multimodal detection data and the detection task. The allocation unit is configured to allocate corresponding fault detection agents to the multimodal detection data according to the type and detection task. The fault detection agents perform fault detection on the allocated multimodal detection data based on the corresponding detection task to obtain the fault detection results. The fault detection agents include appearance defect detection agents, internal structure detection agents and operating status detection agents. The integrated decision-making unit is configured to summarize the fault detection results and input them into the decision and report generation agent. The decision and report generation agent then makes an integrated decision based on the summarized fault detection results to obtain the tire fault detection results.

[0012] Optionally, the data acquisition unit is further configured to: The system calls upon data acquisition and standardized intelligent agents to aggregate multimodal detection data for the target tire, including image data, X-ray data, and time-series signals. The multimodal detection data is standardized to obtain standard data packets, and the standard data packets are labeled with the corresponding type and detection task.

[0013] Optionally, the allocation unit is further configured to: A standard data package, labeled as image data and as surface defect detection task, is assigned to the appearance defect detection agent. Assign standard data packets, labeled as X-ray data and with the detection task labeled as detecting internal defects, to the internal structure detection agent. Standard data packets, labeled as time-series signals and as service status detection tasks, are assigned to the operational status detection agent.

[0014] Optionally, the allocation unit is further configured to: The appearance defect detection agent extracts multi-level features of the assigned standard data packets according to the detection task corresponding to the assigned standard data packets; For each level of feature in the multi-level feature set, the feature distance field between the level feature and the nearest neighbor normal feature prototype in the multi-level normal feature prototype library is calculated. The multi-level normal feature prototype is obtained by clustering the multi-level features of multiple good product images of new specification tires in the feature space. The feature distance field is input into the anomaly amplification network, which outputs an anomaly heatmap and defect classification confidence based on the input feature distance field.

[0015] Optionally, the defect classification confidence level is obtained based on the following method: The anomaly amplification network calculates the mean and standard deviation of the heatmap corresponding to the anomaly heatmap; Adaptive threshold segmentation of abnormal heatmaps is performed based on the mean and standard deviation of the heatmaps to obtain candidate defect regions. For each candidate defect region, extract the abnormal intensity statistics, geometric features, and texture features; Defect types are obtained by feature pattern matching based on abnormal intensity statistics, geometric features, and texture features, and the defect classification confidence score corresponding to each defect type is calculated.

[0016] Optionally, the allocation unit is further configured to: The running status detection agent maps the multimodal signals in the allocated standard data packets onto a virtual, equally spaced time axis to obtain aligned multimodal signals that retain the timing relationship. For each modal signal in the aligned multimodal signal that preserves the temporal relationship, high-level spatiotemporal features corresponding to the modal signal are extracted, and modal key vectors representing the global state of the modal signal are generated based on the high-level spatiotemporal features. For each detection task corresponding to the standard data packet allocated to the operational status detection agent, determine the modality importance query vector corresponding to the detection task, and determine the attention weight combination for the detection task based on the modality importance query vector and each modality key vector; Based on the combination of each modality key vector and attention weights, weighted attention fusion is performed to generate a fusion feature vector for the detection task; The lightweight task heads for all detection tasks are computed in parallel based on their respective fused feature vectors during a single forward propagation, and each task outputs its own running status detection results.

[0017] Optionally, the tire fault detection device also includes a detection task addition processing unit, configured to: In response to a request to add a new detection task, a new detection task is added, and a modal importance query vector and a lightweight task header are configured for the new detection task.

[0018] In addition, this application also provides a tire fault detection electronic device, including: one or more processors; and a storage device for storing one or more programs, which, when executed by one or more processors, enable the one or more processors to implement the tire fault detection method as described above.

[0019] In addition, this application also provides a computer-readable medium having a computer program stored thereon, which, when executed by a processor, implements the tire fault detection method as described above.

[0020] To achieve the above objectives, according to another aspect of the embodiments of this application, a computer program product is provided.

[0021] A computer program product according to an embodiment of this application includes a computer program that, when executed by a processor, implements the tire fault detection method provided in an embodiment of this application.

[0022] One embodiment of the above invention has the following advantages or beneficial effects: This application, in response to triggering a tire inspection task, acquires multimodal inspection data for a target tire, determines the type of multimodal inspection data and the inspection task; assigns corresponding fault detection agents to the multimodal inspection data according to the type and inspection task, and the fault detection agents perform fault detection on the assigned multimodal inspection data based on the corresponding inspection task to obtain fault detection results. The fault detection agents include appearance defect detection agents, internal structure detection agents, and operating status detection agents; the fault detection results are summarized and input to a decision and report generation agent, which makes a comprehensive decision based on the summarized fault detection results to obtain the tire fault detection result. This can effectively reduce tire fault detection costs and improve tire fault detection efficiency and accuracy.

[0023] The further effects of the aforementioned unconventional alternative methods will be explained below in conjunction with specific implementation methods. Attached Figure Description

[0024] The accompanying drawings are provided to better understand this application and do not constitute an undue limitation thereof. Wherein: Figure 1 This is a schematic diagram of the main flow of a tire fault detection method according to an embodiment of this application; Figure 2 This is a schematic diagram of the main flow of a tire fault detection method according to an embodiment of this application; Figure 3 This is a schematic diagram of the three-layer architecture and complete data flow involved in the implementation of a tire fault detection method according to an embodiment of this application; Figure 4 This is a schematic diagram of the main units of a tire fault detection device according to an embodiment of this application; Figure 5 This is an exemplary system architecture diagram to which embodiments of this application can be applied; Figure 6 This is a schematic diagram of the structure of a computer system suitable for implementing terminal devices or servers in the embodiments of this application. Detailed Implementation

[0025] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of this application, including various details to aid understanding. These embodiments should be considered merely exemplary. Therefore, those skilled in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description. It should be noted that the acquisition, transmission, storage, use, and processing of data in the technical solutions of this application comply with relevant national laws and regulations. It should also be noted that certain software, components, models, and other existing industry solutions may be mentioned in the embodiments of this application. These should be considered exemplary, intended only to illustrate the feasibility of implementing the technical solutions of this application, and do not imply that the applicant has already used or necessarily used such solutions. The collection, gathering, updating, analysis, processing, use, transmission, and storage of user personal information involved in the technical solutions of this application all comply with relevant laws and regulations, are used for legal and reasonable purposes, do not violate public order and good morals, are not shared, disclosed, or sold outside of these legal uses, and are subject to supervision and management by regulatory authorities. Necessary measures should be taken to prevent unauthorized access to user personal information, safeguard user personal information security, cybersecurity, and national security, and ensure that those authorized to access personal information comply with relevant laws and regulations. Once this user personal information is no longer needed, risks should be minimized by restricting or even prohibiting data collection and / or deleting the data.

[0026] When used, including in certain relevant applications, data is deidentified to protect user privacy, for example by removing specific identifiers, controlling the amount or specificity of stored data, controlling how data is stored, and / or other methods.

[0027] Figure 1 This is a schematic diagram of the main flow of a tire fault detection method according to an embodiment of this application, as shown below. Figure 1 As shown, the tire fault detection method mainly includes the following steps S101-S103.

[0028] Step S101: In response to triggering a tire detection task, acquire multimodal detection data for the target tire, and determine the type of multimodal detection data and the detection task.

[0029] In this embodiment, the execution entity of the tire fault detection method (e.g., a server) responds to the triggering of the tire detection task by calling a data acquisition and standardization agent to obtain multimodal detection data for the target tire, such as image data, X-ray data, and time-series signals, and determines the data type of the multimodal detection data and the required detection task.

[0030] Step S102: Assign corresponding fault detection agents to the multimodal detection data according to the type and detection task. The fault detection agents perform fault detection on the assigned multimodal detection data based on the corresponding detection task to obtain the fault detection results. The fault detection agents include appearance defect detection agents, internal structure detection agents and operating status detection agents.

[0031] For example, multimodal detection data is routed and distributed via a message middleware (RabbitMQ / Kafka). The message middleware allocates the corresponding data (e.g., image data) to a fault detection agent (e.g., an appearance defect detection agent) capable of performing the corresponding detection task (e.g., detecting surface defects) based on the type of multimodal detection data (e.g., image data) and the detection task (e.g., detecting surface defects). The fault detection agent (e.g., the appearance defect detection agent) performs fault detection on the allocated multimodal detection data (e.g., image data) based on the corresponding detection task (e.g., detecting surface defects), obtaining fault detection results, such as anomaly heatmaps and defect classification confidence scores.

[0032] Step S103: Summarize the fault detection results and input them into the decision and report generation agent. The decision and report generation agent makes a comprehensive decision based on the summarized fault detection results to obtain the tire fault detection results.

[0033] The system aggregates the fault detection results obtained from the appearance defect detection agent, internal structure detection agent, and operational status detection agent, and inputs the aggregated fault detection results into the decision and report generation agent to obtain the tire fault detection results output by the decision and report generation agent. For example, the tire fault detection results can exist in various forms, such as: structured reports: delivered to the production MES system to guide process improvement; real-time early warnings: sent to the vehicle HMI or service station to ensure driving safety; tire digital twin files: stored on a cloud platform to form a full life cycle health record.

[0034] This embodiment, in response to a triggered tire inspection task, acquires multimodal inspection data for the target tire, determines the type of multimodal inspection data and the inspection task, and assigns corresponding fault detection agents to the multimodal inspection data according to the type and inspection task. The fault detection agents perform fault detection on the assigned multimodal inspection data based on the corresponding inspection task, obtaining fault detection results. The fault detection agents include appearance defect detection agents, internal structure detection agents, and operating status detection agents. The fault detection results are summarized and input into a decision and report generation agent, which makes a comprehensive decision based on the summarized fault detection results to obtain the tire fault detection result. This can effectively reduce tire fault detection costs and improve tire fault detection efficiency and accuracy.

[0035] Figure 2 This is a schematic flowchart of the main process of a tire fault detection method according to an embodiment of this application, as follows: Figure 2 As shown, the tire fault detection method mainly includes the following steps S201-S207.

[0036] Step S201: In response to triggering the tire inspection task, the data acquisition and standardization agent is invoked to gather multimodal inspection data for the target tire, including image data, X-ray data and time-series signals.

[0037] The target tire is the tire that needs to undergo intelligent inspection throughout its entire lifecycle. When a tire inspection task is triggered, the data acquisition and standardized intelligent agent can be invoked to gather multimodal inspection data for the target tire collected by various devices (e.g., production-end devices: production line vision cameras, tire X-ray machines; service-end devices: vehicle CAN bus, tire pressure / vibration sensors), such as image data, X-ray data, and time-series signals.

[0038] Step S202: Standardize the multimodal detection data to obtain standard data packets, and label the standard data packets with the corresponding type and detection task.

[0039] Step S203: Assign a standard data packet, labeled as image data and labeled as surface defect detection task, to the appearance defect detection agent.

[0040] Step S204: Assign a standard data packet, labeled as X-ray data and labeled as internal defect detection task, to the internal structure detection agent.

[0041] Step S205: Assign a standard data packet, tagged as a timing signal and tagged as a service status detection task, to the operational status detection agent.

[0042] Step S206: The fault detection agent performs fault detection on the assigned multimodal detection data based on the corresponding detection task to obtain the fault detection result. The fault detection agent includes an appearance defect detection agent, an internal structure detection agent, and an operating status detection agent.

[0043] Specifically, the fault detection agent performs fault detection on the assigned multimodal detection data based on the corresponding detection task, obtaining fault detection results. This includes: the appearance defect detection agent extracts multi-level features from the assigned standard data package according to the detection task corresponding to the assigned standard data package (e.g., detecting surface defects). For example, shallow features (layers 3-6): capturing texture, granularity, and local patterns; mid-level features (layers 7-10): extracting structural features and edge information; deep features (layers 11-12): containing semantic information and global context; and for multi-level features... For each level of feature (in this example, the current level of feature is a shallow feature), a feature distance field is calculated between the level of feature (in this example, the current level of feature is a shallow feature) and the nearest neighbor normal feature prototypes (e.g., nearest neighbor normal feature prototype 1, nearest neighbor normal feature prototype 3, ...) in the multi-level normal feature prototype library. The multi-level normal feature prototypes are obtained by clustering the multi-level features in the feature space based on multiple good-quality images of the new specification tire. For example, the normal prototype library is constructed as follows: for each new specification tire (e.g., the target tire of this application), only a few dozen good-quality images need to be collected. Multi-level features are extracted using a pre-trained Visual Transformer (ViT) and clustered in the feature space to form a "multi-level normal feature prototype library" for that specification tire (e.g., the target tire of this application). The feature distance field is input into an anomaly amplification network. The anomaly amplification network, through learnable upsampling and attention fusion, outputs an anomaly heatmap and a defect classification confidence score corresponding to the defect type of the anomaly heatmap based on the input feature distance field.

[0044] Specifically, the defect classification confidence is obtained as follows: the anomaly amplification network calculates the heatmap mean (e.g., μ_H) and standard deviation (e.g., σ_H) corresponding to the anomaly heatmap; adaptive thresholding is performed on the anomaly heatmap based on the heatmap mean and standard deviation to obtain candidate defect regions. For example, adaptive thresholding is performed on the anomaly heatmap H: T = μ_H + α·σ_H, where α is adjusted according to the specifications to obtain multiple candidate defect regions; for each candidate defect region, anomaly intensity statistics (e.g., mean, maximum, variance), geometric features (e.g., area, aspect ratio, density), and texture features (e.g., contrast with surrounding normal regions) are extracted; feature pattern matching is performed with the anomaly intensity statistics, geometric features, and texture features against a pseudo-anomaly database to obtain the defect type, and the defect classification confidence is calculated for each defect type. The pseudo-anomaly database is obtained as follows: "pseudo-anomaly" data is created using normal samples through data augmentation. Pseudo-anomaly database generation methods: random local occlusion (simulating blemishes and damage), texture replacement (copying and pasting from other areas), Gaussian noise injection (simulating image acquisition noise), and elastic deformation (simulating surface deformation).

[0045] Specifically, the fault detection agent performs fault detection on the assigned multimodal detection data based on the corresponding detection task to obtain fault detection results. This includes: the operation status detection agent uniformly maps the multimodal signals (e.g., vibration, sound, pressure, etc.) in the assigned standard data packets onto a virtual, equally spaced time axis to obtain aligned multimodal signals that retain their temporal relationships. For example, through an asynchronous signal alignment module: for signals such as vibration, sound, and pressure with different sampling rates and incompletely synchronized timestamps, a learnable time interpolation alignment layer is introduced to uniformly map each signal onto a virtual, equally spaced time axis, preserving their temporal relationships. For each modal signal in the aligned multimodal signal that preserves temporal relationships (e.g., the current modal signal is an audio signal), high-level spatiotemporal features corresponding to the modal signal (e.g., audio signal) are extracted. Based on the high-level spatiotemporal features, a modal key vector representing the global state of the modal signal is generated. For example, this high-level spatiotemporal feature is used by a lightweight modal self-description module to generate a modal key vector representing the global state of the current modal signal. For example, the modal key vector of the audio modality may encode the high-frequency energy information of the current segment. For each detection task corresponding to the standard data packet allocated to the running state detection agent, the modal importance query vector corresponding to the detection task is determined. Based on the modal importance query vector and each modal key vector, the attention weight combination for the detection task is determined. For example, for the detection task T, its "modal importance query vector" is similar to the "modal key vectors" of all modalities (e.g., dot product followed by Softmax) to obtain a set of normalized attention weights [α_vibration, α_sound, α_pressure]_T, which is the attention weight combination. This set of weights indicates that, given the current input sample, the detection task T should assign relative importance to each modal feature when determining a fault. For example, for the nailing task, the network might automatically learn α_sound ≈ 0.7, α_vibration ≈ 0.2, and α_pressure ≈ 0.1. Based on the combination of each modal key vector and attention weights, weighted attention fusion is performed to generate a fusion feature vector tailored to the detection task. For example, weighted attention fusion involves summing each modal key vector according to its task-specific attention weights to generate a fusion feature vector customized for the detection task T. That is, the same input sample will generate different fusion feature vectors for different detection tasks. This achieves accurate perception; the "nailing" task focuses on high-frequency sound, and the "wear" task focuses on the vibration envelope, achieving optimal filtering of multimodal information. All lightweight task heads corresponding to the detection tasks perform parallel computation based on their respective fusion feature vectors during a single forward propagation, simultaneously outputting their respective operational status detection results. All lightweight task heads are computed in parallel during a single forward propagation, achieving a single forward computation while simultaneously obtaining the detection results of all tire faults, meeting the real-time requirement of millisecond-level response.

[0046] Specifically, before obtaining the fault detection results, the method also includes: in response to a new detection task request, adding a new detection task and configuring a modal importance query vector and a lightweight task header for the new detection task.

[0047] When adding a new diagnostic task, you only need to add a corresponding "modal attention query vector" and a lightweight task header. The main body does not need to be modified, making it flexible, scalable, and low-cost to expand.

[0048] Step S207: Summarize the fault detection results and input them into the decision and report generation agent. The decision and report generation agent makes a comprehensive decision based on the summarized fault detection results to obtain the tire fault detection results.

[0049] The system aggregates the fault detection results obtained from the appearance defect detection agent, internal structure detection agent, and operational status detection agent, and inputs the aggregated fault detection results into the decision and report generation agent to obtain the tire fault detection results output by the decision and report generation agent. For example, the tire fault detection results can exist in various forms, such as: structured reports: delivered to the production MES system to guide process improvement; real-time early warnings: sent to the vehicle HMI or service station to ensure driving safety; tire digital twin files: stored on a cloud platform to form a full life cycle health record.

[0050] Figure 3 This is a schematic diagram of the three-layer architecture and complete data flow involved in the implementation of a tire fault detection method according to an embodiment of this application. Figure 3As shown, the data source layer (bottom layer) provides multimodal raw data input, covering the entire tire lifecycle, including: production-end equipment: production line vision cameras (acquiring exterior images), tire X-ray machines (acquiring internal structure images); service-end equipment: vehicle-mounted CAN bus (acquiring vehicle status data), tire pressure / vibration sensors (acquiring real-time tire operating data). The intelligent agent collaboration layer (core layer) is the core processing layer, where various intelligent agents collaborate to complete inspection tasks. Data flow: All raw data flows to the data acquisition and standardization agent, i.e., the data acquisition and standardization intelligent agent, for unified processing. Standardized data (i.e., standard data packets) is routed and distributed through message middleware (RabbitMQ / Kafka). In this application, Agent refers to an intelligent agent. Three types of detection agents, i.e., three types of detection intelligent agents, work in parallel: Exterior defect detection agent, i.e., exterior defect detection intelligent agent: processes image data and detects surface defects; Internal structure detection agent, i.e., internal structure detection intelligent agent: processes X-ray data and detects internal defects; Operating status detection agent, i.e., operating status detection intelligent agent: processes timing signals and detects service status. All diagnostic results are aggregated to the Decision and Report Generation Agent, i.e., the decision and report generation intelligent agent, for comprehensive judgment. Output and Application Layer (Top Layer): The intelligent diagnostic results (i.e., the tire fault detection results output by the decision and report generation intelligent agent of this application) are applied to different scenarios. Output formats: Structured report: delivered to the production MES system to guide process improvement; Real-time warning: sent to the vehicle HMI or repair station to ensure driving safety; Tire digital twin file: stored on the cloud platform to form a full life cycle health record.

[0051] As an example, in the tire production or service phase, in response to triggering a tire inspection task, all raw data (images, X-rays, sensor signals) is aggregated and processed into standard data packets by a data acquisition and standardization agent. These standard data packets, based on their type and task requirements, are distributed via message queues (such as RabbitMQ) to the corresponding fault detection agents (e.g., appearance defect detection agents, internal structure detection agents, and operational status detection agents) for parallel analysis. Each fault detection agent submits preliminary results to a decision-making and report-generating agent, which performs information fusion and comprehensive decision-making, ultimately generating a structured report, pushing alerts to the MES / HMI, and updating the tire's digital profile. The entire process achieves an automated closed-loop.

[0052] Specifically, the appearance defect detection agent is based on a two-stage algorithm of normal sample prototype comparison and texture self-supervised anomaly amplification, which can solve the problem of rapid online deployment when zero defect samples are collected for new tire specifications. First, a normal prototype library is constructed. For each new tire specification (e.g., the target tire of this application), only a few dozen good product images need to be collected. Multi-level features are extracted through a pre-trained Visual Transformer (ViT), and clustered in the feature space to form a "multi-level normal feature prototype library" for that tire specification (e.g., the target tire of this application). Then, online detection and anomaly amplification are performed. For the image to be inspected of the target tire, multi-level features are also extracted. Then, on each layer of feature map, the feature distance field between it and the nearest neighbor prototype in the "multi-level normal feature prototype library" is calculated. Subsequently, a lightweight anomaly amplification network is introduced. This network takes the multi-level feature distance field as input and outputs a high-resolution, pixel-level anomaly heatmap and defect classification confidence score through learnable upsampling and attention fusion.

[0053] This application enables zero-defect sample learning and amplification of abnormal signals, and can be applied to scenarios involving the rapid deployment of new tire specifications. Specifically, it employs a two-stage pipeline: offline modeling stage: requiring only normal samples → building a specification-specific normal prototype library; online detection stage: real-time image → anomaly distance calculation → adaptive amplification → precise positioning.

[0054] Specifically, the normal prototype library construction (unsupervised learning) involves: multi-level feature extraction: backbone network: using a Vision Transformer (ViT-B / 16 or ViT-L / 16) pre-trained on ImageNet; feature layer selection: shallow features (layers 3-6): capturing texture, granularity, and local patterns; mid-level features (layers 7-10): extracting structural features and edge information; deep features (layers 11-12): containing semantic information and global context. Feature map processing: fusing the cls token and patch token of each Transformer block to generate an H×W×C feature map. Adaptive prototype clustering algorithm: for each feature level k: feature dimensionality reduction: using PCA or random projection to reduce high-dimensional features to a clusterable dimension (e.g., 64-128 dimensions). Density-aware clustering: using DBSCAN or OPTICS algorithms to automatically determine the number of "clusters" for normal patterns. Adaptive radius setting: dynamically adjusting the neighborhood radius based on the feature space density. Outlier exclusion: filtering outlier feature points in sparse regions. Prototype Extraction: For each cluster, calculate its feature center (mean or median) and save the cluster's covariance matrix to describe the feature distribution. Record the "representative weight" of each prototype (based on cluster size and density). Prototype Library Optimization Strategies: Redundant Prototype Removal: Merge prototypes that are too close in space; Boundary Prototype Enhancement: Pay special attention to boundary prototypes in the normal feature space; Multi-Scale Association: Establish correspondences between prototypes at different levels to form a hierarchical prototype map.

[0055] Specifically, online detection and anomaly amplification: feature distance field calculation: local nearest neighbor distance: for the feature map of the image to be detected. The feature vector at each position (i,j) : ,in: It is the m-th normal prototype in the k-th layer. It is the Mahalanobis distance. Considering the prototype covariance matrix, the distance field is obtained for each level. .

[0056] Multi-scale range fusion: for each range field Bilinear upsampling to a uniform resolution is performed, and scale weights are introduced: shallow distances are given higher weight (more sensitive to fine textures), generating an initial multi-scale anomaly score map. .

[0057] Lightweight Anomaly Amplification Network: Network Architecture: Input: Multi-layer feature distance field Feature pyramid encoder (3-layer CNN): Each layer: The system progressively fuses multi-scale distance information. The attention-guided upsampling module includes: cross-scale attention (high-level semantic guidance for low-level detail enhancement), spatial attention (focusing on regions of high distance difference), and channel attention (adaptively weighting contributions from different levels). The anomaly refinement decoder progressively upsamples to restore resolution, and skip connections fuse detailed information from the original distance field. The final output is a high-resolution anomaly heatmap. .

[0058] Self-supervised training strategy: Training data: Only normal samples are used, and "pseudo-anomaly" data is created through data augmentation. Pseudo-anomaly database generation methods: random local occlusion (simulating blemishes and damage), texture replacement (copying and pasting from other areas), Gaussian noise injection (simulating image acquisition noise), elastic deformation (simulating surface deformation).

[0059] Loss function: ,in, : Reconstruction error in pseudo-anomaly regions (encouragement to amplify this error). Smoothness constraint in normal regions (suppressing false alarms).

[0060] Decision and Classification Module: Pixel-level Defect Segmentation: Adaptive threshold segmentation of the abnormal heatmap H: ,in These are the mean and standard deviation of the heatmap. Adjustments based on specifications. Morphological post-processing: Remove small connected regions and fill holes. Defect classification confidence: Region feature extraction: For each candidate defect region, extract: abnormal intensity statistics (mean, maximum, variance), geometric features (area, aspect ratio, density), and texture features (contrast with surrounding normal regions). Confidence calculation: Confidence = f(abnormal intensity, region consistency, boundary clarity). Defect type inference: Based on feature pattern matching (no classifier training required): dotted highlights → bubbles, impurities; linear continuity → scratches, cracks; sheet-like uniformity → delamination, stains. This application can achieve zero-defect sample start-up: completely eliminating the need to collect defect samples, solving the cold start problem; achieve multi-scale prototype comparison: overcoming the limitations of single-scale features for complex defects; achieve learnable anomaly amplification: explicitly amplifying weak anomaly signals to improve detection sensitivity; achieve self-supervised training: training the amplification network using only normal samples to ensure generalization ability; achieve rapid deployment: new specifications require only dozens of normal images and can be deployed within hours.

[0061] The operational status detection agent enables the fusion and accurate diagnosis of multi-source asynchronous sensor signals in resource-constrained in-vehicle environments. This agent detects tire operational status through a lightweight spatiotemporal asynchronous multi-sensor fusion network. The lightweight spatiotemporal asynchronous multi-sensor fusion network includes: an asynchronous signal alignment module: for signals with different sampling rates and incompletely synchronized timestamps, such as vibration, sound, and pressure, a learnable time interpolation alignment layer is introduced to uniformly map each signal onto a virtual, equally spaced time axis, preserving their temporal relationships. A task-adaptive modal attention fusion module: this module does not simply splice features. It maintains a modal attention weight vector (i.e., the query vector for the subtask) for each diagnostic subtask (such as wear, nail puncture, or imbalance). The network automatically learns to focus more on high-frequency sound features when judging "nail puncture" and more on the low-frequency envelope features of vibration signals when judging "wear." It features a lightweight multi-task head: the network shares most of the feature extraction layers, and configures a minimalist dedicated layer for different tasks in the terminal branches, enabling the simultaneous output of multiple diagnostic results in a single forward propagation, thus meeting the real-time requirements of vehicle edge computing.

[0062] Specifically: Asynchronous signal alignment module: from "hard synchronization" to "soft alignment"; introduces a learnable time interpolation alignment layer; Virtual time axis: the network internally defines a virtual time series with equal intervals, higher than the original sampling rate of all input signals. This series does not directly correspond to any physical clock, but is a common reference frame for feature fusion. Learnable time interpolation: for each sensor signal, this layer does not use fixed linear or spline interpolation, but instead uses a small neural network (such as a multilayer perceptron) as the interpolation network. This interpolation network takes a set of original asynchronous data points near the target virtual time point and their relative time offsets as input, and outputs the signal estimate at that virtual time point. End-to-end learning: the parameters of the interpolation network are trained together with the subsequent diagnostic network. Its goal is to learn an optimal interpolation method, not to accurately reconstruct the original waveform, but to extract and retain the temporal context and features most beneficial to subsequent diagnosis. For example, for impulse signals, it learns to retain their sharp leading edge; for stationary signals, it smooths them. This preserves information in asynchronicity: overcoming the blurring effect of fixed interpolation on transient features. Data-driven: The alignment method is optimized by reverse-engineering the diagnostic task objectives, forming the optimal signal representation oriented towards the task.

[0063] Specifically, the task-adaptive modal attention fusion module moves from "static fusion" to "dynamic focusing"; employs a task-based dynamic attention mechanism; and features are encoded by passing each aligned modal signal (vibration, sound, pressure) through its own lightweight feature encoder (such as a one-dimensional CNN or Temporal Convolutional Network) to extract high-level spatiotemporal features. Task-specific attention vectors are maintained by the network for each predefined diagnostic subtask (e.g., "prick T," "wear W," "imbalance B"). Dynamic weight calculation occurs during forward propagation: the feature sequences of each modality are processed by a lightweight "modal self-description" module to generate a modal key vector representing the global state of the current modality (e.g., the key vector for the sound modality might encode the high-frequency energy information of the current segment). For the detection task T, its "modal importance query vector" is compared with the "modal key vectors" of all modalities (e.g., a dot product followed by Softmax) to obtain a set of normalized attention weights [α_vibration, α_sound, α_pressure]_T. This set of weights indicates that, given the current input sample, the detection task T should assign relative importance to each modal feature when determining the fault. For example, for the nailing task, the network may automatically learn α_sound ≈ 0.7, α_vibration ≈ 0.2, and α_pressure ≈ 0.1. Weighted attention fusion: The modal key vectors are weighted and summed according to their task-specific attention weights to generate a fusion feature vector tailored to the detection task T. This means that the same input sample will generate different fusion feature vectors for different detection tasks. This achieves accurate perception; the "nailing" task focuses on high-frequency sound, and the "wear" task focuses on the vibration envelope, achieving optimal filtering of multimodal information. Decoupling and anti-interference: When a certain modality is contaminated by noise (such as wind noise during driving), other tasks can reduce its impact by adjusting their weights, while tasks dependent on that modality (if the noise is related to the fault features) can still be effectively utilized.

[0064] Employing a lightweight multi-task head: from "serial inference" to "parallel symbiosis." Achieving synergy between a shared feature backbone and minimally sized task-specific layers. Deep shared feature backbone: Asynchronous alignment and modal attention fusion constitute the network's shared feature backbone, responsible for extracting information-rich and task-adaptively modulated general high-level features from raw multi-source asynchronous signals. Sharing significantly improves efficiency. Minimalist task-specific heads: Each detection task is connected to only one extremely lightweight output layer, typically 1-2 fully connected layers, or even a linear classifier. Its input is a "task-adaptive fused feature vector." Joint optimization and loss balancing: All lightweight task heads are computed in parallel during a single forward propagation. The total loss function is a weighted sum of the losses of each task (e.g., cross-entropy). A dynamic loss balancing strategy is employed during training (e.g., automatically adjusting weights based on task difficulty or gradient magnitude) to ensure all tasks are learned evenly, avoiding simple tasks dominating training and harming complex tasks. This achieves extreme efficiency: a single forward computation simultaneously obtains the detection results for all tire faults, meeting the real-time requirement of millisecond-level response. Knowledge symbiosis: Sharing a backbone of features enables the network to learn general representations that are beneficial for various tire faults. Different tasks achieve positive transfer at the feature level, improving overall generalization ability and few-shot learning performance. Flexible and scalable: When adding a new diagnostic task, only a corresponding "modal attention query vector" and a lightweight task head need to be added. The backbone does not need to be modified, resulting in low expansion costs.

[0065] The introduction of the operational status detection agent solves the problem of asynchronous data from multiple vehicle sensors; through "task-adaptive modal attention," the interpretability of the model and its diagnostic specificity under different fault conditions are enhanced. The introduction of the appearance defect detection agent directly performs "prototype comparison" in a high-dimensional feature space, making it more sensitive to subtle texture differences; the "anomaly amplification network" can effectively suppress background interference, accurately highlight defect areas, and significantly improve the detection rate of small defects.

[0066] In this application, "big data" refers to data from the Internet of Vehicles (IoV). "Multi-agent maintenance" refers to a multi-agent collaborative architecture and dynamic knowledge enhancement mechanism that enables a fundamental shift in tire inspection and maintenance from experience-driven to data-driven intelligence.

[0067] This application enables seamless integration of the "production-service" data chain, constructing a unified digital identity and comprehensive health record for tires. Through a clearly defined agent-based collaborative mechanism, complex multimodal detection tasks are decoupled, achieving efficient and accurate parallel processing and comprehensive decision-making. By integrating few-shot learning and multi-source information attention fusion technologies, it achieves highly robust and accurate defect diagnosis and condition assessment even in the absence of defect samples and under complex operating conditions. It provides a dynamically scalable and flexibly adaptable intelligent detection platform architecture, supporting rapid deployment for new production lines, new vehicle models, and new defect types.

[0068] Figure 4 This is a schematic diagram of the main units of a tire fault detection device according to an embodiment of this application. Figure 4 As shown, the tire fault detection device 400 includes a data acquisition unit 401, an allocation unit 402, and a comprehensive decision-making unit 403.

[0069] The data acquisition unit 401 is configured to acquire multimodal detection data for the target tire in response to triggering a tire detection task, and to determine the type of multimodal detection data and the detection task.

[0070] The allocation unit 402 is configured to allocate corresponding fault detection agents to the multimodal detection data according to the type and detection task. The fault detection agents perform fault detection on the allocated multimodal detection data based on the corresponding detection task to obtain the fault detection results. The fault detection agents include appearance defect detection agents, internal structure detection agents and operating status detection agents.

[0071] The integrated decision-making unit 403 is configured to summarize the fault detection results and input them into the decision and report generation agent. The decision and report generation agent makes an integrated decision based on the summarized fault detection results to obtain the tire fault detection results.

[0072] In some embodiments, the data acquisition unit 401 is further configured to: invoke a data acquisition and standardization agent to aggregate multimodal detection data for the target tire, wherein the multimodal detection data includes image data, X-ray data and time-series signals; perform standardization processing on the multimodal detection data to obtain standard data packets, and label the standard data packets with the corresponding type and detection task.

[0073] In some embodiments, the allocation unit 402 is further configured to: allocate standard data packets labeled as image data and with detection tasks labeled as surface defect detection to the appearance defect detection agent; allocate standard data packets labeled as X-ray data and with detection tasks labeled as internal defect detection to the internal structure detection agent; and allocate standard data packets labeled as timing signals and with detection tasks labeled as service status detection to the operating status detection agent.

[0074] In some embodiments, the allocation unit 402 is further configured to: extract multi-level features of the allocated standard data packets according to the detection task corresponding to the allocated standard data packets; calculate the feature distance field between the multi-level features and the nearest neighbor normal feature prototypes in the multi-level normal feature prototype library for each multi-level feature; wherein the multi-level normal feature prototypes are obtained by clustering the multi-level features of multiple good product images of new specification tires in the feature space; input the feature distance field into the anomaly amplification network, and the anomaly amplification network outputs an anomaly heatmap and defect classification confidence based on the input feature distance field.

[0075] In some embodiments, the defect classification confidence is obtained as follows: the anomaly amplification network calculates the mean and standard deviation of the heatmap corresponding to the anomaly heatmap; adaptive threshold segmentation is performed on the anomaly heatmap based on the mean and standard deviation of the heatmap to obtain candidate defect regions; for each candidate defect region, anomaly intensity statistics, geometric features, and texture features are extracted; feature pattern matching is performed based on the anomaly intensity statistics, geometric features, and texture features to obtain the defect type, and the defect classification confidence corresponding to each defect type is calculated.

[0076] In some embodiments, the allocation unit 402 is further configured to: map the multimodal signals in the allocated standard data packets to a virtual, equally spaced time axis to obtain aligned multimodal signals that retain temporal relationships; for each modal signal in the aligned multimodal signals that retain temporal relationships, extract the high-level spatiotemporal features corresponding to the modal signals, and generate a modality key vector representing the global state of the modal signals based on the high-level spatiotemporal features; for each detection task corresponding to the standard data packets allocated to the running state detection agent, determine the modality importance query vector corresponding to the detection task, and determine the attention weight combination for the detection task based on the modality importance query vector and each modality key vector; perform weighted attention fusion based on each modality key vector and the attention weight combination to generate a fusion feature vector for the detection task; and perform parallel computation of the lightweight task heads corresponding to all detection tasks based on the corresponding fusion feature vectors in a single forward propagation, while outputting their respective running state detection results.

[0077] In some embodiments, the tire fault detection device further includes Figure 4 The detection task addition processing unit (not shown) is configured to: add a new detection task in response to a detection task addition request, and configure a modal importance query vector and a lightweight task header for the new detection task.

[0078] It should be noted that the tire fault detection method and tire fault detection device in this application are related in terms of specific implementation, so the repeated content will not be described again.

[0079] Figure 5 An exemplary system architecture 500 is shown that can be applied to the tire fault detection method or tire fault detection device according to the embodiments of this application.

[0080] like Figure 5 As shown, system architecture 500 may include terminal devices 501, 502, and 503, a network 504, and a server 505. Network 504 serves as the medium for providing communication links between terminal devices 501, 502, and 503 and server 505. Network 504 may include various connection types, such as wired or wireless communication links, or fiber optic cables, etc.

[0081] Users can use terminal devices 501, 502, and 503 to interact with server 505 via network 504 to receive or send messages, etc. Various communication client applications can be installed on terminal devices 501, 502, and 503, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, social media platform software, etc. (for example only).

[0082] Terminal devices 501, 502, and 503 can be various electronic devices with a tire fault detection and processing screen that supports web browsing, including but not limited to smartphones, tablets, laptops, and desktop computers.

[0083] Server 505 can be a server providing various services, such as a backend management server supporting tire inspection tasks triggered by users using terminal devices 501, 502, and 503 (for example only). In response to a triggered tire inspection task, the backend management server can acquire multimodal inspection data for the target tire, determine the type of multimodal inspection data and the inspection task; assign corresponding fault detection agents to the multimodal inspection data according to the type and inspection task; the fault detection agents perform fault detection on the assigned multimodal inspection data based on the corresponding inspection task, obtaining fault detection results. These fault detection agents include appearance defect detection agents, internal structure detection agents, and operating status detection agents; the fault detection results are summarized and input into a decision and report generation agent; the decision and report generation agent makes a comprehensive decision based on the summarized fault detection results, obtaining the tire fault detection result. This can effectively reduce tire fault detection costs and improve tire fault detection efficiency and accuracy.

[0084] It should be noted that the tire fault detection method provided in this application embodiment is generally executed by server 505, and correspondingly, the tire fault detection device is generally set in server 505.

[0085] It should be understood that Figure 5The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0086] The following is for reference. Figure 6 It shows a schematic diagram of the structure of a computer system 600 suitable for implementing a terminal device according to the embodiments of this application. Figure 6 The terminal device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0087] like Figure 6 As shown, the computer system 600 includes a central processing unit (CPU) 601, which can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) 602 or programs loaded from storage section 608 into random access memory (RAM) 603. The RAM 603 also stores various programs and data required for the operation of the computer system 600. The CPU 601, ROM 602, and RAM 603 are interconnected via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0088] The following components are connected to I / O interface 605: an input section 606 including a keyboard, mouse, etc.; an output section 607 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 608 including a hard disk, etc.; and a communication section 609 including a network interface card such as a LAN card, modem, etc. The communication section 609 performs communication processing via a network such as the Internet. A drive 610 is also connected to I / O interface 605 as needed. A removable medium 611, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on drive 610 as needed so that computer programs read from it can be installed into storage section 608 as needed.

[0089] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 609, and / or installed from removable medium 611. When the computer program is executed by central processing unit (CPU) 601, it performs the functions defined above in the system of this application.

[0090] It should be noted that the computer-readable medium shown in this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. Computer-readable storage media can be, for example, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media can also be any computer-readable medium other than computer-readable storage media, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0091] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0092] The units described in the embodiments of this application can be implemented in software or hardware. The described units can also be housed in a processor; for example, a processor can be described as including a data acquisition unit, an allocation unit, and a comprehensive decision-making unit. The names of these units do not necessarily limit the specific unit itself.

[0093] In another aspect, this application also provides a computer-readable medium, which may be included in the device described in the above embodiments; or it may exist independently and not assembled into the device. The computer-readable medium carries one or more programs that, when executed by the device, cause the device to respond to triggering a tire detection task, acquire multimodal detection data for a target tire, determine the type of multimodal detection data and the detection task; assign corresponding fault detection agents to the multimodal detection data according to the type and detection task, and have the fault detection agents perform fault detection on the assigned multimodal detection data based on the corresponding detection task to obtain fault detection results. The fault detection agents include an appearance defect detection agent, an internal structure detection agent, and an operating status detection agent; summarize the fault detection results and input them to a decision and report generation agent, which makes a comprehensive decision based on the summarized fault detection results to obtain the tire fault detection result.

[0094] The computer program product of this application includes a computer program that, when executed by a processor, implements the tire fault detection method in the embodiments of this application.

[0095] The technical solution according to the embodiments of this application can effectively reduce the cost of tire fault detection and improve the efficiency and accuracy of tire fault detection.

[0096] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can occur depending on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A tire fault detection method, characterized in that, include: In response to triggering a tire inspection task, acquire multimodal inspection data for the target tire, and determine the type of the multimodal inspection data and the inspection task; According to the type and the detection task, a corresponding fault detection agent is assigned to the multimodal detection data. The fault detection agent performs fault detection on the assigned multimodal detection data based on the corresponding detection task to obtain the fault detection result. The fault detection agent includes an appearance defect detection agent, an internal structure detection agent, and an operating status detection agent. The fault detection results are summarized and input into the decision and report generation agent. The decision and report generation agent makes a comprehensive decision based on the summarized fault detection results to obtain the tire fault detection results.

2. The method according to claim 1, characterized in that, The process of acquiring multimodal detection data for the target tire and determining the type of the multimodal detection data and the detection task includes: The system calls upon a data acquisition and standardized intelligent agent to aggregate multimodal detection data for the target tire, wherein the multimodal detection data includes image data, X-ray data, and time-series signals; The multimodal detection data is standardized to obtain a standard data package, and the standard data package is labeled with the corresponding type and detection task.

3. The method according to claim 2, characterized in that, The step of assigning a corresponding fault detection agent to the multimodal detection data according to the type and the detection task includes: A standard data package, labeled as image data and as surface defect detection task, is assigned to the appearance defect detection agent. Assign standard data packets, labeled as X-ray data and with the detection task labeled as detecting internal defects, to the internal structure detection agent. Standard data packets, labeled as time-series signals and as service status detection tasks, are assigned to the operational status detection agent.

4. The method according to claim 3, characterized in that, The fault detection agent performs fault detection on the assigned multimodal detection data based on the corresponding detection task to obtain fault detection results, including: The appearance defect detection agent extracts multi-level features of the assigned standard data packets according to the detection task corresponding to the assigned standard data packets; For each level feature in the multi-level features, the feature distance field between the level feature and the nearest neighbor normal feature prototype in the multi-level normal feature prototype library is calculated, wherein the multi-level normal feature prototype is obtained by clustering the multi-level features of multiple good product images of new specification tires in the feature space. The feature distance field is input into the anomaly amplification network, which outputs an anomaly heatmap and defect classification confidence based on the input feature distance field.

5. The method according to claim 4, characterized in that, The defect classification confidence level is obtained based on the following method: The anomaly amplification network calculates the mean and standard deviation of the heatmap corresponding to the anomaly heatmap; Based on the mean and standard deviation of the heatmap, an adaptive threshold segmentation is performed on the abnormal heatmap to obtain candidate defect regions; For each candidate defect region, extract the abnormal intensity statistics, geometric features, and texture features; Based on the abnormal intensity statistics, the geometric features, and the texture features, feature pattern matching is performed to obtain the defect type, and the defect classification confidence level corresponding to each defect type is calculated.

6. The method according to claim 3, characterized in that, The fault detection agent performs fault detection on the assigned multimodal detection data based on the corresponding detection task to obtain fault detection results, including: The running status detection agent maps the multimodal signals in the allocated standard data packets onto a virtual, equally spaced time axis to obtain aligned multimodal signals that retain the timing relationship. For each modal signal in the aligned multimodal signal that preserves the temporal relationship, high-level spatiotemporal features corresponding to the modal signal are extracted, and a modal key vector representing the global state of the modal signal is generated based on the high-level spatiotemporal features. For each detection task corresponding to the standard data packet allocated to the operational status detection agent, a modality importance query vector corresponding to the detection task is determined, and an attention weight combination for the detection task is determined based on the modality importance query vector and each modality key vector. Based on the combination of each modality key vector and the attention weight, weighted attention fusion is performed to generate a fusion feature vector for the detection task; The lightweight task heads for all detection tasks are computed in parallel based on their respective fused feature vectors during a single forward propagation, and each task outputs its own running status detection results.

7. The method according to claim 6, characterized in that, Before obtaining the fault detection result, the method further includes: In response to a request to add a new detection task, a new detection task is added, and a modal importance query vector and a lightweight task header are configured for the new detection task.

8. A tire fault detection device, characterized in that, include: The data acquisition unit is configured to acquire multimodal detection data for a target tire in response to triggering a tire detection task, and to determine the type of the multimodal detection data and the detection task. The allocation unit is configured to allocate a corresponding fault detection agent to the multimodal detection data according to the type and the detection task. The fault detection agent performs fault detection on the allocated multimodal detection data based on the corresponding detection task to obtain a fault detection result. The fault detection agent includes an appearance defect detection agent, an internal structure detection agent, and an operating status detection agent. The integrated decision-making unit is configured to summarize the fault detection results and input them into the decision and report generation agent, which then makes an integrated decision based on the summarized fault detection results to obtain the tire fault detection results.

9. An electronic device for tire fault detection, characterized in that, include: One or more processors; Storage device for storing one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-7.

10. A computer-readable medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-7.

11. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-7.

Images