Ultrasonic probe maintenance service automatic response processing method and system

Abstract

The invention provides an ultrasonic probe maintenance service automatic response processing method and system, and belongs to the technical field of semantic analysis, and the method comprises the steps: firstly carrying out data acquisition operation, constructing an initial maintenance knowledge graph through the data acquisition operation, carrying out the dynamic enhancement of the initial maintenance knowledge graph, and obtaining a maintenance enhanced knowledge graph; then obtaining client context data, and obtaining semantic analysis data based on the client context data and the maintenance enhanced knowledge graph; executing a query strategy based on the semantic analysis data, obtaining final result data, and obtaining response material data based on the final result data; and finally, response service is carried out based on the response material data, and a prospective and integrated automatic response processing method is provided, so that maintenance response service can be better carried out on clients.

Description

Technical Field

[0001] This invention relates to the field of semantic analysis technology, and more specifically, to an automatic response processing method and system for ultrasonic probe repair services. Background Technology

[0002] Currently, customer inquiries in the equipment repair service field generally rely on general online customer service systems or pre-scripted chatbots. However, when dealing with specialized equipment such as ultrasonic probes, which are highly technically complex, have long maintenance cycles, and whose maintenance progress is strongly correlated with performance parameters, general systems often have limitations.

[0003] In the context of ultrasonic probe repair, the repair process is not simply a replacement of parts, but a holistic process. Conventional systems often roughly abstract this whole process into nodes such as start, process, and completion, ignoring the correlation of procedures and the quantitative data representation. Furthermore, there is a lag in the tracking of repairs, which makes it impossible to effectively anticipate customer questions. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide an automatic response processing method for ultrasonic probe repair services, the method comprising: Perform data acquisition operations and construct an initial maintenance knowledge graph through these operations. Dynamically enhance the initial maintenance knowledge graph to obtain an enhanced maintenance knowledge graph. Obtain customer context data, and then obtain semantic parsing data based on the customer context data and the enhanced maintenance knowledge graph; The query strategy is executed based on the semantic parsing data to obtain the final result data, and the response material data is obtained based on the final result data. Response service is provided based on response material data.

[0005] Preferably, the step of performing data acquisition operations, constructing an initial maintenance knowledge graph through data acquisition operations, dynamically enhancing the initial maintenance knowledge graph, and obtaining an enhanced maintenance knowledge graph includes: Acquire physical test data, electrical test data, maintenance business data, and business text data, while performing anomaly handling, alignment, and mapping operations to obtain maintenance baseline data and construct an initial maintenance knowledge graph based on the maintenance baseline data; Import temporal causal relationships into the initial maintenance knowledge graph and generate a maintenance dynamic knowledge graph and a maintenance dynamic subgraph. Construct a maintenance knowledge model based on the initial maintenance knowledge graph and the maintenance dynamic subgraph, and output maintenance node vector data through the maintenance knowledge model. Based on maintenance node vector data and maintenance dynamic subgraph, dynamic path data is obtained by using planning and inference strategies. Then, an enhanced maintenance knowledge graph is obtained by combining the dynamic path data with the maintenance dynamic knowledge graph.

[0006] Preferably, the step of obtaining dynamic path data based on maintenance node vector data and maintenance dynamic subgraph using a planning and reasoning strategy, and obtaining an enhanced maintenance knowledge graph through the dynamic path data and maintenance dynamic knowledge graph, includes: Obtain feasible path data based on maintenance dynamic subgraph; Based on feasible path data, maintenance node vectors and maintenance dynamic subgraphs, and by obtaining optimized feasible path data through maintenance agents, the feasible path data is simulated and deduced to generate path deduction data. Optimized feasible path data and path deduction data are imported into the maintenance dynamic knowledge graph to generate an enhanced maintenance knowledge graph.

[0007] Preferably, the step of obtaining customer context data and obtaining semantic parsing data based on customer context data and maintenance enhanced knowledge graph includes: Acquire service input data, perform standardized operations on the service input data, build service customer profiles, and obtain customer context data; Based on the maintenance-enhanced knowledge graph and customer context data, link entity data, intent parsing data, and time logic data are obtained respectively, and integrated into semantic parsing data.

[0008] Preferably, the step of acquiring linked entity data, intent parsing data, and time logic data based on the maintenance-enhanced knowledge graph and customer context data respectively includes: Entity recognition is performed on customer context data to obtain context entity data. Entity matching is performed on the context entity data based on the maintenance enhanced knowledge graph, and linked entity data is obtained. Perform intent analysis on customer context data to obtain intent parsing data; Perform time-series and logical analysis on customer context data to obtain time-logic data.

[0009] Preferably, the step of executing a query strategy based on semantic parsing data to obtain final result data, and simultaneously obtaining response material data based on the final result data, includes: Obtain graph query instructions by semantically parsing the data, and obtain the final result data based on the graph query instructions and the maintenance-enhanced knowledge graph. Quantitative data of the results are obtained based on the final result data, and response material data is obtained by combining the quantitative data of the results with the final result data.

[0010] Preferably, the step of obtaining the graph query instruction through semantic parsing data, and obtaining the final result data based on the graph query instruction and the maintenance enhanced knowledge graph, includes: Perform complexity analysis and access planning on the graph query command to obtain the query execution plan; Adaptive queries are performed on the enhanced maintenance knowledge graph based on the query execution plan to obtain the result data; Enhanced result data is obtained by performing rule-based reasoning using the result data and the maintenance-enhanced knowledge graph. The enhanced result data is sorted by relevance to obtain the final result data.

[0011] Preferably, the step of obtaining result quantification data based on the final result data, and obtaining response material data through the result quantification data and the final result data, includes: Quantitative calculations and deviation analysis are performed on the final result data to obtain quantitative data. Summary data is obtained by semantically parsing the data and quantifying the results. Interaction data is obtained by semantically parsing the data and quantifying the results. The quantitative data, summary data, semantic parsing data, and interaction data of the results are integrated into response material data.

[0012] Furthermore, embodiments of the present invention also provide an automatic response and processing system for ultrasonic probe repair services, comprising: At least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the methods described above.

[0013] This embodiment first performs data acquisition operations and constructs an initial maintenance knowledge graph through these operations. The initial maintenance knowledge graph is then dynamically enhanced to obtain an enhanced maintenance knowledge graph. Next, customer context data is acquired, and semantic parsing data is obtained based on the customer context data and the enhanced maintenance knowledge graph. Then, a query strategy is executed based on the semantic parsing data to obtain final result data, and response material data is also obtained based on the final result data. Finally, response services are provided based on the response material data. By establishing a knowledge graph, various relevant data throughout the maintenance process can be linked, and causal inferences can be made. In subsequent response material package construction, key data is quantified, and the knowledge graph is enhanced through planning and inference strategies. This provides a foundation for proactive responses, offering a proactive and integrated automatic response processing method to better provide maintenance response services to customers and solve the problems of segmentation and lag in general response systems. Attached Figure Description

[0014] Figure 1 This is a flowchart of the steps of an automatic response processing method for ultrasonic probe repair service according to the present invention; Figure 2 This is a schematic diagram of an automatic response processing system for ultrasonic probe repair services according to the present invention. Detailed Implementation

[0015] It should be noted that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0016] It should be noted that, in the embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0017] The present invention will now be described in detail with reference to the accompanying drawings, as shown in the attached drawings. Figure 1 As shown, a detailed introduction is given to the automatic response processing method and system for ultrasonic probe repair services.

[0018] In this embodiment, an automatic response processing method for ultrasonic probe repair services includes steps A, B, C, and D.

[0019] Step A involves performing data acquisition operations and constructing an initial maintenance knowledge graph. This initial maintenance knowledge graph is then dynamically enhanced to obtain an enhanced maintenance knowledge graph.

[0020] Furthermore, step A includes steps A-1, A-2, and A-3.

[0021] Step A-1: ​​Obtain physical test data, electrical test data, maintenance business data, and business text data. Simultaneously, perform anomaly handling, alignment, and mapping operations to obtain maintenance baseline data. Construct an initial maintenance knowledge graph based on the maintenance baseline data.

[0022] Understandably, after each probe is sent for repair, it is inspected by an engineer. During the inspection, corresponding test data for the probe is generated, including but not limited to physical test data, electrical test data, maintenance business data, and business text data. Physical test data refers to data of physical entities, such as high-precision 3D scan point clouds of the probe and microscopic images of external damage. Electrical test data refers to electroacoustic data, such as array element impedance matrix, pulse echo waveform, and crosstalk matrix. Maintenance business data refers to business and historical data, such as work order information, real-time maintenance records, maintenance history records, and similar model failure cases. Business text data refers to unstructured text data, such as fault descriptions filled in by customers, engineer remarks, and maintenance manuals. After the data is acquired, data preprocessing operations can be performed on the above data, including but not limited to anomaly handling, alignment, and mapping operations.

[0023] The anomaly handling process involves identifying and discarding identical data in physical and electrical test data caused by network jitter, etc. For periodic data, if a certain time is missing, it can be filled using methods such as forward padding or linear interpolation. If there is critical mandatory data, such as missing corresponding probes, manual processing and filling are required. Then, the digital format is unified, and the date and time format is also unified. Then, preset rules are applied for anomaly detection. The preset rules may include physical filtering rules, outlier filtering rules, and correction rules. For physical filtering rules, conventional rules are applied. For example, the impedance matrix of an array element is generally tens of ohms. If there is 0.001 ohms, it is considered an anomaly and is removed. For outlier filtering rules, they are applied to continuous data. For example, a sliding window is used to calculate the standard deviation, and extreme values ​​exceeding 2-3 times the standard deviation are temporarily considered as suspected values. For correction rules, they are applied to suspected values. Suspected values ​​are compared with historical measurements of the same probe and the same location to determine whether to retain or remove them. During implementation, technicians can make adaptive changes to the preset rules based on the actual situation to adapt to the actual situation.

[0024] The alignment process involves the following steps: After the anomaly handling is completed, the physical and electrical test data, which have been preliminarily cleaned and formatted uniformly, are output. At this point, all data can be timestamped. For physical and electrical test data, time alignment methods such as NTP and PTP can be used. For file data without reliable timestamps, such as maintenance business data and business text data, their internal logical order, such as page numbers, test serial numbers, and limited reference time points, such as report dates, can be used to approximate alignment on the time axis. After timestamp alignment is completed, a three-dimensional coordinate system is established based on the reference positioning surface of the uniform fixture used by the probe during testing as the origin. The transformation matrix from the probe's own coordinate system to the global coordinate system is obtained, and all test points are transformed using this matrix. Then, their xyz coordinates in the global coordinate system are calculated, and a spatial index is established to obtain spatiotemporally aligned data, facilitating subsequent regional queries and analysis.

[0025] The mapping process involves extracting entities from structured data such as maintenance business data, physical test data, and electrical test data after the alignment process is complete. For example, the work order serial number can be extracted from the work order information in the maintenance business data, or array element 37 can be extracted from the physical test data and electrical test data after alignment and anomaly handling. At the same time, an alias mapping table is needed to map aliases to standard entities, such as mapping SN and S / N to serial numbers. For unstructured data such as business text data, tools such as NER are used to process it, thereby identifying and extracting entities from the text description. In some texts, there may be expressions such as "use," "the former," and "this part" for reference. These need to be linked to the specific entities identified in the previous text. Each identified entity is assigned a UID, and a mapping table for UIDs is established. Then, syntactic analysis in natural language processing is used to extract preliminary attribute relationships from the text, which can also be called triplet relationships, such as array element A7 - impedance value - too high, fault - possible cause - poor solder joint, etc.

[0026] After the exception handling, alignment, and mapping operations, data encapsulation can be performed to obtain maintenance baseline data. The maintenance baseline data includes a metadata layer, an entity layer, a data layer, and a semantic layer. The metadata layer is used to store timestamps, UIDs, spatial coordinates, etc. The entity layer is used to store the extracted entities. The data layer is used to store all the data detected. The semantic layer is used to store attribute relationships.

[0027] Understandably, the entity layer is retrieved from the maintenance baseline data and decomposed from top to bottom. At the same time, the entities appearing in the entity layer are analyzed to identify commonalities and abstract them into classes. For example, the object category includes ultrasonic probes, probe components, and testing equipment. Ultrasonic probes include linear array probes, convex array probes, and phased array probes. Probe components include acoustic lenses, piezoelectric elements, etc., and testing equipment includes impedance analyzers, etc., thereby establishing category relationships.

[0028] After establishing the category relationships, define category attributes and object attributes. Category attributes describe the characteristics of the entity itself. For example, the array element class includes array element A1, array element B1, etc., and all of them have attributes, with the unit being ohms and having numerical values. Object attributes describe the relationships between entities. For example, a probe component includes an acoustic lens. Then, you can define the spectrum rules, such as each probe component must belong to at least one ultrasonic probe, a linear array probe is a subclass of an ultrasonic probe, and if the fault mode of an array element is open circuit and the impedance is greater than 10kΩ, then the repair action is re-welding, etc. The above are just examples to illustrate the spectrum rules for ease of understanding. The spectrum rules need to be adjusted for adaptability based on the actual situation.

[0029] Once defined, mapping rules are constructed. For structured data, mapping can be performed directly. For unstructured data, the triple relationship mentioned above is used to match it with the defined classes and attributes. For example, array element A7 - impedance value - too high will be mapped to finding the entity of array element A7 and linking it to an entity that is too high. A knowledge extraction mapping engine, such as the Apache Jena tool, is run through the maintenance baseline data to construct an initial maintenance knowledge graph.

[0030] After the initial maintenance knowledge graph is constructed, there may be conflicts. For example, data source A is array element A7 with an impedance of 45Ω, while data source B is array element B with an impedance of 50Ω. In this case, the conflicts need to be eliminated. These conflicts can be eliminated by means of timeliness, weighted average comparison of initial credibility, manual processing, etc.

[0031] Step A-2: Import temporal causal relationships into the initial maintenance knowledge graph and generate a maintenance dynamic knowledge graph and a maintenance dynamic subgraph. Construct a maintenance knowledge model based on the initial maintenance knowledge graph and the maintenance dynamic subgraph, and output maintenance node vector data through the maintenance knowledge model.

[0032] The types of temporal relationships are defined to expand a set of relationships describing temporal sequence, such as connection relationships, ordering relationships, overlapping relationships, and duration relationships. Connection relationships can be understood as process A must be completed before process B begins. Ordering relationships can be understood as the end time of process A equals the start time of process B. Overlapping relationships can be understood as the execution times of processes A and B partially overlap and are allowed to run concurrently. Duration relationships can be understood as the standard time required to link processes to each other. For connection relationships, they can be automatically derived from the input and output of the process; for example, if the output of process X is the input of process Y, then X-connect-Y. For ordering and overlapping relationships, process mining algorithms, such as inductive mining analysis of historical work orders, can be used to automatically discover sequential patterns and concurrent patterns, and these patterns can be used as ordering and overlapping relationships. Duration relationships can be obtained from maintenance history records.

[0033] The types of causal relationships are defined, and causal relationships such as cause and constraint are introduced. Causal relationships can be obtained based on fault tree analysis and the experience of maintenance technicians, and then transformed into causal relationships. For example, cleanliness less than N leads to welding failure. Then, correlation analysis is performed on historical data to calculate the conditional probability or lift of co-occurrence of two time periods, and these are used as the strength weight of the causal relationship. For example, the correlation between historical cleanliness failure and subsequent welding failure can be analyzed. Then, a maintenance dynamic subgraph containing the current process, related processes, time constraints and causal chains can be extracted around the specific fault point of the probe currently being repaired. The maintenance dynamic subgraph is the direct operation object for subsequent planning.

[0034] Understandably, the initial maintenance knowledge graph is static, only able to answer questions of what is and what it is, but not about what to do first and what to do next, or why to do it this way. By importing time sequence and causal relationships, a dynamic maintenance knowledge graph can be obtained, which can endow the graph with process logic and diagnostic reasoning capabilities, making it understandable and able to simulate the evolution of the maintenance process.

[0035] Based on the initial maintenance knowledge graph with imported causal and temporal relationships, node features and edge features are constructed. The category and attributes of each entity node are encoded as feature vectors, i.e., node features. For example, the feature vector of an entity node may include component type, initial impedance value, region, and whether it is a fault point. Different feature codes are assigned to different types of relationships, and the strength weight of the relationship can be used as edge features. At this point, the node features and edge features are completed. Then, a heterogeneous graph neural network architecture, such as the RGCN architecture, can be selected to build the maintenance knowledge model to effectively handle multiple types of relationships. The heterogeneous graph neural network architecture can perform link prediction and node classification to understand the potential associations between entities and predict the category or attribute of nodes. For training the maintenance knowledge model, historical successful maintenance case graphs are used as positive samples, and some perturbation graphs are introduced as negative samples, such as disconnecting certain necessary relationships. The model is trained using positive and negative samples to obtain a trained maintenance knowledge model. The maintenance dynamic subgraph of the current probe is input into the maintenance knowledge model, and the maintenance knowledge model outputs a low-dimensional vector representation of each node in the graph, i.e., maintenance node vector data.

[0036] Understandably, by capturing complex structural and relational information in the knowledge graph through a maintenance knowledge model and encoding it into vectors that are easy to compute, entities with similar maintenance logic, fault characteristics, or solutions in the vector space will have vectors that are closer together. By calculating the vector distance, reference solutions similar to the current situation can be quickly retrieved from the historical case library. For example, multiple combinations of procedures that have successfully repaired open circuits in array elements in the past will have their corresponding subgraph vectors clustered in the space, thus providing a basis for decision-making for subsequent planning.

[0037] Step A-3: Based on the maintenance node vector data and the maintenance dynamic subgraph, dynamic path data is obtained using a planning and inference strategy. Then, an enhanced maintenance knowledge graph is obtained by combining the dynamic path data with the maintenance dynamic knowledge graph.

[0038] Furthermore, step A-3 includes steps A3-1-1, A3-1-2, and A3-1-3.

[0039] Step A3-1-1: Obtain feasible path data based on the maintenance dynamic subgraph.

[0040] Step A3-1-2: Based on feasible path data, maintenance node vectors and maintenance dynamic subgraphs, optimize feasible path data through maintenance agents, simulate and extrapolate feasible path data to generate path extrapolation data.

[0041] Step A3-1-3: Import the optimized feasible path data and path inference data into the maintenance dynamic knowledge graph to generate the maintenance enhanced knowledge graph.

[0042] First, under the constraints of the graph logic, all feasible operation sequences, i.e., feasible path data, are selected in the space of the maintenance dynamic subgraph using an enumeration method. Then, the maintenance process is modeled as a Markov decision process, where the state is the current graph state, the action is to select the next operation to be executed, and the reward is a score calculated based on completion time, cost, and quality. A reinforcement learning agent, i.e., the maintenance agent, is pre-trained using historical data or learned online in a simulation environment. The maintenance agent can adopt a three-level architecture, including a top-level controller, a middle-level planner, and a bottom-level executor. The inputs are feasible path data, maintenance node vectors, and maintenance dynamic subgraphs, and the output is optimized feasible path data.

[0043] For the current new task, the agent starts from the initial state, simulates multiple decision sequences, predicts an expected cumulative reward value for each possible path, and quantitatively evaluates its potential performance in multiple dimensions such as time, cost, and quality. Then, a multi-objective optimization algorithm is used to generate a set of Pareto optimal candidate solutions that achieve a balance among multiple mutually constraining objectives, such as different strategies that are fastest but slightly more expensive or most stable but slightly slower, and reveals the trade-offs of each choice. This outputs N optimized possible candidate maintenance paths, i.e., optimized feasible path data. The optimized feasible path data is integrated back into the maintenance dynamic knowledge graph, thus providing a basis for decision-making in subsequent communication and dynamic adjustments. Then, a relevant maintenance simulation model can be built using digital twin technology. The candidate maintenance paths are simulated using the maintenance simulation model. In each simulation, the time consumption of each process is randomly sampled from its time distribution. Finally, the simulation results are analyzed, including completion time analysis and impact analysis. The completion time analysis represents the probability distribution of the total maintenance time, for example, an 80% probability of completion within 66-72 hours. The impact analysis represents which process's time fluctuation has the greatest impact on the total project duration. This yields path projection data, which is then integrated back into the maintenance dynamic knowledge graph to form a graph that supports interactive hypothesis analysis. This allows for dynamic projection of the quantitative results of different decisions on maintenance outcomes, providing a basis for subsequent interactive hypothesis questioning by customers or engineers. For example, for questions such as "If the fault of array element A7 is more complex than expected, requiring an additional repair process, what impact will this have on the subsequent schedule?", the above operations generate an enhanced maintenance knowledge graph, and an API interface is configured for easy access. Each time a new probe is inspected or a probe maintenance process moves to the next stage, new data is imported into the enhanced maintenance knowledge graph to ensure continuous updates.

[0044] Step B: Obtain customer context data, and obtain semantic parsing data based on customer context data and the maintenance enhanced knowledge graph.

[0045] Furthermore, step B includes steps B-1 and B-2.

[0046] Step B-1: Obtain service input data, perform standardization operations on the service input data, construct a service customer profile, and obtain customer context data.

[0047] Raw customer input data is obtained through channels such as customer service chat windows, mobile apps, voice assistants, and emails. The raw input data is in the format of voice, text, or structured data. Metadata from the channels is also acquired, such as source identifiers, timestamps, and client information. The raw input data and metadata are then integrated via an API interface. At this point, a domain dictionary needs to be constructed using a maintenance-enhanced knowledge graph. For example, array element impedance values ​​and ultrasonic probe information can be extracted from the maintenance-enhanced knowledge graph. For text data, all encodings are standardized, and the text is cleaned, such as removing extra spaces and line breaks. For typos, a general spell check model is used to correct common spelling and terminology errors. Then, colloquialisms are separated based on the domain dictionary. Abbreviations or aliases are standardized into standard terms, such as unifying "B-ultrasound head" as "ultrasound probe." After processing text-formatted data, speech-formatted data can be processed. Speech can be converted into text using automatic speech recognition services, and then processed based on the aforementioned text processing steps. For structured data, these steps can be skipped. Then, for multiple independent questions that customers may ask in a single message, such as "How is the progress? Have the array elements been replaced?", sentence boundary detection in natural language processing can be used to process them, thereby separating the two questions about progress and array elements into two independent query units for subsequent processing. At this point, the output channel's metadata, processed contextual text data, and query units are included.

[0048] Next, the customer history database is invoked. This database includes data such as the corresponding customer's query records, interaction behaviors, and historical feedback. Based on the channel's metadata, the customer's identity is identified and confirmed. Data of the corresponding customer is retrieved from the historical customer database to build a customer profile, which may include static and dynamic attributes. Static attributes include information such as customer role, affiliated organization, and registration time, while dynamic attributes include information such as recent activity level, common query patterns, and technical depth. After the customer profile is built, a classification model such as FastText or BERT can be used to perform primary intent classification on the query unit. This process is used for coarse-grained division to provide prior guidance for subsequent parsing. Common classification types may include status progress queries, such as "where is the repair progressed?", reason for abnormality queries, such as "why is it delayed?", and prediction estimation requests, such as "when will it be completed?". The above customer profile, primary intent, query unit, and contextual text data are encapsulated into customer contextual data. For customer-related questions, there may also be questions that refer to terms such as "using," "the former," or "this component" as mentioned above. In this case, the methods described above can be used to process them.

[0049] Step B-2: Based on the maintenance enhanced knowledge graph and customer context data, obtain the linked entity data, intent parsing data and time logic data respectively, and integrate them into semantic parsing data.

[0050] Furthermore, step B-2 includes steps B-2-1, B-2-2, and B-2-3.

[0051] Step B-2-1: Perform entity recognition on customer context data to obtain context entity data, perform entity matching on context entity data based on maintenance enhanced knowledge graph, and obtain linked entity data.

[0052] The entity extraction method described above is used to extract entities from the query unit and context text data in the customer context data, extracting all possible entities. At this time, the extracted entities may point to components of different probes or have different meanings in the context. Therefore, for each identified entity, all possible candidate entities are searched in the maintenance enhancement knowledge graph. For example, entity A7 may be array element A7 or workstation A7. Then, the semantic similarity between the context text data and the context of each candidate entity in the graph is calculated, such as adjacent entities, descriptive text, etc. Using the entity vector embedding in step A-2 above, the cosine similarity between the entity vector and the candidate entity vector is calculated. Combining semantic similarity, semantic similarity, the frequency of the entity in the knowledge graph, and the entities already linked in the dialogue history, the optimal possible target entity is selected and the confidence is recorded. This is used to obtain the linked entity data, where each entry includes the text mentioned by the entity, the linked graph entity, the entity type, and the confidence.

[0053] Step B-2-2: Perform intent analysis on the customer context data to obtain intent parsing data.

[0054] The pre-trained intent analysis model is invoked. The intent analysis model can adopt a multi-task learning architecture, such as the BERT model, which is responsible for deep semantic encoding of customer context data. On this basis, the model executes three sub-tasks in parallel, including intent classification, argument role labeling, and semantic dependency parsing. Intent classification further judges sub-intents under the guidance of the first-level intent. Argument role labeling is to identify and label the entities corresponding to each slot in the intent framework. Semantic dependency parsing is to capture the logical relationships between words, such as comparison and causality. The model can learn the shared representation of these three tasks through joint training and use conditional random fields to ensure the consistency of labels in the output sequence.

[0055] For example, by identifying the top-level intent as comparison through intent analysis model, it can be further refined to compare the numerical changes of a certain indicator of a specific entity at different points in time, and automatically fill in the argument slots such as comparison subject, comparison indicator, and time base. Through this intent modeling, different information needs such as querying the current value, checking whether the standard is met, comparing historical data, and evaluating the trend of change can be clearly distinguished, thereby transforming the originally vague and general customer inquiry into a structured intent representation object, that is, intent parsing data.

[0056] In some possible implementations, a function can also be added that infers deeper information needs that customers have not explicitly expressed based on their historical conversations based on customer profiles. For example, if a clinician repeatedly asks about uniformity, it may imply concern about the impact on specific usage.

[0057] Step B-2-3: Perform time-series and logical analysis on the customer context data to obtain time-logic data.

[0058] All vague and relative descriptions of time, comparison, and logic in customer context data are transformed into standardized time expressions and logical relationship parsing results required for knowledge graph retrieval. For time expressions, such as "last Wednesday, two hours after the start of maintenance" in customer context data, including natural language descriptions, are normalized to 20xx-10-25 based on the specific time the conversation occurred. For logical relationship parsing results, logical relationships in customer context data are identified, such as comparison relationships, trend relationships, and conditional relationships. The benchmark object and target attribute of the comparison are parsed. For example, in the sentence "how much has the sensitivity improved compared to before maintenance," sensitivity is identified as an attribute, and the value before maintenance is the benchmark object. The time expression and logical relationship parsing results define a precise query scope for subsequent knowledge graph retrieval.

[0059] The linked entity data, intent parsing data, and temporal logic data are integrated into a unified semantic parsing graph. This graph, centered on the intent parsing data, connects related entities, attributes, time ranges, and logical relationships. At this point, the consistency within the integrated semantic graph needs to be checked. For example, it's necessary to check if all the necessary arguments for the intent have been filled, and if the entity types match the expected types of the intent. After the consistency check, the rationality of the current semantic parsing graph is verified using historical context text data. For instance, if a customer has been consistently asking about probe A, but suddenly asks about a component of probe B without explicit mention, it may indicate an incorrect referencing. In this case, queryability needs to be restricted; for example, the customer should not be able to query other customers' repair information. Then, a comprehensive confidence score can be calculated for the entire semantic parsing graph. This comprehensive confidence score integrates the aforementioned sub-steps, the consistency score of the integrated semantic graph, and the consistency score with the context and history. Finally, an integrated and verified semantic parsing graph, i.e., the semantic parsing data, is output.

[0060] Step C involves executing a query strategy based on the semantic parsing data to obtain the final result data, and simultaneously obtaining the response material data based on the final result data.

[0061] Furthermore, step C includes steps C-1 and C-2.

[0062] Step C-1: Obtain graph query instructions through semantic parsing data, and obtain the final result data based on the graph query instructions and the maintenance enhancement knowledge graph.

[0063] Understandably, semantic parsing data is transformed into a set of structured query instructions for maintenance-enhanced knowledge graphs, namely graph query instructions. Graph query instructions represent the retrieval target, filtering conditions, comparison operations, and preferred format. The retrieval target represents what information is to be obtained, the filtering conditions represent the spatiotemporal range, the comparison operations represent how the results should be processed, and the preferred format can be based on customer profiles, indicating the expected level of detail and presentation of the generated data.

[0064] Furthermore, step C-1 includes steps C-1-1, C-1-2, C-1-3, and C-1-4.

[0065] Step C-1-1: Perform complexity analysis and access planning on the graph query command to obtain the query execution plan.

[0066] The complexity of the spectral query commands is analyzed, including the number of entities involved, the depth of relationships, the time span, and the spatial range. Complex queries are marked as high-order. Simple queries can be used as a basis and directly subqueries. For complex queries, they need to be decomposed into multiple parallel subqueries. For example, querying the variation trend of the array element impedance matrix and pulse echo waveform of all array elements at the probe tip over the past week can be decomposed into subquery 1 and subquery 2. Subquery 1 means obtaining the array element impedance matrix data sequence of all array elements in the probe tip region within the past week's time window, and subquery 2 means obtaining the pulse echo waveform data sequence of all array elements within the same time window and spatial range. Based on the difficulty involved in the query, the optimal index combination is selected, including time dimension queries, Spatial dimension queries, attribute value queries, and graph structure queries are supported. For time dimension queries, fast filtering by time range is supported. For spatial dimension queries, spatial indexes can be used to quickly locate entities within a specific physical area. For attribute value queries, attribute indexes can be used for specific attributes, such as impedance greater than 50Ω. For graph structure queries, if multi-hop relationships are involved, adjacency indexes or path indexes of the graph structure can be used. The real-time requirements of the retrieval also need to be determined according to the query intent, including strong real-time queries, such as current progress, near real-time queries, such as measurement data from the past few minutes, and analytical queries, such as historical trend analysis. Finally, a query execution plan is obtained that clearly defines the access path, indexes used, data sources, and expected execution order for each subquery.

[0067] Step C-1-2: Perform adaptive queries on the maintenance enhanced knowledge graph based on the query execution plan to obtain the result data.

[0068] The enhanced maintenance knowledge graph is invoked, and the appropriate query method is adaptively selected based on the specific content within the query execution plan. This includes graph traversal queries, spatiotemporal range queries, vector similarity queries, and aggregate statistical queries. For graph traversal queries, Gremlin queries are executed within the enhanced maintenance knowledge graph to find entity relationships and paths. For spatiotemporal range queries, data points within a specific time and space range are filtered. For vector similarity queries, entities or historical cases most similar to the target vector are found. Aggregate statistical queries are used to calculate averages, trends, distributions, etc. After the query task is distributed, for dependent subqueries, such as simultaneously querying the array element impedance matrix and pulse echo waveform, the task is distributed to different... The computing nodes execute in parallel. For subqueries with dependencies, a pipeline is established for execution. For example, a graph traversal is first performed to find the list of related array element IDs. Then, the ID list is passed as input to the temporal and spatiotemporal range query for the historical data of these array elements. The query results can then be obtained. The results need to be processed in a unified format and different results need to be connected, such as probe ID-timestamp-spatial coordinates. During the processing, duplicate records that may be returned due to different query paths are eliminated, thereby generating the final results. The results may contain a structured intermediate result as a preliminary result view, which represents the integration of all related data into a logical view, but has not yet undergone final sorting and pruning.

[0069] Step C-1-3: Use the result data and the maintenance enhancement knowledge graph to perform rule reasoning to obtain enhanced result data.

[0070] The system invokes attributes and transitive relationships from the maintenance enhancement knowledge graph and inherits them into the result data. Simultaneously, it identifies frequent sequential patterns in the result data sorted by time, such as low cleanliness scores often being followed by welding failures. Based on the causal relationships and statistical correlations established earlier, it infers possible causal chains from the results. For example, if both excessively high environmental humidity and insufficient glue curing are found, it may indicate a causal relationship between the two. Thus, the system enhances the results data by adding facts, associations, or annotations obtained through reasoning.

[0071] Step C-1-4: Sort the enhanced result data by relevance to obtain the final result data.

[0072] In some possible embodiments, multi-factor scoring is applied to enhance the result data. This multi-factor scoring may include, for example, semantic relevance scoring, time scoring, source scoring, and preference scoring. For semantic relevance scoring, the entity vector embedding described above is used to calculate the cosine similarity between the query intent and the vector of each result entity. For time scoring, the query is performed on the time frame, giving higher weight to the most recent time. A time decay function, such as exponential decay, can be used to calculate the time score. For source scoring, different weights are assigned to different data sources; for example, the weight of measurement data from automated equipment may be greater than that of manually entered data, and the score is obtained based on these weights. For preference scoring, weights can be adjusted based on customer profiles; for example, engineers may pay more attention to raw data, and a score is obtained accordingly. Finally, a comprehensive ranking score can be obtained based on the above scores. The scores discussed above are not all possible scoring options; they are merely examples to better illustrate the scoring process. During implementation by technical personnel, selections, additions, deletions, or additional additions can be made based on the actual situation.

[0073] After obtaining the comprehensive ranking score, the enhanced ranking results data can be obtained. However, highly similar results may still exist. In this case, clustering is required, and only representative results of each cluster should be returned. It is also necessary to ensure that the top-ranked results cover multiple dimensions of the query intent. For example, for a query to assess quality, it is necessary to ensure that there are multiple different test data in the results, rather than just returning a large number of similar results of a certain type of data. In addition, some results that are related to the main query but have a novel perspective can be inserted at appropriate positions in the results list, such as the 3rd to 5th positions. For example, a similar historical case can be inserted to broaden the subsequent answer materials, thereby generating the final result data.

[0074] Step C-2: Obtain result quantification data based on the final result data, and obtain response material data by combining the result quantification data with the final result data.

[0075] Furthermore, step C-2 includes steps C-2-1, C-2-2, and C-2-3.

[0076] Step C-2-1 involves performing quantitative calculations and deviation analysis on the final result data to obtain quantitative result data.

[0077] The core indicators in the final result data are quantified. For numerical data sequences, such as a set of impedance values, the mean, median, standard deviation, maximum, minimum, and range of variation are calculated. For time series data, the linear regression slope, moving average, and inflection point are calculated to determine whether the trend is upward, downward, stable, or fluctuating. The data distribution pattern is also analyzed, such as normal, skewed, or multimodal, and skewness and kurtosis are calculated. For comparative queries, differences, percentage changes, and ratios are calculated. Then, the quantified data is compared with a standard baseline to perform deviation analysis, obtaining absolute and relative deviations. Absolute deviation represents the difference between the measured value and the standard value, such as the absolute difference between the factory specifications and the relative deviation. The deviation is characterized as a percentage deviation. In the context of statistical process control, the process capability index Cpk is calculated, and based on a preset threshold, each data point and the overall status are judged as qualified, warning, or unqualified. After the above process is executed, simple rules can be applied to identify key relationships in the data, including identifying anomalies in the data, such as three consecutive points showing an upward trend or a sudden jump in value exceeding three times the standard deviation; correlation indicators, such as calculating the correlation coefficient between different variables to indicate possible strong correlations, such as a negative correlation between temperature and impedance; and speculative indicators, which can generate a list of possible root cause hypotheses based on the causal chain inferred above. Thus, the results of the quantitative calculation and key relationships are integrated into the quantitative data of the results.

[0078] Step C-2-2: Obtain summary data by semantic parsing data and result quantification results.

[0079] Content selection is performed based on the query intent in the semantic parsing data. Relevant data within the quantified results are selected. For example, for progress queries, the completion status and time estimates of the process are output first, while for quality queries, the comparison between test data and standards is output first. For engineers, more original data references are retained. Then, the data can be input through commonly used large language models, such as text generators, and converted into natural language paragraphs. Probe P2: The current acoustic calibration node is 85% complete. Of the 128 completed array elements, 125 meet the sensitivity standard, and 3 are being fine-tuned. The average impedance measurement value is 45.2Ω, which meets the standard.

[0080] Step C-2-3: Obtain interactive data by semantically parsing the data and quantifying the results.

[0081] To construct visualized data, time-series data can be used to create line charts showing the trend of parameters changing over time, while category comparison data can be used to create bar charts comparing indicators of different processes, array elements, or probes. This process can be repeated to generate different visualized data. The configuration of the visualized data must be within a reasonable range, such as determining appropriate coordinate axis ranges, scales, and units. An interactive framework must also be constructed, including defining drill-down paths and linkage rules. For drill-down paths, drill-down paths are set for elements in the visualized icons; for example, unqualified array element A7 can be drilled down to its detailed historical measurement data. For linkage rules, linkage rules between multiple views are defined; for example, selecting a time range on the time axis will automatically update the spatial distribution to the average value within that time period. Once the interactive data, summary data, semantic parsing data, and result quantification data are acquired, they can be packaged into response material data.

[0082] Step D: Provide response services based on the response material data.

[0083] For step D, the data in the response material library can be used to organize a coherent narrative according to logical levels, such as core assertions, evidence, background, and prospects. Then, based on user preferences and the terminal environment, an appropriate expression format can be adaptively selected. For example, technical reports and text summaries containing raw data can be generated for engineers, or concise text summaries and graphics can be pushed to mobile users. During the response service, the maintenance status can be continuously monitored to update the knowledge graph and response material data in real time, so as to send customers real-time information and forward-looking responses.

[0084] Figure 2 The diagram illustrates an automatic response processing system for ultrasonic probe repair services, which can realize the ideas of this application, according to some embodiments of this application.

[0085] Specifically, an automated response and processing system for ultrasonic probe repair services includes: At least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the methods described above.

[0086] In this context, the processor is the control center of the electronic device. It can be a single processor or a collective term for multiple processing elements. For example, a processor can be one or more central processing units (CPUs), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement Embodiment 1 of this invention, such as one or more digital signal processors (DSPs) or one or more field-programmable gate arrays (FPGAs).

[0087] The processor can perform various functions of the electronic device by running or executing software programs stored in the memory and calling data stored in the memory. The memory is used to store the software programs that execute the present invention and is controlled by the processor for execution. Specific implementation methods can be referred to the above method embodiments, which will not be repeated here. The memory can be a real-only memory (ROM) or other types of static storage devices that can store static information and instructions, a random access memory (RAM) or other types of dynamic storage devices that can store information and instructions, or an electrically erasable programmable read-only memory (EEPROM), a compact disc read-only (CD-ROM), or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), disk storage media, or other magnetic storage devices, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and that can be accessed by a computer, but is not limited thereto. The memory can be integrated with the processor or exist independently and coupled to the processor through the interface circuit of the electronic device. This embodiment of the invention does not specifically limit this.

[0088] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0089] It should be understood that, in the embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0090] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. An automatic response processing method for ultrasonic probe repair services, characterized in that, It includes the following steps: Perform data acquisition operations and construct an initial maintenance knowledge graph through these operations. Dynamically enhance the initial maintenance knowledge graph to obtain an enhanced maintenance knowledge graph. Obtain customer context data, and then obtain semantic parsing data based on the customer context data and the enhanced maintenance knowledge graph; The query strategy is executed based on the semantic parsing data to obtain the final result data, and the response material data is obtained based on the final result data. Response service is provided based on response material data.

2. The automatic response processing method for ultrasonic probe repair service according to claim 1, characterized in that, The process of performing data acquisition operations, constructing an initial maintenance knowledge graph through these operations, dynamically enhancing the initial maintenance knowledge graph, and obtaining an enhanced maintenance knowledge graph includes: Acquire physical test data, electrical test data, maintenance business data, and business text data, while performing anomaly handling, alignment, and mapping operations to obtain maintenance baseline data and construct an initial maintenance knowledge graph based on the maintenance baseline data; Import temporal causal relationships into the initial maintenance knowledge graph and generate a maintenance dynamic knowledge graph and a maintenance dynamic subgraph. Construct a maintenance knowledge model based on the initial maintenance knowledge graph and the maintenance dynamic subgraph, and output maintenance node vector data through the maintenance knowledge model. Based on maintenance node vector data and maintenance dynamic subgraph, dynamic path data is obtained by using planning and inference strategies. Then, an enhanced maintenance knowledge graph is obtained by combining the dynamic path data with the maintenance dynamic knowledge graph.

3. The automatic response processing method for ultrasonic probe repair service according to claim 2, characterized in that, The process of obtaining dynamic path data based on maintenance node vector data and dynamic maintenance subgraphs using a planning and reasoning strategy, and then obtaining an enhanced maintenance knowledge graph through the dynamic path data and the dynamic maintenance knowledge graph, includes: Obtain feasible path data based on maintenance dynamic subgraph; Based on feasible path data, maintenance node vectors and maintenance dynamic subgraphs, and by obtaining optimized feasible path data through maintenance agents, the feasible path data is simulated and deduced to generate path deduction data. Optimized feasible path data and path deduction data are imported into the maintenance dynamic knowledge graph to generate an enhanced maintenance knowledge graph.

4. The automatic response processing method for ultrasonic probe repair service according to claim 1, characterized in that, The process of obtaining customer context data and semantic parsing data based on the customer context data and the enhanced maintenance knowledge graph includes: Acquire service input data, perform standardized operations on the service input data, build service customer profiles, and obtain customer context data; Based on the maintenance-enhanced knowledge graph and customer context data, link entity data, intent parsing data, and time logic data are obtained respectively, and integrated into semantic parsing data.

5. The automatic response processing method for ultrasonic probe repair service according to claim 4, characterized in that, The process of acquiring linked entity data, intent parsing data, and time logic data based on the maintenance-enhanced knowledge graph and customer context data includes: Entity recognition is performed on customer context data to obtain context entity data. Entity matching is performed on the context entity data based on the maintenance enhanced knowledge graph, and linked entity data is obtained. Perform intent analysis on customer context data to obtain intent parsing data; Perform time-series and logical analysis on customer context data to obtain time-logic data.

6. The automatic response processing method for ultrasonic probe repair service according to claim 1, characterized in that, The process of executing a query strategy based on semantic parsing data to obtain final result data, and simultaneously obtaining response material data based on the final result data, includes: Obtain graph query instructions by semantically parsing the data, and obtain the final result data based on the graph query instructions and the maintenance-enhanced knowledge graph. Quantitative data of the results are obtained based on the final result data, and response material data is obtained by combining the quantitative data of the results with the final result data.

7. The automatic response processing method for ultrasonic probe repair service according to claim 6, characterized in that, The process of obtaining graph query instructions through semantic parsing data, and obtaining final result data based on the graph query instructions and the maintenance-enhanced knowledge graph, includes: Perform complexity analysis and access planning on the graph query command to obtain the query execution plan; Adaptive queries are performed on the enhanced maintenance knowledge graph based on the query execution plan to obtain the result data; Enhanced result data is obtained by performing rule-based reasoning using the result data and the maintenance-enhanced knowledge graph. The enhanced result data is sorted by relevance to obtain the final result data.

8. The automatic response processing method for ultrasonic probe repair service according to claim 6, characterized in that, The process of obtaining quantitative result data based on the final result data, and obtaining response material data by comparing the quantitative result data with the final result data, includes: Quantitative calculations and deviation analysis are performed on the final result data to obtain quantitative data. Summary data is obtained by semantically parsing the data and quantifying the results. Interaction data is obtained by semantically parsing the data and quantifying the results. The results quantification data, summary data, semantic parsing data, and interaction data are integrated into response material data.

9. An automatic response and processing system for ultrasonic probe repair services, characterized in that, include: At least one processor; And a memory communicatively connected to at least one processor; wherein the memory stores instructions executable by at least one processor, the instructions being executed by at least one processor to enable at least one processor to perform any of the methods described in claims 1-8.

Images