Intelligent flaw detection method, system and device based on data warehouse mode

Abstract

The application discloses a wisdom flaw detection method, system and device based on a data warehouse mode, relates to the technical field of data management, and comprises the following steps: receiving a flaw detection task, wherein the detection task contains a task object and a task intention; generating corresponding flaw semantic information based on a flaw semantic reasoning model and according to semantic deduction of the flaw detection task; performing semantic matching on the flaw semantic information based on a dynamic semantic addressing mapping library to obtain a sub-data acquisition instruction; acquiring an original flaw signal through a virtualized fact generation layer and according to the sub-data acquisition instruction; dynamically calling a flaw physical operator and an operator parameter based on the task intention, performing feature extraction on the original flaw signal to generate a damage feature view; and generating a flaw detection report based on the damage feature view. The application has the effect of improving the detection logic flexibility of a flaw detection scene.

Description

Technical Field

[0001] This application relates to the technical field of data governance, and in particular to intelligent flaw detection methods, systems and equipment based on data warehouse models. Background Technology

[0002] In the field of industrial nondestructive testing, especially in the in-service flaw detection of critical facilities such as railways, aviation, and energy, the inspection scenarios are significantly diverse and dynamic. The inspection objects involve various physical components such as rails, wheels, bearings, and blades, whose materials, structures, and potential damage types are complex and varied, including cracks, peeling, porosity, and wear. The inspection methods encompass multimodal sensing technologies such as ultrasonic, eddy current, magnetic particle, X-ray, and vision, and the signals generated by different methods differ significantly in waveform characteristics, data dimensions, and physical meaning. Simultaneously, the objectives of on-site inspection tasks vary, ranging from large-scale rapid surveys and screenings to detailed diagnosis of suspicious areas, or qualitative confirmation and quantitative assessment of specific types of damage. Faced with such scenarios, the flaw detection system needs to flexibly organize and coordinate multiple detection signals to form a complete damage judgment basis.

[0003] Currently, flaw detection operations generally employ testing equipment and systems based on fixed procedures. The core of these systems lies in pre-embedding signal acquisition, feature processing, and damage assessment logic into the program or testing process. During actual testing, operators select the appropriate testing instruments and probes based on the object and purpose of the inspection, and perform scanning according to a predetermined standardized procedure. The instruments process the signals in real time according to preset thresholds and algorithms, providing indications of the presence or absence of defects, or saving the acquired raw waveforms and image data locally in a fixed format for later playback and analysis by professionals. To achieve centralized data management, some systems upload testing data to databases or data warehouses; however, these data warehouses are primarily used for archiving, querying, and generating statistical reports of historical data, and do not participate in the dynamic driving of the testing process. In multi-method collaborative diagnosis, it is usually necessary to independently complete data acquisition using different instruments such as ultrasound and eddy current. Operators or analysts manually export data from each instrument, and with the help of general analysis software or relying solely on personal experience, manually compare different signals along the time axis and spatial location to comprehensively determine the damage status. Different acquisition devices generate data with different formats, interface protocols, and parameter configurations. Typically, it is necessary to develop dedicated data reading and format conversion modules for each device, which is a lot of work and difficult to expand.

[0004] The limitations of this existing technology are becoming increasingly apparent in complex flaw detection applications. First, the detection logic is highly coupled with specific detection equipment, algorithms, parameters, and data storage formats. Whenever the detection object, method, or damage assessment criteria change, it often requires redevelopment or extensive configuration of software processes, making it difficult for the system to quickly adapt to changes in the detection task. Second, the process of determining the required detection evidence from the task objective relies almost entirely on professionals' understanding and manual design of detection standards, procedures, and their own experience. The system cannot automatically interpret the detection intent and plan which types and combinations of signal data need to be acquired, resulting in a high degree of manual dependence on conversion between different tasks, leading to low efficiency and difficulty in ensuring consistency. Furthermore, in the comprehensive diagnosis of multi-source heterogeneous signals, the original signals are scattered across different file directories and storage locations on different instruments. Analysts need to know the specific storage path and format of the required data, manually searching, exporting, and loading them one by one. It is impossible to automatically and efficiently acquire data based on semantic requirements such as "a certain type of signal in a certain part." Furthermore, due to the lack of a unified organization and access mechanism for heterogeneous data, flaw detection signals with various formats and acquisition parameters expose a large number of underlying details when used. Analysts not only need to be familiar with damage identification knowledge, but also need to understand the access methods of different data sources, which greatly increases the workload and limits the development of automated and intelligent integrated diagnosis.

[0005] Therefore, there is an urgent need in this field for a technical solution that can improve the flexibility of flaw detection, reduce reliance on manual labor, and improve the efficiency of multi-source evidence collaboration. It is hoped that the required detection evidence can be automatically organized according to different detection task intentions, the original flaw detection signals distributed in heterogeneous storage can be dynamically acquired, and the corresponding processing methods can be flexibly combined to generate the required damage analysis view, thereby improving the level of automation and intelligence of detection in complex flaw detection scenarios and the adaptability to changing tasks. Summary of the Invention

[0006] To address the issue of rigid detection logic in flaw detection scenarios, this application provides a smart flaw detection method, system, and equipment based on a data warehouse model.

[0007] Firstly, this application provides a smart flaw detection method based on a data warehouse model, employing the following technical solution: Intelligent flaw detection methods based on the data warehouse model include: Receive a flaw detection task, wherein the detection task includes a task object and a task intent; Based on the flaw detection semantic reasoning model, and based on the flaw detection task, semantic inference is performed to generate corresponding flaw detection semantic information. The flaw detection semantic reasoning model is used to characterize the semantic association pattern between dynamically expandable physical objects, detection methods and damage states, as well as the semantic reasoning rules corresponding to different task intentions. The flaw detection semantic information is used to characterize the evidence logic structure required to complete the task intention, as well as the semantic roles and collaboration patterns of each piece of evidence. Based on the dynamic semantic addressing mapping library, semantic matching is performed on the flaw detection semantic information to obtain sub-data acquisition instructions. The dynamic semantic addressing mapping library is used to characterize the many-to-many dynamic conversion rules between the semantic roles and physical storage locations. The original flaw detection signal is obtained through the virtualized fact generation layer and according to the sub-data acquisition instruction. The virtualized fact generation layer is used to characterize the virtualized organization logic for multi-source heterogeneous flaw detection data. Based on the task intent, the physical operators for flaw detection and their parameters are dynamically invoked to extract features from the original flaw detection signal and generate a damage feature view. Based on the damage feature view, a flaw detection report is generated.

[0008] By adopting the above technical solution, a flaw detection task is received. The detection task includes a task object and a task intent. Then, based on the flaw detection semantic reasoning model, semantic inference is performed according to the flaw detection task to generate corresponding flaw detection semantic information. The flaw detection semantic reasoning model is used to represent the semantic association pattern between dynamically expandable physical objects, detection methods, and damage states, as well as the semantic reasoning rules corresponding to different task intents. The flaw detection semantic information is used to represent the logical structure of evidence required to complete the task intent, as well as the semantic roles and collaboration patterns of each piece of evidence. Then, based on the dynamic semantic addressing mapping library, semantic matching is performed on the flaw detection semantic information to obtain sub-data acquisition instructions. The dynamic semantic addressing mapping library is used to represent the many-to-many dynamic conversion rules between semantic roles and physical storage locations. Then, through the virtualized fact generation layer, the original flaw detection signal is acquired according to the sub-data acquisition instructions. The virtualized fact generation layer is used to represent the virtualized organization logic for multi-source heterogeneous flaw detection data. Then, based on the task intent, the flaw detection physical operators and operator parameters are dynamically called to extract features from the original flaw detection signal and generate a damage feature view. Finally, based on the damage feature view, a flaw detection report is generated. This invention transforms the traditional fixed detection logic into an intent-driven dynamic detection mechanism. That is, it automatically organizes detection evidence according to the detection task intent, dynamically acquires the original flaw detection signals in heterogeneous storage, and flexibly combines multiple processing methods to generate a damage analysis view. This improves the flexibility of flaw detection, reduces reliance on human experience, enhances the collaborative efficiency of multi-source evidence, and thus improves the level of automation, intelligence, and adaptability to changing tasks in complex flaw detection scenarios.

[0009] Optionally, the step of generating corresponding flaw detection semantic information based on the flaw detection semantic reasoning model and performing semantic inference according to the flaw detection task includes: The flaw detection task is loaded into the flaw detection semantic reasoning model, wherein the flaw detection semantic reasoning model includes dynamically expandable physical object nodes, detection method nodes and damage state nodes, as well as dynamically expandable semantic association patterns and semantic reasoning rules. Semantic parsing is performed on the flaw detection task to activate the target physical object node corresponding to the task object, the target detection method node associated with the task intent, and the target damage status node associated with the task intent. Based on the semantic association pattern, all reachable paths from the target physical object node to the target damage state node via the target detection method node are extracted as candidate evidence chains. Based on the semantic reasoning rules, confidence is evaluated for each candidate evidence chain, and chain weights are assigned to each candidate evidence chain according to the evaluation results; and candidate evidence chains whose chain weights satisfy preset conditions are selected to form a weighted evidence chain set. Semantic roles are assigned to nodes on each candidate evidence chain in the weighted evidence chain set, and a weighted collaboration mode between each semantic role is determined based on the connection relationship between nodes and the chain weight. The weighted collaboration mode is a multi-evidence chain collaboration mechanism based on the chain weight, which is used to constrain the collaboration relationship between the semantic roles of different candidate evidence chains. Based on the topological structure of each candidate evidence chain in the weighted evidence chain set, as well as the corresponding semantic roles and weighted collaboration modes, flaw detection semantic information is generated.

[0010] By adopting the above technical solution, in order to achieve semantic inference, the flaw detection task is loaded into the flaw detection semantic reasoning model. This model includes dynamically expandable physical object nodes, detection method nodes, and damage state nodes, as well as dynamically expandable semantic association patterns and semantic reasoning rules. Then, the flaw detection task is semantically parsed, activating the target physical object node corresponding to the task object, the target detection method node associated with the task intent, and the target damage state node associated with the task intent. Based on the semantic association pattern, all reachable paths from the target physical object node through the target detection method node to the target damage state node are extracted as candidate evidence chains. Finally, based on semantic... The reasoning rules involve evaluating the confidence level of each candidate evidence chain and assigning a chain weight to each candidate evidence chain based on the evaluation results; selecting candidate evidence chains whose chain weights meet preset conditions to form a weighted evidence chain set; assigning semantic roles to nodes on each candidate evidence chain in the weighted evidence chain set; and determining the weighted collaboration mode between each semantic role based on the connection relationship between nodes and the chain weight. The weighted collaboration mode is a multi-evidence chain collaboration mechanism based on chain weight, used to constrain the collaboration relationship between the semantic roles of different candidate evidence chains. Finally, based on the topology of each candidate evidence chain in the weighted evidence chain set, as well as the corresponding semantic roles and weighted collaboration modes, flaw detection semantic information is generated.

[0011] Optionally, the step of evaluating the confidence of each candidate evidence chain based on the semantic reasoning rules and assigning a chain weight to each candidate evidence chain according to the evaluation results includes: Based on the flaw detection semantic reasoning model, the reasoning success rate of each candidate evidence chain is obtained, wherein the reasoning success rate is used to characterize the frequency with which the reasoning conclusion of the candidate evidence chain in historical flaw detection tasks is consistent with the subsequent confirmation result. Based on the reasoning success rate, the chain weight of each candidate evidence chain is determined, wherein the reasoning success rate is positively correlated with the chain weight; After the completion of the flaw detection task, the reasoning conclusions of each candidate evidence chain in the weighted evidence chain set are compared with the corresponding post-confirmation results, and the reasoning success rate of each candidate evidence chain in the flaw detection semantic reasoning model is updated according to the correctness of the comparison.

[0012] By adopting the above technical solution, in order to generate chain weights, the reasoning success rate of each candidate evidence chain is obtained based on the flaw detection semantic reasoning model. The reasoning success rate is used to characterize the frequency with which the reasoning conclusions of the candidate evidence chain in historical flaw detection tasks are consistent with the post-confirmation results. Then, based on the reasoning success rate, the chain weight of each candidate evidence chain is determined. The reasoning success rate is positively correlated with the chain weight. After the current flaw detection task is completed, the reasoning conclusions of each candidate evidence chain in the current weighted evidence chain set are compared with the corresponding post-confirmation results, and the reasoning success rate of each candidate evidence chain in the flaw detection semantic reasoning model is updated according to the correctness of the comparison.

[0013] Optionally, the step of semantically matching the flaw detection semantic information based on a dynamic semantic addressing mapping library to obtain sub-data acquisition instructions includes: Semantic roles are extracted from the flaw detection semantic information to obtain target semantic roles, wherein the target semantic roles include the inspected object, the detection method, and the target defect; Based on the dynamic semantic addressing mapping library, and according to the target semantic roles, a structured addressing result is obtained. The structured addressing result is generated by the many-to-many dynamic conversion rules maintained in the dynamic semantic addressing mapping library. The structured addressing result is grouped according to the physical storage location. Each group corresponds to a physical storage location, and each group contains all the target semantic roles mapped to that physical storage location and the signal segment extraction range or metadata filtering conditions corresponding to each target semantic role. Based on the structured addressing result, a sub-data retrieval instruction is generated, wherein the sub-data retrieval instruction contains only one access request targeting the physical storage location corresponding to the group for each group in the structured addressing result.

[0014] By adopting the above technical solution, in order to achieve semantic matching, semantic roles are extracted from the flaw detection semantic information to obtain target semantic roles. The target semantic roles include the inspected object, the detection method, and the target defect. Then, based on the dynamic semantic addressing mapping library, semantic addressing is performed according to the target semantic roles to obtain structured addressing results. The structured addressing results are generated by the many-to-many dynamic transformation rules maintained in the dynamic semantic addressing mapping library. The structured addressing results are grouped according to physical storage location. Each group corresponds to a physical storage location, and each group contains all target semantic roles mapped to that physical storage location and the signal segment extraction range or metadata filtering conditions corresponding to each target semantic role. Then, based on the structured addressing results, sub-data acquisition instructions are generated. The sub-data acquisition instructions contain only one access request targeting the physical storage location corresponding to each group in the structured addressing results.

[0015] Optionally, the step of acquiring the original flaw detection signal through the virtualized fact generation layer and according to the sub-data acquisition instruction includes: In the virtualization fact generation layer, the sub-data acquisition instructions are parsed to obtain the target physical storage location corresponding to each access request, as well as the target semantic role associated with each target physical storage location and the corresponding extraction conditions. The extraction conditions include the signal segment extraction range or the metadata filtering conditions. The virtualization fact generation layer, based on the data source type corresponding to each target physical storage location, calls the data access interface that matches the data source type to initiate a data read request to the corresponding target physical storage location; The virtualization fact generation layer receives the raw data returned from each of the target physical storage locations, and extracts the raw data according to the extraction conditions associated with each of the target physical storage locations to obtain the raw flaw detection signal. The raw flaw detection signal is uniformly managed by the virtualization organization logic of the virtualization fact generation layer to shield the underlying storage differences of multi-source heterogeneous flaw detection data.

[0016] By adopting the above technical solution, in order to obtain the original flaw detection signal, the sub-data acquisition instructions are parsed in the virtualization fact generation layer to obtain the target physical storage location corresponding to each access request, as well as the target semantic role associated with each target physical storage location and the corresponding extraction conditions. The extraction conditions include the signal segment extraction range or metadata filtering conditions. Then, according to the data source type corresponding to each target physical storage location, the virtualization fact generation layer calls the data access interface that matches the data source type to initiate a data read request to the corresponding target physical storage location. Then, the virtualization fact generation layer receives the original data returned by each target physical storage location and, according to the extraction conditions associated with each target physical storage location, truncates the original data to obtain the original flaw detection signal. The original flaw detection signal is uniformly managed by the virtualization organization logic of the virtualization fact generation layer to shield the underlying storage differences of multi-source heterogeneous flaw detection data.

[0017] Optionally, the step of dynamically invoking the flaw detection physical operator and operator parameters based on the task intent to extract features from the original flaw detection signal and generate a damage feature view includes: Based on the task intent, a search is performed in the operator orchestration knowledge base to obtain multiple candidate operator templates that match the task intent. The candidate operator template is a directed acyclic graph structure. The candidate operator template is used to define the signal processing technology. The candidate operator template includes a flaw detection physical operator and a predetermined connection relationship and data flow between the flaw detection physical operators. Moreover, each flaw detection physical operator in the same candidate operator template is for the same detection signal type. Based on the collaborative mode of each semantic role, the candidate operator templates are arranged to generate an operator execution pipeline, and initial parameters are configured for each flaw detection physical operator in the operator execution pipeline. The collaborative mode is used to guide the arrangement of multiple candidate operator templates, and each flaw detection physical operator in the operator execution pipeline has an arrangement hierarchy and data flow path determined by the arrangement. Based on the operator execution pipeline, feature extraction is performed on the original flaw detection signal, and the current output feature quality index of each flaw detection physical operator is monitored in real time during the execution process. The output feature quality index includes at least one of signal-to-noise ratio, waveform smoothness, spectral peak clarity, energy concentration and feature class separability. When the output feature quality index of any flaw detection physical operator is lower than a preset threshold, a parameter adaptive adjustment mechanism is triggered. Based on the deviation between the current output feature quality index and the expected quality index, the parameters of the downstream flaw detection physical operator are corrected until the operator execution pipeline is completed and a damage feature view is generated.

[0018] By adopting the above technical solution, in order to generate a damage feature view, a search is performed in the operator orchestration knowledge base according to the task intent to obtain multiple candidate operator templates that match the task intent. The candidate operator templates are directed acyclic graph structures and are used to define the signal processing technology. Each candidate operator template includes flaw detection physical operators and predetermined connection relationships and data flow directions between them. Furthermore, each flaw detection physical operator within the same candidate operator template targets the same detection signal type. Then, based on the collaborative mode of each semantic role, the candidate operator templates are orchestrated to generate an operator execution pipeline, and initial parameters are configured for each flaw detection physical operator in the pipeline. The collaborative mode guides the orchestration method among the multiple candidate operator templates. The various physical operators in the operator execution pipeline have a predetermined arrangement hierarchy and data flow path. Then, based on the operator execution pipeline, features are extracted from the original flaw detection signal. During the execution process, the current output feature quality index of each physical operator is monitored in real time. The output feature quality index includes at least one of the following: signal-to-noise ratio, waveform smoothness, peak spectral clarity, energy concentration, and feature class separability. When the output feature quality index of any physical operator is lower than a preset threshold, a parameter adaptive adjustment mechanism is triggered. Based on the deviation between the current output feature quality index and the expected quality index, the parameters of the downstream physical operators are corrected until the operator execution pipeline is completed and a damage feature view is generated.

[0019] Optionally, the step of orchestrating the candidate operator templates based on the collaborative patterns of each semantic role to generate an operator execution pipeline includes: The collaborative mode is analyzed to determine the corresponding orchestration logic. The orchestration logic is used to characterize the orchestration rules between different signal processing techniques. The orchestration logic defines the connection method between each flaw detection physical operator in the operator execution pipeline. The connection method includes parallel connection and fusion connection. Based on the orchestration logic, the candidate operator templates are orchestrated to generate an operator execution pipeline.

[0020] By adopting the above technical solution, in order to generate the operator execution pipeline, the cooperative mode is analyzed and the corresponding orchestration logic is determined. The orchestration logic is used to characterize the orchestration rules between different signal processing processes. The orchestration logic defines the connection method between each flaw detection physical operator in the operator execution pipeline, and the connection method includes parallel connection and fusion connection. Then, based on the orchestration logic, the candidate operator templates are orchestrated to generate the operator execution pipeline.

[0021] Optionally, the step of correcting the parameters of the downstream flaw detection physical operator based on the deviation between the current output feature quality index and the expected quality index includes: Based on the current output feature quality index and the expected quality index, the deviation value is determined; Based on the deviation value, a parameter correction strategy corresponding to the deviation value is determined from a preset parameter correction rule base. The parameter correction strategy includes the target parameter to be corrected in the downstream flaw detection physical operator, as well as the correction amount and correction direction of the target parameter. Based on the aforementioned parameter correction strategy, the downstream flaw detection physical operator is corrected.

[0022] By adopting the above technical solution, in order to correct the parameters of the downstream flaw detection physical operator, the deviation value is determined based on the current output feature quality index and the expected quality index. Then, based on the deviation value, the parameter correction strategy corresponding to the deviation value is determined from the preset parameter correction rule base. The parameter correction strategy includes the target parameter to be corrected in the downstream flaw detection physical operator, as well as the correction amount and correction direction of the target parameter. Then, the downstream flaw detection physical operator is corrected based on the parameter correction strategy.

[0023] Secondly, this application also provides an intelligent flaw detection system based on a data warehouse model, which adopts the following technical solution: A smart flaw detection system based on a data warehouse model includes: The task receiving module is used to receive flaw detection tasks, wherein the detection task includes a task object and a task intent; The semantic reasoning module is used to generate corresponding flaw detection semantic information based on the flaw detection semantic reasoning model and the flaw detection task. The flaw detection semantic reasoning model is used to characterize the semantic association pattern between dynamically expandable physical objects, detection methods and damage states, as well as the semantic reasoning rules corresponding to different task intentions. The flaw detection semantic information is used to characterize the evidence logic structure required to complete the task intention, as well as the semantic role and collaboration mode of each piece of evidence. The semantic addressing module is used to perform semantic matching on the flaw detection semantic information based on the dynamic semantic addressing mapping library to obtain sub-data acquisition instructions. The dynamic semantic addressing mapping library is used to represent the many-to-many dynamic conversion rules between the semantic roles and physical storage locations. The virtualized data acquisition module is used to acquire the original flaw detection signal through the virtualized fact generation layer and according to the sub-data acquisition instruction, wherein the virtualized fact generation layer is used to characterize the virtualized organization logic for multi-source heterogeneous flaw detection data; The feature extraction module is used to dynamically call the flaw detection physical operator and operator parameters based on the task intent, extract features from the original flaw detection signal, and generate a damage feature view. The inspection report generation module is used to generate a flaw detection report based on the damage feature view.

[0024] Thirdly, this application also provides a computer device, which adopts the following technical solution: A computer device includes a memory and a processor, the memory storing a computer program executable on the processor, the processor executing the computer program to implement the method described in the first aspect.

[0025] In summary, this application includes at least the following beneficial technical effects: receiving a flaw detection task, wherein the detection task includes a task object and a task intent; then, based on a flaw detection semantic reasoning model and according to the flaw detection task, generating corresponding flaw detection semantic information; wherein the flaw detection semantic reasoning model is used to characterize the semantic association pattern between dynamically expandable physical objects, detection methods, and damage states, as well as the semantic reasoning rules corresponding to different task intents; the flaw detection semantic information is used to characterize the logical structure of evidence required to complete the task intent, as well as the semantic roles and collaborative patterns of each piece of evidence; and then, based on dynamic semantic search... The invention employs a dynamic semantic addressing mapping library to perform semantic matching on flaw detection semantic information, generating sub-data acquisition instructions. This library represents the many-to-many dynamic conversion rules between semantic roles and physical storage locations. Then, a virtualized fact generation layer, based on the sub-data acquisition instructions, acquires the original flaw detection signal. This virtualized fact generation layer represents the virtualized organization logic for multi-source heterogeneous flaw detection data. Based on the task intent, flaw detection physical operators and their parameters are dynamically invoked to extract features from the original flaw detection signal, generating a damage feature view. Finally, a flaw detection report is generated based on this damage feature view. This invention transforms the traditional fixed detection logic into an intent-driven dynamic detection mechanism. It automatically organizes detection evidence according to the detection task intent, dynamically acquires original flaw detection signals from heterogeneous storage, and flexibly combines multiple processing methods to generate a damage analysis view. This improves the flexibility of flaw detection, reduces reliance on human experience, enhances the collaborative efficiency of multi-source evidence, and ultimately improves the automation level, intelligence, and adaptability to changing tasks in complex flaw detection scenarios. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the overall process of an embodiment of this application.

[0027] Figure 2 This is a schematic diagram of the system structure of this application.

[0028] Figure 3 This is a structural block diagram of the computer device described in this application. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0030] This application discloses an intelligent flaw detection method based on a data warehouse model.

[0031] Reference Figure 1 Intelligent flaw detection methods based on the data warehouse model include: Step S11: Receive a flaw detection task, wherein the detection task includes a task object and a task intent.

[0032] It's important to note that in complex industrial flaw detection scenarios, the specific objectives of on-site operators vary greatly. They might need to conduct a large-scale, rapid survey of an entire section of rail, or perform a detailed diagnosis of a specific part of a bearing to confirm the presence of fatigue cracks. Therefore, the received inspection task must clearly include two layers of information: the "task object" and the "task intent." In current technology, the system's processing logic is fixed; a single inspection command can only trigger a pre-defined, single device action, failing to convey the difference in intent—whether it's a "general survey" or a "detailed survey"—to the backend data organization and analysis methods. This step, however, explicitly uses the task intent as input, enabling subsequent steps to automatically distinguish between them.

[0033] Step S12: Based on the flaw detection semantic reasoning model and according to the flaw detection task, semantic inference is performed to generate corresponding flaw detection semantic information. The flaw detection semantic reasoning model is used to characterize the semantic association pattern between dynamically expandable physical objects, detection methods and damage states, as well as the semantic reasoning rules corresponding to different task intentions. The flaw detection semantic information is used to characterize the logical structure of evidence required to complete the task intention, as well as the semantic roles and collaborative patterns of each piece of evidence.

[0034] It should be noted that step S12 aims to address the pain point of current technology where "evidence organization relies entirely on human experience." In traditional operations, when a specific diagnostic intent needs to be achieved, analysts must rely on a deep understanding of testing standards, damage mechanisms, and process characteristics to manually plan the combination of required testing methods, signal types, and the collaborative logic between signals to corroborate each other or eliminate false defects. This method of evidence arrangement, entirely dependent on expert judgment, is inefficient and inconsistent. In this step, the flaw detection semantic reasoning model explicitly represents the inherent and scalable correlation patterns between physical objects, testing methods, and damage states, as well as the reasoning rules corresponding to different intents. After receiving the task object and task intent, the model automatically derives a complete set of flaw detection semantic information necessary to complete the task—a structured list of evidence and collaborative logic—providing precise high-level semantic guidance for subsequent automatic data acquisition.

[0035] Step S13: Based on the dynamic semantic addressing mapping library, perform semantic matching on the flaw detection semantic information to obtain sub-data acquisition instructions. The dynamic semantic addressing mapping library is used to represent the many-to-many dynamic conversion rules between semantic roles and physical storage locations.

[0036] It should be noted that step S13 addresses the problem of "rigid physical data addressing" in existing technologies. In real-world environments where multi-source heterogeneous data coexists, the aforementioned derived evidence requirements may not be stored in a fixed folder or database; they may be scattered across A-scan waveform files, full-matrix capture data arrays, or even comparison data from different versions. Traditionally, operators must know exactly which disk directory these data are located in and manually load them. The dynamic semantic addressing mapping library in this step establishes a flexible conversion rule from semantic roles to physical storage locations. This mapping relationship is many-to-many and dynamically changeable. It intelligently matches and decomposes the abstract evidence requirements obtained in the previous step into a set of specific, executable sub-data acquisition instructions. These instructions precisely point to specific data fragments in the physical storage layer, preparing for virtualized data acquisition.

[0037] Step S14: Obtain the original flaw detection signal through the virtualized fact generation layer and according to the sub-data acquisition instruction. The virtualized fact generation layer is used to characterize the virtualized organization logic for multi-source heterogeneous flaw detection data.

[0038] It should be noted that step S14 aims to address the problem of "varied access methods and integration difficulties for heterogeneous data" in current technologies. In the traditional model, flaw detection data from different manufacturers' equipment, such as ultrasonic instruments, eddy current analyzers, and vision systems, have completely different underlying formats, sampling rates, and coordinate definitions. Analysis software needs to interface with numerous dedicated data interfaces and conversion modules, resulting in significant development effort and limited scalability. The virtualized fact generation layer in this step constructs a unified virtualized organization logic for multi-source heterogeneous flaw detection data. It receives sub-data acquisition instructions from the upper layer and automatically performs shielded access to various heterogeneous data sources internally. It reorganizes the scattered raw bytes in different physical locations and formats according to a unified standard into consistent raw flaw detection signals that can be directly used by the upper-level processing stage, thus providing a clean and unified data foundation for subsequent feature analysis.

[0039] Step S15: Based on the task intent, dynamically call the flaw detection physical operator and operator parameters to extract features from the original flaw detection signal and generate a damage feature view.

[0040] It should be noted that step S15 enables flexible and on-demand configuration of the flaw detection processing. Traditional flaw detection equipment executes a fixed feature extraction process, including filtering, gain, and time-frequency analysis, regardless of the detection intent. However, in this step, the system dynamically adjusts the processing logic based on the task intent: for a "rapid general survey" intent, it may only call the computationally inefficient amplitude statistics operator to obtain attenuation and amplitude characteristics; for a "refined diagnosis" intent, it will call a high-resolution time-frequency transformation operator or spatial imaging operator. By flexibly calling physical operators and parameters, the raw flaw detection signal obtained in the aforementioned steps can be converted into a multi-dimensional, multi-view damage feature view that highly matches the current task objective, avoiding information loss or wasted computing power caused by fixed algorithms.

[0041] Step S16: Generate a flaw detection report based on the damage feature view.

[0042] It should be noted that step S16 aggregates and expresses the multi-dimensional feature view formed after intelligent processing. Unlike traditional instruments that only provide a single peak damage assessment curve or defect equivalent value, this step can generate a structured flaw detection report based on the damage feature view that integrates the results of multiple pieces of evidence. This report contains rich information such as the background of the detected object, the task intent, the evidence composition, the multi-dimensional features of the damage, and its spatial distribution. This report not only gives the final conclusion of whether the damage exists, but also clearly presents the complete chain of evidence supporting the conclusion, improving the interpretability and traceability of the detection results. It facilitates historical comparison and verification, and allows for future expansion to build a more intelligent damage assessment model.

[0043] In the above implementation, a flaw detection task is received, wherein the detection task includes a task object and a task intent. Then, based on the flaw detection semantic reasoning model, semantic inference is performed according to the flaw detection task to generate corresponding flaw detection semantic information. The flaw detection semantic reasoning model is used to characterize the semantic association pattern between dynamically expandable physical objects, detection methods and damage states, and semantic reasoning rules corresponding to different task intents. The flaw detection semantic information is used to characterize the evidence logic structure required to complete the task intent, as well as the semantic roles and collaboration patterns of each piece of evidence. Then, based on a dynamic semantic addressing mapping library, semantic matching is performed on the flaw detection semantic information to obtain sub-data acquisition instructions. The dynamic semantic addressing mapping library is used to characterize the many-to-many dynamic conversion rules between semantic roles and physical storage locations. Then, through a virtualized fact generation layer, the original flaw detection signal is acquired according to the sub-data acquisition instructions. The virtualized fact generation layer is used to characterize the virtualized organization logic for multi-source heterogeneous flaw detection data. Then, based on the task intent, the flaw detection physical operator and operator parameters are dynamically called to extract features from the original flaw detection signal and generate a damage feature view. Finally, based on the damage feature view, a flaw detection report is generated. This invention transforms the traditional fixed detection logic into an intent-driven dynamic detection mechanism. That is, it automatically organizes detection evidence according to the detection task intent, dynamically acquires the original flaw detection signals in heterogeneous storage, and flexibly combines multiple processing methods to generate a damage analysis view. This improves the flexibility of flaw detection, reduces reliance on human experience, enhances the collaborative efficiency of multi-source evidence, and thus improves the level of automation, intelligence, and adaptability to changing tasks in complex flaw detection scenarios.

[0044] As a further implementation of the method, the step of generating corresponding flaw detection semantic information based on the flaw detection semantic reasoning model and semantic derivation according to the flaw detection task includes: Step S21: Load the flaw detection task into the flaw detection semantic reasoning model. The flaw detection semantic reasoning model includes dynamically expandable physical object nodes, detection method nodes, and damage state nodes, as well as dynamically expandable semantic association patterns and semantic reasoning rules.

[0045] Step S22: Perform semantic parsing on the flaw detection task, and activate the target physical object node corresponding to the task object, the target detection method node associated with the task intent, and the target damage status node associated with the task intent.

[0046] Step S23: Based on semantic association patterns, extract all reachable paths from the target physical object node to the target damage state node via the target detection method node, as candidate evidence chains.

[0047] Step S24: Based on semantic reasoning rules, evaluate the confidence of each candidate evidence chain and assign chain weights to each candidate evidence chain according to the evaluation results; and select candidate evidence chains whose chain weights meet preset conditions to form a weighted evidence chain set.

[0048] Step S25: Assign semantic roles to nodes on each candidate evidence chain in the weighted evidence chain set, and determine the weighted collaboration mode between each semantic role based on the connection relationship between nodes and the chain weight. The weighted collaboration mode is a multi-evidence chain collaboration mechanism based on chain weight, which is used to constrain the collaboration relationship between semantic roles of different candidate evidence chains.

[0049] Step S26: Generate flaw detection semantic information based on the topological structure of each candidate evidence chain in the weighted evidence chain set, as well as the corresponding semantic roles and weighted collaboration modes.

[0050] It should be noted that the core idea of ​​this implementation method is to explicitly construct the reasoning process from task intent to evidence organization in flaw detection as a dynamically expandable semantic network. In traditional flaw detection operations, when faced with a damage diagnosis task for a specific component, the organization of evidence relies entirely on experts searching relevant detection knowledge in their minds, matching applicable methods, and planning the synergistic relationships between various signals. This thought process is implicit and rigid, making it difficult for the system to automatically reproduce and flexibly adjust. This implementation abstracts physical objects, detection methods, and damage states into three types of nodes in a semantic network, and models their inherent relationships as scalable semantic association patterns, enabling explicit representation and computation of expert thought processes. When a detection task is input, the system uses the task object and task intent as activation starting points, automatically searches the network for all reachable paths from the target physical object to the target damage state via applicable detection methods, filters and combines the optimal set of evidence chains through confidence assessment and weighting mechanisms, and finally determines the semantic roles and collaborative modes based on the topological structure of each evidence chain. This transforms the originally vague and empirical detection strategy into structured and executable flaw detection semantic information, realizing the shift from "manual experience planning" to "semantic reasoning-driven" approaches.

[0051] In the above implementation, to achieve semantic inference, the flaw detection task is loaded into the flaw detection semantic reasoning model. This model includes dynamically expandable physical object nodes, detection method nodes, and damage state nodes, as well as dynamically expandable semantic association patterns and semantic reasoning rules. Then, the flaw detection task is semantically parsed, activating the target physical object node corresponding to the task object, the target detection method node associated with the task intent, and the target damage state node associated with the task intent. Based on the semantic association pattern, all reachable paths from the target physical object node through the target detection method node to the target damage state node are extracted as candidate evidence chains. Finally, based on semantic inference... The system follows a set of rules to assess the confidence level of each candidate evidence chain and assign chain weights to each candidate evidence chain based on the assessment results. Candidate evidence chains whose chain weights meet preset conditions are selected to form a weighted evidence chain set. Semantic roles are then assigned to nodes on each candidate evidence chain in the weighted evidence chain set. Based on the connection relationships between nodes and the chain weights, a weighted collaboration mode is determined between each semantic role. This weighted collaboration mode is a multi-evidence chain collaboration mechanism based on chain weights, used to constrain the collaboration relationships between the semantic roles of different candidate evidence chains. Finally, based on the topology of each candidate evidence chain in the weighted evidence chain set, as well as the corresponding semantic roles and weighted collaboration modes, flaw detection semantic information is generated.

[0052] As a further implementation of the method, the steps of evaluating the confidence of each candidate evidence chain based on semantic reasoning rules and assigning chain weights to each candidate evidence chain according to the evaluation results include: Step S31: Based on the flaw detection semantic reasoning model, obtain the reasoning success rate of each candidate evidence chain, wherein the reasoning success rate is used to characterize the frequency with which the reasoning conclusion of the candidate evidence chain in the historical flaw detection task is consistent with the subsequent confirmation result.

[0053] Step S32: Based on the reasoning success rate, determine the chain weight of each candidate evidence chain, wherein the reasoning success rate is positively correlated with the chain weight.

[0054] Step S33: After the completion of this flaw detection task, compare the reasoning conclusions of each candidate evidence chain in the weighted evidence chain set with the corresponding post-confirmation results, and update the reasoning success rate of each candidate evidence chain in the flaw detection semantic reasoning model according to the correctness of the comparison.

[0055] It should be noted that this implementation provides a dynamic evaluation and self-evolution mechanism for evidence chain weights based on empirical feedback. In flaw detection practice, the actual diagnostic reliability of different evidence chains varies due to factors such as the detection object, working conditions, and damage evolution patterns. Current technologies typically employ fixed judgment rules or preset priorities, which cannot adaptively correct themselves with the accumulation of actual data. Therefore, this implementation uses the reasoning success rate as a quantitative indicator to characterize the frequency with which the reasoning conclusions of each candidate evidence chain are consistent with the subsequent confirmation results in historical tasks, and determines the chain weights accordingly, so that the confidence assessment is based on objective statistical verification rather than subjective experience presets. At the same time, after each detection task, the reasoning conclusions of this time are compared with the subsequent confirmation results, and the reasoning success rate is updated in reverse according to the correctness of the comparison, forming a closed-loop learning circuit of "evaluation-application-verification-update", so that the weights of each evidence chain are automatically adjusted with the accumulation of data, driving the flaw detection semantic reasoning model to continuously approach the best actual detection practice.

[0056] In the above implementation, in order to generate chain weights, the reasoning success rate of each candidate evidence chain is obtained based on the flaw detection semantic reasoning model. The reasoning success rate is used to characterize the frequency with which the reasoning conclusions of the candidate evidence chain in historical flaw detection tasks are consistent with the post-confirmation results. Then, based on the reasoning success rate, the chain weight of each candidate evidence chain is determined. The reasoning success rate is positively correlated with the chain weight. After the current flaw detection task is completed, the reasoning conclusions of each candidate evidence chain in the current weighted evidence chain set are compared with the corresponding post-confirmation results, and the reasoning success rate of each candidate evidence chain in the flaw detection semantic reasoning model is updated according to the correctness of the comparison.

[0057] As a further implementation of the method, the step of semantically matching the flaw detection semantic information based on the dynamic semantic addressing mapping library to obtain sub-data acquisition instructions includes: Step S41: Extract semantic roles from the flaw detection semantic information to obtain target semantic roles, where the target semantic roles include the inspected object, the detection method, and the target defect.

[0058] Step S42: Based on the dynamic semantic addressing mapping library, semantic addressing is performed according to the target semantic roles to obtain structured addressing results. The structured addressing results are generated by the many-to-many dynamic conversion rules maintained in the dynamic semantic addressing mapping library. The structured addressing results are grouped according to physical storage location. Each group corresponds to a physical storage location, and each group contains all target semantic roles mapped to that physical storage location and the signal segment extraction range or metadata filtering conditions corresponding to each target semantic role.

[0059] Step S43: Based on the structured addressing result, generate a sub-data retrieval instruction. The sub-data retrieval instruction contains only one access request for each group in the structured addressing result, targeting the physical storage location corresponding to that group.

[0060] It should be noted that this embodiment provides an efficient, merged addressing and instruction generation mechanism from semantic roles to physical storage. In current technologies, even after analysts have identified the required signal types, they still need to determine the specific file directories or data tables where each signal is located and initiate multiple independent access requests, resulting in a cumbersome and inefficient data acquisition process, and the access path is highly coupled with semantic requirements. To solve this problem, this embodiment first extracts target semantic roles such as the inspected object, detection method, and target defect from the flaw detection semantic information. Then, it performs semantic addressing through many-to-many dynamic conversion rules maintained in the dynamic semantic addressing mapping library to obtain structured addressing results grouped by physical storage location. This merges data requirements scattered across different storage locations according to the access target, avoiding repeated access to the same storage location. Subsequently, based on the structured addressing results, sub-data acquisition instructions are generated, ensuring that each physical storage location corresponds to only one access request. This maximizes the reduction of data acquisition interactions while ensuring complete coverage of semantic requirements, achieving efficient and accurate acquisition from semantic requirements to physical data.

[0061] In the above implementation, in order to achieve semantic matching, semantic roles are extracted from the flaw detection semantic information to obtain target semantic roles. The target semantic roles include the inspected object, the detection method, and the target defect. Then, based on the dynamic semantic addressing mapping library, semantic addressing is performed according to the target semantic roles to obtain structured addressing results. The structured addressing results are generated by many-to-many dynamic transformation rules maintained in the dynamic semantic addressing mapping library. The structured addressing results are grouped according to physical storage locations, with each group corresponding to a physical storage location. Each group contains all target semantic roles mapped to that physical storage location and the signal segment extraction range or metadata filtering conditions corresponding to each target semantic role. Then, based on the structured addressing results, sub-data acquisition instructions are generated. The sub-data acquisition instructions contain only one access request targeting the physical storage location corresponding to each group in the structured addressing results.

[0062] As a further implementation of the method, the step of acquiring the original flaw detection signal through a virtualized fact generation layer and according to the sub-data acquisition instruction includes: Step S51: In the virtualization fact generation layer, the sub-data acquisition instructions are parsed to obtain the target physical storage location corresponding to each access request, as well as the target semantic role associated with each target physical storage location and the corresponding extraction conditions. The extraction conditions include the signal segment extraction range or metadata filtering conditions.

[0063] In step S52, the virtualization fact generation layer calls the data access interface that matches the data source type to initiate a data read request to the corresponding target physical storage location based on the data source type corresponding to each target physical storage location.

[0064] In step S53, the virtualization fact generation layer receives the raw data returned by each target physical storage location and extracts the raw data according to the extraction conditions associated with each target physical storage location to obtain the raw flaw detection signal. The raw flaw detection signal is uniformly managed by the virtualization organization logic of the virtualization fact generation layer to shield the underlying storage differences of multi-source heterogeneous flaw detection data.

[0065] It should be noted that this implementation method achieves a unified virtualized access and on-demand signal extraction mechanism for multi-source heterogeneous flaw detection data. Currently, data generated by different flaw detection devices differ in storage format, interface protocol, and parameter definition. Upper-layer applications need to interface with various dedicated interfaces and handle the underlying differences in raw data, resulting in high development and maintenance costs and limited scalability. To solve this problem, this implementation method parses sub-data acquisition instructions through a virtualized fact generation layer, clarifying the physical storage location targeted by each access request and its associated semantic role and extraction conditions. Subsequently, based on the data source type corresponding to each physical storage location, it automatically calls the matching data access interface to initiate a read request, achieving unified access to different modal data sources such as ultrasonic, eddy current, and visual data. After receiving the returned raw data, it precisely extracts it according to the extraction conditions to obtain the raw flaw detection signal that accurately corresponds to the semantic requirements, and presents it uniformly to the upper layer through virtualized organization logic. Thus, the upper-layer processing only needs to operate on a unified virtualized view, without needing to be aware of the physical distribution and format differences of the underlying data, achieving efficient integration and transparent access to multi-source heterogeneous flaw detection data.

[0066] In the above implementation, in order to obtain the original flaw detection signal, the sub-data acquisition instructions are parsed in the virtualization fact generation layer to obtain the target physical storage location corresponding to each access request, as well as the target semantic role associated with each target physical storage location and the corresponding extraction conditions. The extraction conditions include the signal segment extraction range or metadata filtering conditions. Then, the virtualization fact generation layer calls the data access interface that matches the data source type corresponding to each target physical storage location to initiate a data read request to the corresponding target physical storage location. Then, the virtualization fact generation layer receives the original data returned by each target physical storage location and truncates the original data according to the extraction conditions associated with each target physical storage location to obtain the original flaw detection signal. The original flaw detection signal is uniformly managed by the virtualization organization logic of the virtualization fact generation layer to shield the underlying storage differences of multi-source heterogeneous flaw detection data.

[0067] As a further implementation of the method, the step of dynamically invoking the flaw detection physical operator and its parameters based on the task intent to extract features from the original flaw detection signal and generate a damage feature view includes: Step S61: Based on the task intent, a search is performed in the operator orchestration knowledge base to obtain multiple candidate operator templates that match the task intent. The candidate operator template is a directed acyclic graph structure. The candidate operator template is used to define the signal processing technology. The candidate operator template includes the flaw detection physical operator and the predetermined connection relationship and data flow between the flaw detection physical operators. Moreover, each flaw detection physical operator in the same candidate operator template is for the same detection signal type.

[0068] Step S62: Based on the collaborative mode of each semantic role, the candidate operator templates are arranged to generate the operator execution pipeline, and initial parameters are configured for each flaw detection physical operator in the operator execution pipeline. The collaborative mode is used to guide the arrangement of multiple candidate operator templates. Each flaw detection physical operator in the operator execution pipeline has a determined arrangement hierarchy and data flow path.

[0069] Step S63: Based on the operator execution pipeline, feature extraction is performed on the original flaw detection signal, and the current output feature quality index of each flaw detection physical operator is monitored in real time during the execution process. The output feature quality index includes at least one of signal-to-noise ratio, waveform smoothness, spectral peak clarity, energy concentration and feature class separability.

[0070] Step S64: When the output feature quality index of any flaw detection physical operator is lower than the preset threshold, the parameter adaptive adjustment mechanism is triggered. Based on the deviation between the current output feature quality index and the expected quality index, the parameters of the downstream flaw detection physical operator are corrected until the operator execution pipeline is completed and a damage feature view is generated.

[0071] It should be noted that this implementation method realizes a task intent-driven, quality-closed-loop adaptive flaw detection signal processing and feature view generation mechanism. Its core idea is to construct the flaw detection signal processing flow as a programmable and controllable operator execution pipeline, enabling the processing strategy to be flexibly adjusted according to the task intent. Simultaneously, quality monitoring and adaptive parameter correction are introduced during execution to ensure the reliability and specificity of the output results. Specifically, firstly, matching candidate operator templates are retrieved from the operator orchestration knowledge base according to the task intent. Each template encapsulates a standardized processing technology for a specific signal type using a directed acyclic graph structure. Subsequently, multiple templates are orchestrated according to the collaborative patterns between semantic roles, forming an operator execution pipeline with clear hierarchical relationships and data flow paths. Initial parameters are configured for each operator, thereby enabling the collaborative integration of the signal extraction logic obtained from different detection methods. During execution, the output feature quality indicators of each operator are monitored in real time. When the quality of a certain link does not meet expectations, a parameter adjustment mechanism is automatically triggered, correcting the downstream operator parameters based on deviation information, forming a closed-loop control of "monitoring-feedback-correction". This mechanism enables the signal processing process to have both flexible orchestration capabilities for task-oriented tasks and autonomous optimization capabilities for signal quality, ultimately generating a damage feature view that highly matches the detection intent.

[0072] In the above implementation, to generate a damage feature view, a search is performed in the operator orchestration knowledge base according to the task intent to obtain multiple candidate operator templates matching the task intent. Each candidate operator template is a directed acyclic graph structure, used to define the signal processing technology. The candidate operator template includes flaw detection physical operators and predetermined connection relationships and data flow directions between them. Each flaw detection physical operator within the same candidate operator template targets the same detection signal type. Then, based on the collaborative mode of each semantic role, the candidate operator templates are orchestrated to generate an operator execution pipeline, and initial parameters are configured for each flaw detection physical operator in the pipeline. The collaborative mode guides the orchestration method among the multiple candidate operator templates. The various physical operators in the operator execution pipeline have a predetermined arrangement hierarchy and data flow path. Then, based on the operator execution pipeline, features are extracted from the original physical signal. During the execution process, the current output feature quality index of each physical operator is monitored in real time. The output feature quality index includes at least one of the following: signal-to-noise ratio, waveform smoothness, peak spectral clarity, energy concentration, and feature class separability. When the output feature quality index of any physical operator is lower than a preset threshold, a parameter adaptive adjustment mechanism is triggered. Based on the deviation between the current output feature quality index and the expected quality index, the parameters of the downstream physical operators are corrected until the operator execution pipeline is completed and a damage feature view is generated.

[0073] As a further implementation of the method, the step of orchestrating candidate operator templates based on the collaborative patterns of various semantic roles to generate an operator execution pipeline includes: Step S71: Analyze the collaborative mode and determine the corresponding orchestration logic. The orchestration logic is used to characterize the orchestration rules between different signal processing processes. The orchestration logic defines the connection method between each flaw detection physical operator in the operator execution pipeline, and the connection method includes parallel connection and fusion connection.

[0074] Step S72: Based on the orchestration logic, the candidate operator templates are orchestrated to generate the operator execution pipeline.

[0075] It should be noted that this implementation provides a structured orchestration mechanism for multi-source signal processing driven by collaborative patterns. Its core idea is to concretize the collaborative patterns among various pieces of evidence identified in the flaw detection semantic information into executable operator pipeline orchestration logic, thereby enabling the signal processing techniques corresponding to different detection methods to be organically combined according to the inherent collaborative relationships between the evidence. Specifically, firstly, the collaborative patterns are analyzed to extract the orchestration rules between different signal processing techniques. This orchestration logic clarifies the connection methods between various flaw detection physical operators in the operator execution pipeline. Parallel connections correspond to scenarios where multiple pieces of evidence need to be processed independently and then aggregated simultaneously, while fusion connections correspond to scenarios where different pieces of evidence need to be cross-fused at the feature level. Subsequently, based on this orchestration logic, multiple candidate operator templates obtained from the retrieval are orchestrated, integrating the originally independent signal processing techniques into an operator execution pipeline with a clear structure, hierarchical relationship, and data flow path. This ensures that the final generated feature view reflects the collaborative logic between multiple pieces of evidence, rather than simply a stacking of signal processing results.

[0076] In the above implementation, in order to generate the operator execution pipeline, the cooperative mode is parsed to determine the corresponding orchestration logic. The orchestration logic is used to characterize the orchestration rules between different signal processing processes. The orchestration logic defines the connection method between each flaw detection physical operator in the operator execution pipeline. The connection method includes parallel connection and fusion connection. Then, based on the orchestration logic, the candidate operator templates are orchestrated to generate the operator execution pipeline.

[0077] As a further implementation of the method, the step of correcting the parameters of the downstream flaw detection physical operator based on the deviation between the current output feature quality index and the expected quality index includes: Step S81: Determine the deviation value based on the current output feature quality index and the expected quality index.

[0078] Step S82: Based on the deviation value, determine the parameter correction strategy corresponding to the deviation value from the preset parameter correction rule base. The parameter correction strategy includes the target parameter to be corrected in the downstream flaw detection physical operator, as well as the correction amount and correction direction of the target parameter.

[0079] Step S83: Based on the parameter correction strategy, the downstream flaw detection physical operator is corrected.

[0080] In the above implementation, in order to correct the parameters of the downstream flaw detection physical operator, a deviation value is determined based on the current output feature quality index and the expected quality index. Then, based on the deviation value, a parameter correction strategy corresponding to the deviation value is determined from a preset parameter correction rule base. The parameter correction strategy includes the target parameter to be corrected in the downstream flaw detection physical operator, as well as the correction amount and correction direction of the target parameter. Then, the downstream flaw detection physical operator is corrected based on the parameter correction strategy.

[0081] This application also discloses an intelligent flaw detection system based on a data warehouse model.

[0082] refer to Figure 2 A smart flaw detection system based on a data warehouse model includes: The task receiving module is used to receive flaw detection tasks, where the detection task includes the task object and the task intent; The semantic reasoning module is used to generate corresponding flaw detection semantic information based on the flaw detection semantic reasoning model and the flaw detection task. The flaw detection semantic reasoning model is used to represent the semantic association pattern between dynamically expandable physical objects, detection methods and damage states, as well as the semantic reasoning rules corresponding to different task intentions. The flaw detection semantic information is used to represent the logical structure of evidence required to complete the task intention, as well as the semantic role and collaboration mode of each piece of evidence. The semantic addressing module is used to perform semantic matching on flaw detection semantic information based on the dynamic semantic addressing mapping library to obtain sub-data acquisition instructions. The dynamic semantic addressing mapping library is used to represent the many-to-many dynamic conversion rules between semantic roles and physical storage locations. The virtualized data acquisition module is used to acquire the original flaw detection signal through the virtualized fact generation layer and according to the sub-data acquisition instructions. The virtualized fact generation layer is used to characterize the virtualized organization logic for multi-source heterogeneous flaw detection data. The feature extraction module is used to dynamically call the physical operators and operator parameters of the flaw detection based on the task intent, extract features from the original flaw detection signal, and generate a damage feature view. The inspection report generation module is used to generate flaw detection reports based on the damage feature view.

[0083] The intelligent flaw detection system based on the data warehouse model of the present invention can implement any of the intelligent flaw detection methods based on the data warehouse model, and the specific working process of the intelligent flaw detection system based on the data warehouse model of the present invention can refer to the corresponding process in the above-mentioned intelligent flaw detection methods based on the data warehouse model.

[0084] This application also discloses a computer device.

[0085] refer to Figure 3 A computer device includes a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the computer program to implement any of the above-described intelligent flaw detection methods based on the data warehouse model.

[0086] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. A smart flaw detection method based on a data warehouse model, characterized in that, include: Receive a flaw detection task, wherein the detection task includes a task object and a task intent; Based on the flaw detection semantic reasoning model, and based on the flaw detection task, semantic inference is performed to generate corresponding flaw detection semantic information. The flaw detection semantic reasoning model is used to characterize the semantic association pattern between dynamically expandable physical objects, detection methods and damage states, as well as the semantic reasoning rules corresponding to different task intentions. The flaw detection semantic information is used to characterize the evidence logic structure required to complete the task intention, as well as the semantic roles and collaboration patterns of each piece of evidence. Based on the dynamic semantic addressing mapping library, semantic matching is performed on the flaw detection semantic information to obtain sub-data acquisition instructions. The dynamic semantic addressing mapping library is used to characterize the many-to-many dynamic conversion rules between the semantic roles and physical storage locations. The original flaw detection signal is obtained through the virtualized fact generation layer and according to the sub-data acquisition instruction. The virtualized fact generation layer is used to characterize the virtualized organization logic for multi-source heterogeneous flaw detection data. Based on the task intent, the physical operators for flaw detection and their parameters are dynamically invoked to extract features from the original flaw detection signal and generate a damage feature view. Based on the damage feature view, a flaw detection report is generated.

2. The intelligent flaw detection method based on a data warehouse model according to claim 1, characterized in that, The step of generating corresponding flaw detection semantic information based on the flaw detection semantic reasoning model and semantic inference according to the flaw detection task includes: The flaw detection task is loaded into the flaw detection semantic reasoning model, wherein the flaw detection semantic reasoning model includes dynamically expandable physical object nodes, detection method nodes and damage state nodes, as well as dynamically expandable semantic association patterns and semantic reasoning rules. Semantic parsing is performed on the flaw detection task to activate the target physical object node corresponding to the task object, the target detection method node associated with the task intent, and the target damage status node associated with the task intent. Based on the semantic association pattern, all reachable paths from the target physical object node to the target damage state node via the target detection method node are extracted as candidate evidence chains. Based on the semantic reasoning rules, confidence is evaluated for each candidate evidence chain, and chain weights are assigned to each candidate evidence chain according to the evaluation results; and candidate evidence chains whose chain weights satisfy preset conditions are selected to form a weighted evidence chain set. Semantic roles are assigned to nodes on each candidate evidence chain in the weighted evidence chain set, and a weighted collaboration mode between each semantic role is determined based on the connection relationship between nodes and the chain weight. The weighted collaboration mode is a multi-evidence chain collaboration mechanism based on the chain weight, which is used to constrain the collaboration relationship between the semantic roles of different candidate evidence chains. Based on the topological structure of each candidate evidence chain in the weighted evidence chain set, as well as the corresponding semantic roles and weighted collaboration modes, flaw detection semantic information is generated.

3. The intelligent flaw detection method based on a data warehouse model according to claim 2, characterized in that, The step of evaluating the confidence level of each candidate evidence chain based on the semantic reasoning rules, and assigning a chain weight to each candidate evidence chain according to the evaluation results, includes: Based on the flaw detection semantic reasoning model, the reasoning success rate of each candidate evidence chain is obtained, wherein the reasoning success rate is used to characterize the frequency with which the reasoning conclusion of the candidate evidence chain in historical flaw detection tasks is consistent with the subsequent confirmation result. Based on the reasoning success rate, the chain weight of each candidate evidence chain is determined, wherein the reasoning success rate is positively correlated with the chain weight; After the completion of the flaw detection task, the reasoning conclusions of each candidate evidence chain in the weighted evidence chain set are compared with the corresponding post-confirmation results, and the reasoning success rate of each candidate evidence chain in the flaw detection semantic reasoning model is updated according to the correctness of the comparison.

4. The intelligent flaw detection method based on a data warehouse model according to claim 1, characterized in that, The step of semantically matching the flaw detection semantic information based on the dynamic semantic addressing mapping library to obtain sub-data acquisition instructions includes: Semantic roles are extracted from the flaw detection semantic information to obtain target semantic roles, wherein the target semantic roles include the inspected object, the detection method, and the target defect; Based on the dynamic semantic addressing mapping library, and according to the target semantic roles, a structured addressing result is obtained. The structured addressing result is generated by the many-to-many dynamic conversion rules maintained in the dynamic semantic addressing mapping library. The structured addressing result is grouped according to the physical storage location. Each group corresponds to a physical storage location, and each group contains all the target semantic roles mapped to that physical storage location and the signal segment extraction range or metadata filtering conditions corresponding to each target semantic role. Based on the structured addressing result, a sub-data retrieval instruction is generated, wherein the sub-data retrieval instruction contains only one access request targeting the physical storage location corresponding to the group for each group in the structured addressing result.

5. The intelligent flaw detection method based on a data warehouse model according to claim 4, characterized in that, The step of acquiring the original flaw detection signal through the virtualized fact generation layer and according to the sub-data acquisition instruction includes: In the virtualization fact generation layer, the sub-data acquisition instructions are parsed to obtain the target physical storage location corresponding to each access request, as well as the target semantic role associated with each target physical storage location and the corresponding extraction conditions. The extraction conditions include the signal segment extraction range or the metadata filtering conditions. The virtualization fact generation layer, based on the data source type corresponding to each target physical storage location, calls the data access interface that matches the data source type to initiate a data read request to the corresponding target physical storage location; The virtualization fact generation layer receives the raw data returned from each of the target physical storage locations, and extracts the raw data according to the extraction conditions associated with each of the target physical storage locations to obtain the raw flaw detection signal. The raw flaw detection signal is uniformly managed by the virtualization organization logic of the virtualization fact generation layer to shield the underlying storage differences of multi-source heterogeneous flaw detection data.

6. The intelligent flaw detection method based on a data warehouse model according to claim 1, characterized in that, The step of dynamically invoking the flaw detection physical operator and its parameters based on the task intent to extract features from the original flaw detection signal and generate a damage feature view includes: Based on the task intent, a search is performed in the operator orchestration knowledge base to obtain multiple candidate operator templates that match the task intent. The candidate operator template is a directed acyclic graph structure. The candidate operator template is used to define the signal processing technology. The candidate operator template includes a flaw detection physical operator and a predetermined connection relationship and data flow between the flaw detection physical operators. Moreover, each flaw detection physical operator in the same candidate operator template is for the same detection signal type. Based on the collaborative mode of each semantic role, the candidate operator templates are arranged to generate an operator execution pipeline, and initial parameters are configured for each flaw detection physical operator in the operator execution pipeline. The collaborative mode is used to guide the arrangement of multiple candidate operator templates, and each flaw detection physical operator in the operator execution pipeline has an arrangement hierarchy and data flow path determined by the arrangement. Based on the operator execution pipeline, feature extraction is performed on the original flaw detection signal, and the current output feature quality index of each flaw detection physical operator is monitored in real time during the execution process. The output feature quality index includes at least one of signal-to-noise ratio, waveform smoothness, spectral peak clarity, energy concentration and feature class separability. When the output feature quality index of any flaw detection physical operator is lower than a preset threshold, a parameter adaptive adjustment mechanism is triggered. Based on the deviation between the current output feature quality index and the expected quality index, the parameters of the downstream flaw detection physical operator are corrected until the operator execution pipeline is completed and a damage feature view is generated.

7. The intelligent flaw detection method based on a data warehouse model according to claim 6, characterized in that, The step of orchestrating the candidate operator templates based on the collaborative patterns of various semantic roles to generate an operator execution pipeline includes: The collaborative mode is analyzed to determine the corresponding orchestration logic. The orchestration logic is used to characterize the orchestration rules between different signal processing techniques. The orchestration logic defines the connection method between each flaw detection physical operator in the operator execution pipeline. The connection method includes parallel connection and fusion connection. Based on the orchestration logic, the candidate operator templates are orchestrated to generate an operator execution pipeline.

8. The intelligent flaw detection method based on a data warehouse model according to claim 6, characterized in that, The step of correcting the parameters of the downstream flaw detection physical operator based on the deviation between the current output feature quality index and the expected quality index includes: Based on the current output feature quality index and the expected quality index, the deviation value is determined; Based on the deviation value, a parameter correction strategy corresponding to the deviation value is determined from a preset parameter correction rule base. The parameter correction strategy includes the target parameter to be corrected in the downstream flaw detection physical operator, as well as the correction amount and correction direction of the target parameter. Based on the aforementioned parameter correction strategy, the downstream flaw detection physical operator is corrected.

9. A smart flaw detection system based on a data warehouse model, characterized in that, include: The task receiving module is used to receive flaw detection tasks, wherein the detection task includes a task object and a task intent; The semantic reasoning module is used to generate corresponding flaw detection semantic information based on the flaw detection semantic reasoning model and the flaw detection task. The flaw detection semantic reasoning model is used to characterize the semantic association pattern between dynamically expandable physical objects, detection methods and damage states, as well as the semantic reasoning rules corresponding to different task intentions. The flaw detection semantic information is used to characterize the evidence logic structure required to complete the task intention, as well as the semantic role and collaboration mode of each piece of evidence. The semantic addressing module is used to perform semantic matching on the flaw detection semantic information based on the dynamic semantic addressing mapping library to obtain sub-data acquisition instructions. The dynamic semantic addressing mapping library is used to represent the many-to-many dynamic conversion rules between the semantic roles and physical storage locations. The virtualized data acquisition module is used to acquire the original flaw detection signal through the virtualized fact generation layer and according to the sub-data acquisition instruction, wherein the virtualized fact generation layer is used to characterize the virtualized organization logic for multi-source heterogeneous flaw detection data; The feature extraction module is used to dynamically call the flaw detection physical operator and operator parameters based on the task intent, extract features from the original flaw detection signal, and generate a damage feature view. The inspection report generation module is used to generate a flaw detection report based on the damage feature view.

10. A computer device, characterized in that, The method includes a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the computer program to implement the method of any one of claims 1 to 8.

Images