Natural language-based operation and maintenance fault reporting intention understanding and order building method, device and equipment

Abstract

The application discloses a kind of based on natural language's operation and maintenance report barrier intention understanding and method, device and equipment, the method includes: after loading pre-trained field natural language understanding model, receiving natural language report barrier;Intention understanding and slot extraction are carried out to the natural language report barrier by the field natural language understanding model, obtain intention information and slot data;Similarity of target asset archive is calculated based on the slot data searches one thing one file asset library, and target asset archive is positioned to obtain;Based on the intention information, the slot data and the similarity of target asset archive, calculate the confidence of building single;According to the confidence of building single, carry out single building process shunt, and enter the single building process after shunt, including confirmation single building process, clarification dialogue process or transfer artificial process.The application forms complete automation, systematization closed loop from spoken language input to final single building disposal path, and greatly reduces the threshold of artificial intervention.

Description

Technical Field

[0001] This invention relates to the field of operation and maintenance security technology, and in particular to a method, device and equipment for understanding and creating operation and maintenance fault reports based on natural language. Background Technology

[0002] The unified service desk and customer self-service portal for digital infrastructure operations and maintenance receive a large number of verbal fault reports daily. Reporters use short sentences to describe the location, symptoms, and approximate equipment, requiring human operators to repeatedly ask for the asset number, project affiliation, and fault type before a service order can be created. Current processing methods include: general customer service dialogue and form filling; large-scale model intent recognition followed by Q&A or automated tasks; and proximity tags or QR codes to directly retrieve account information. None of the existing technologies have created a feasible closed-loop operation and maintenance service order creation process: "verbal fault report → domain intent and slot → retrieval of a unique asset → clarification and confirmation of service order creation; low confidence leads to transfer to human operator with completed information."

[0003] The current maintenance fault reporting and creation system has the following shortcomings: 1. Verbal reports of faults cannot generate accurate work orders. Reporters tend to describe the location and symptoms but not the asset number. Agents have to ask repeated questions to fill out the form, which takes a long time to connect. The form fields are often wrong for the project or the device.

[0004] 2. The dialog form does not verify the asset file. The existing automatic form entry directly writes the entities in the dialog into the operation and maintenance work order without searching the one-to-one asset database. It cannot handle locations with the same name, aliases, or incomplete codes, and is easy to create on the wrong assets. 3. Operation and maintenance semantic understanding stops at the question and answer stage. Power-related questions and answers and vehicle fault analysis can extract intent and slots, but they are not included in the operation and maintenance work order. The person reporting the fault still has to go through the separate work order process.

[0005] 4. When required fields are missing, the agent will only ask general follow-up questions or directly transfer the case to human agent without clarifying which required fields are missing or whether the asset is unique. When transferring the case to human agent with low confidence, the agent will not include the already filled information and candidate files, and the agent will ask the same questions repeatedly.

[0006] 5. Lack of pre-creation confirmation: After automatic form filling or automatic task issuance, the person reporting the fault cannot see the draft work order which equipment will be built and what type of fault it is. Incorrect work orders can only be corrected afterward.

[0007] Therefore, there is an urgent need for a new solution for understanding and creating fault reports based on natural language. Summary of the Invention

[0008] In view of the above problems, the present invention is proposed to provide a method, apparatus and device for understanding and creating maintenance fault reports based on natural language to overcome or at least partially solve the above problems.

[0009] Other features and advantages of the invention will become apparent from the following detailed description, or may be learned in part by practice of the invention.

[0010] According to a first aspect of the present invention, a method for understanding and creating maintenance fault reports based on natural language is provided, comprising: After loading a pre-trained domain natural language understanding model, a natural language error report is received, which is text information or text information obtained through speech conversion. The domain natural language understanding model is used to perform intent understanding and slot extraction on the natural language report to obtain intent information and slot data. Based on the slot data, a one-item-one-file asset database is retrieved, the similarity of the target asset file is calculated, and the target asset file is located. Based on the similarity of the intent information, the slot data, and the target asset file, the order creation confidence level is calculated; Based on the confidence level of the order creation process, the order creation process is diverted and then enters the diverted order creation process, including the order confirmation process, the clarification dialogue process, or the transfer to a human agent process.

[0011] In some embodiments of the present invention, the domain natural language understanding model includes at least an encoder, an intent classification head, and a slot labeling head; the step of performing intent understanding and slot extraction on the natural language report using the domain natural language understanding model to obtain intent information and slot data includes: The natural language error report is input into the encoder for encoding processing to obtain sentence vectors and character vector sequences; The sentence vector is classified using the intent classification head to obtain the intent information, which includes intent labels and intent confidence scores. The word vector sequence is sequence-annotated using a slot labeling header to extract the slot data, which includes at least the asset designation, fault phenomenon, location of occurrence, and urgency level.

[0012] In some embodiments of the present invention, the step of retrieving a one-item-one-file asset database based on the slot data, calculating the similarity of the target asset file, and locating the target asset file includes: The asset designations in the slot data are standardized. The standardized asset designations are precisely matched with the asset files in the one-item-one-file asset database. When the asset designation is exactly matched with the asset file and is unique, the matched asset file will be used as the target asset file. If an exact match fails and no unique asset is found, the error reporting statement vector of the natural language error is obtained, and the cosine similarity between the error reporting statement vector and the file vector of each asset file in the one-item-one-file asset database is calculated. The asset file with the highest cosine similarity and whose similarity difference with other candidate asset files reaches a preset interval is selected as the target asset file.

[0013] In some embodiments of the present invention, calculating the order creation confidence level based on the similarity of the intent information, the slot data, and the target asset profile includes: The number of required slots that have been filled and passed the verification in the slot data is counted, and the completeness rate of required slots is calculated. Using preset intention confidence weights, slot completeness weights, and asset similarity weights, the intention confidence, the required slot completeness, and the similarity of the target asset file are weighted and summed to obtain the order creation confidence.

[0014] In some embodiments of the present invention, the step of streamlining the order creation process based on the order creation confidence level includes: When the order creation confidence level is greater than or equal to the preset confirmation order creation threshold, and the number of filled mandatory slots in the slot data is equal to the total number of mandatory slots, and the target asset file is unique, the order is diverted to the confirmation order creation process. When the confidence level of the order is less than or equal to the preset threshold for transferring to human assistance, or when a clear instruction from the person reporting the fault is received to request human assistance, the order is diverted to the human assistance process. When the confidence level of the order creation is between the preset upper threshold for confirming the order creation and the preset lower threshold for transferring to manual processing, or when the asset file is not unique, or when the number of filled slots is not equal to the total number of filled slots, the user is diverted to the clarification dialogue process.

[0015] In some embodiments of the present invention, the diversion into the clarification dialogue process includes: Based on the unfilled required slots or candidate asset list in the slot data, follow-up information is generated. The candidate asset list is a set of multiple asset files retrieved from the one-item-one-file asset database when the target asset file is not uniquely determined. Receive the response from the person reporting the fault based on the follow-up information, and update the response content to the slot data; If the updated slot data is still incomplete or the asset files in the candidate asset list are still not unique, the clarification dialogue process will be repeated. After the number of repeated rounds reaches the preset limit, the process will switch to manual intervention.

[0016] In some embodiments of the present invention, the process of diverting traffic to the confirmation and order creation process includes: Show the fault reporter a draft work order, which includes the asset name and location information of the target asset file, the fault phenomenon and urgency level in the slot data, and the intent tag; Receive confirmation instructions from the person reporting the fault regarding the draft work order, and write it into the maintenance work order after confirmation; If a rejection instruction is received from the person reporting the fault regarding any field in the draft work order, the clarification dialogue process will only be executed for that field.

[0017] In some embodiments of the present invention, the diversion to manual transfer process includes: Generate a structured handover record, which includes at least a dialogue summary, intent confidence, filled slots, candidate asset list, order creation confidence, and transfer reason; The handover record is written into the session handover table with the session identifier as the primary key, and a pending notification is sent to the human operation and maintenance agent so that the human operation and maintenance agent can complete the operation and maintenance work order based on the filled slot.

[0018] According to a second aspect of the present invention, a device for understanding and creating maintenance fault reports based on natural language is provided, the device comprising: The information receiving module is used to receive natural language reports after loading a pre-trained domain natural language understanding model. The natural language reports are text information or text information obtained through speech conversion. The information parsing module is used to perform intent understanding and slot extraction on the natural language report using the domain natural language understanding model to obtain intent information and slot data. The similarity calculation module is used to retrieve the one-item-one-file asset database based on the slot data, calculate the similarity of the target asset file, and locate the target asset file. The confidence calculation module is used to calculate the order creation confidence based on the similarity between the intent information, the slot data, and the target asset file. The traffic diversion execution module is used to divert the order creation process according to the order creation confidence level, and enter the diverted order creation process, including the order confirmation process, the clarification dialogue process, or the transfer to human agent process.

[0019] According to a third aspect of the present invention, a computer device is provided, including a processor and a memory, the memory storing computer program instructions executable by the processor, wherein when the processor executes the computer program instructions, it implements the instructions as described in any of the above methods.

[0020] The technical solutions provided in the embodiments of the present invention have at least the following technical effects or advantages: This invention provides a method, apparatus, and device for understanding the intent of maintenance and maintenance (M&A) fault reports and creating service requests based on natural language. This method uses a pre-trained domain-specific natural language understanding model to parse received natural language fault reports to obtain intent information and slot data. Based on the slot data, it locates the target asset file, calculates the service request confidence level, and routes the service request process according to the confidence level. This allows the system to directly start from the complainant's spoken expression, and after asset confirmation and confidence level assessment, automatically and intelligently guide the process towards service request routing, such as confirming the service request, clarifying the dialogue, or transferring to a human agent. This forms a complete automated and systematic closed loop from spoken input to the final service request handling path, greatly reducing the threshold for human intervention.

[0021] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 A flowchart illustrating a method for understanding and creating maintenance fault reports based on natural language, as provided in an embodiment of the present invention. Figure 2 A schematic diagram of the principle structure of a device for understanding and creating maintenance fault reports based on natural language, provided in an embodiment of the present invention; Figure 3 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0024] Exemplary embodiments of this application will now be described in more detail with reference to the accompanying drawings.

[0025] The accompanying drawings illustrate various structural schematics according to embodiments of this application. These drawings are not to scale, and some details have been enlarged for clarity, and some details may have been omitted. The shapes of the various regions and layers shown in the drawings, as well as their relative sizes and positional relationships, are merely exemplary and may deviate from reality due to manufacturing tolerances or technical limitations. Furthermore, those skilled in the art can design regions / layers with different shapes, sizes, and relative positions as needed.

[0026] It should be noted that the terms "comprising" and "having," and any variations thereof, as used in this application, are intended to cover non-exclusive inclusion. The term "multiple" as used in this application refers to two or more. In the context of this application, similar or identical components may be denoted by the same or similar reference numerals.

[0027] To better understand the above technical solutions, the following will describe the above technical solutions in detail with reference to specific implementation methods. It should be understood that the embodiments of this application and the specific features in the embodiments are detailed descriptions of the technical solutions of the present invention, rather than limitations on the technical solutions of the present invention. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0028] Figure 1 This is a flowchart illustrating a method for understanding and creating maintenance fault reports based on natural language, as provided in an embodiment of the present invention. Figure 1 As shown, this method for understanding and creating service requests based on natural language processing includes the following steps: S1. After loading the pre-trained domain natural language understanding model, receive natural language error reports, wherein the natural language error reports are text information or text information obtained through speech conversion. In this embodiment of the invention, the domain natural language understanding model is obtained through pre-training (e.g., fine-tuning using operational spoken language corpus); the domain natural language understanding model includes at least an encoder, an intent classification head, and a slot labeling head. The encoder uses a pre-trained Chinese language model and is responsible for converting the input natural language fault report (text) into vectors that can be processed by a computer. The intent classification head determines the user's fault report intent based on the sentence vectors output by the encoder, such as repair, inspection anomaly, consultation, or complaint, and provides an intent confidence score (C_int). The slot labeling head performs sequence annotation (BIO) on the word / word vector sequence output by the encoder, extracts the key information fragments required for creating the order, i.e., slots, such as asset reference, fault phenomenon, and location of occurrence, and provides a slot location confidence score for each slot.

[0029] In this embodiment of the invention, the natural language fault report is text information or text information obtained through speech conversion; the person reporting the fault submits a fault description in voice or text form through a self-service portal or service desk session, thereby receiving the natural language fault report. When the natural language fault report is in voice form, it can be converted into text first through a speech-to-text service; the natural language fault report will be written into the current session record and retain historical round information for subsequent reference resolution.

[0030] S2. Perform intent understanding and slot extraction on the natural language report using the domain natural language understanding model to obtain intent information and slot data; This invention embodiment uses the domain natural language understanding model to perform intent understanding and slot extraction on the natural language fault report to obtain intent information and slot data, including: inputting the natural language fault report into the encoder for encoding processing to obtain sentence vectors and word vector sequences; classifying the sentence vectors using the intent classification head to obtain intent information, which includes intent labels and intent confidence; and performing sequence annotation on the word vector sequences using the slot annotation head to extract the slot data, which includes at least the following fields: asset reference, fault phenomenon, location of occurrence, and urgency level.

[0031] For example, when the natural language fault report is "The PTZ camera at the East Gate is displaying a distorted image, it's urgent, I'm on duty.", the obtained intent information and slot data are as follows: Intent tag Intent = Repair, Intent confidence C_int = 0.92; Asset index = East Gate PTZ camera; Fault phenomenon = Display distortion; Location of occurrence = East Gate; Urgency level = High; Contact information = Duty room.

[0032] S3. Based on the slot data, retrieve the one-item-one-file asset database, calculate the similarity of the target asset file, and locate the target asset file. In this embodiment of the invention, the one-item-one-file asset database is a collection of operation and maintenance files established with a single asset file (single asset record) as the smallest unit. That is, each asset file (one item) has an independent and detailed file (one file). The asset file includes, for example, an asset identifier / asset number (such as CAM-DF-01, which is the unique ID number of the asset file), an asset name (such as the No. 1 PTZ camera at the East Gate), an alias (such as the PTZ camera at the East Gate), an installation location (such as the guard post at the East Gate), a project to which the asset file belongs (indicating which project the asset file belongs to), a category (indicating the type of equipment, such as a network camera), and a historical work order entry (associating the historical operation and maintenance work order records of the asset file), etc.

[0033] This invention embodiment retrieves a one-item-one-file asset database based on the slot data, calculates the similarity of the target asset file, and locates the target asset file, including: The asset designations in the slot data are standardized; the standardization process is used to standardize the format of the asset designations (or incomplete numbers), such as removing spaces, converting full-width characters to half-width characters, and unifying the capitalization of letters; The standardized asset designation is precisely matched with the asset file in the one-item-one-file asset database; the standardized asset designation is compared with the fields of the asset file in the one-item-one-file asset database, including, for example, the asset identifier / asset number field. When the asset designation is exactly matched with the asset file and is unique, the matched asset file will be used as the target asset file. If an exact match fails and no unique asset is found, the fault report vector of the natural language fault report is obtained, and the cosine similarity between the fault report vector and the file vector of each asset file in the one-item-one-file asset database is calculated. If an exact match is not found, the fault report vector and the file vector are calculated and obtained separately. The file vector is a fixed-dimensional vector obtained by concatenating multiple fields in the asset file (e.g., including asset name, alias, installation location, project, category, etc.) and encoding it through the encoder. The file vector is generated offline, and when the asset file changes, only the file vector of that asset file needs to be recalculated. The fault report vector... The data is generated only during the processing of the current fault report. When the slot data already contains the fields of asset designation, location, and fault phenomenon, these three fields are concatenated into text in a fixed order (format: designation / location / phenomenon). This concatenated text is then input into the encoder for encoding to obtain a fixed-dimensional vector (with the same dimension as the file vector). The fault report vector and the file vector are both generated based on the same set of pre-trained Chinese language models (encoders), and they extract semantic features from two dimensions: natural language fault reports (i.e., user-generated colloquial fault reports) and asset files, respectively. The cosine similarity is, for example, the inner product of the fault report vector and all candidate file vectors. The asset file with the highest cosine similarity and whose similarity difference with other candidate asset files reaches a preset interval is selected as the target asset file. For the two asset files with the highest cosine similarity and the second highest cosine similarity among other candidate asset files, the similarity difference between the two is calculated to confirm whether the similarity difference reaches a preset interval. If the preset interval is reached, the asset file with the highest cosine similarity is selected as the target asset file. If the preset interval is not reached or the number of asset files in the candidate asset list exceeds the allowed number of displays, it is determined that the asset file is not unique. In this case, the asset files in the current candidate asset list can be added to the draft candidate list, along with their respective cosine similarities. For example, this can be used as one of the judgment conditions in the subsequent order creation process.

[0034] S4. Calculate the order creation confidence level based on the similarity between the intent information, the slot data, and the target asset file; This invention, in its embodiments, calculates the order creation confidence level based on the similarity between the intent information, the slot data, and the target asset profile, including: The number of required slots that have been filled and passed verification in the slot data is counted, and the required slot completeness rate is calculated. For example, the required slot completeness rate is the ratio of the number of required slots that have been filled and passed verification in the slot data (number of slots) to the total number of required slots required in the intent information (number of slots). A slot that has passed verification means that the value is not empty and that the asset reference type slot can form at least one candidate asset file in the one-item-one-file asset library (or the person reporting the fault has selected it from the candidate asset list). Asset references that are only extracted from the fault text but cannot be retrieved for any asset file are not counted as required slots. Using preset intention confidence weights, slot completeness weights, and asset similarity weights, the intention confidence, the required slot completeness, and the similarity of the target asset file are weighted and summed to obtain the order creation confidence.

[0035] For example, the confidence level of the establishment In the formula, the intention confidence weight For example, the slot integrity rate weight is 0.3-0.5. For example, the asset similarity weight is 0.2-0.4. For example, the value is 0.2-0.4, and the sum of the intention confidence weight, slot completeness weight, and asset similarity weight is 1; The confidence level of the stated intent. ; The completeness rate of the required slots. ; Similarity to the target asset files. .

[0036] S5. Based on the order creation confidence level, the order creation process is diverted and the order creation process is entered after diversion, including the order creation confirmation process, the clarification dialogue process, or the transfer to human agent process.

[0037] This invention, based on the order creation confidence level, performs order creation process routing as follows: When the order creation confidence level is greater than or equal to a preset upper threshold for confirmation, and the number of filled mandatory slots equals the total number of mandatory slots, and the target asset file is unique, the order is routed to the confirmation order creation process; when the order creation confidence level is less than or equal to a preset lower threshold for manual intervention (less than the preset upper threshold for confirmation), or when a clear instruction from the reporting party to provide manual intervention is received, the order is routed to the manual intervention process; when the order creation confidence level is between the preset upper threshold for confirmation and the preset lower threshold for manual intervention, or the asset file is not unique, or the number of filled mandatory slots does not equal the total number of mandatory slots, the order is routed to the clarification dialogue process. This invention ensures that even with a high confidence level, a maintenance work order will not be directly generated if key information is missing.

[0038] In this embodiment of the invention, the diversion into the clarification dialogue process includes: Based on the unfilled required slots or candidate asset list in the slot data, follow-up information is generated. The candidate asset list is a set of multiple asset files retrieved from the one-item-one-file asset database when the target asset file is not uniquely determined. Receive the response from the person reporting the fault based on the follow-up information, and update the response content to the slot data; If the updated slot data is still incomplete or the asset files in the candidate asset list are still not unique, the clarification dialogue process will be repeated. After the number of repeated rounds reaches the preset limit, the process will switch to manual intervention.

[0039] In this embodiment of the invention, the process of diverting traffic to confirm order creation includes: Show the fault reporter a draft work order, which includes the asset name and location information of the target asset file, the fault phenomenon and urgency level in the slot data, and the intent tag; Receive confirmation instructions from the person reporting the fault regarding the draft work order, and write it into the maintenance work order after confirmation; If a rejection instruction is received from the person reporting the fault regarding any field in the draft work order, the clarification dialogue process will only be executed for that field.

[0040] In this embodiment of the invention, the diversion to manual transfer process includes: Generate a structured handover record, which includes at least a dialogue summary, intent confidence, filled slots, candidate asset list, order creation confidence, and transfer reason; The handover record is written into the session handover table with the session identifier as the primary key, and a pending notification is sent to the human operation and maintenance agent so that the human operation and maintenance agent can complete the operation and maintenance work order based on the filled slot.

[0041] Example 1 Receive natural language reports of faults from the reporting party, such as "The PTZ camera at Dongmen is displaying a distorted image, it's urgent, I'm in the duty room." The domain natural language understanding model was used to process the data to obtain intent information (intent label = maintenance, intent confidence C_int = 0.92) and slot data (asset designation = East Gate PTZ camera, fault phenomenon = screen flickering, location of occurrence = East Gate, urgency level = high, contact information = duty room). All five mandatory slots were filled and passed the verification, and the completeness rate of mandatory slots was ρ = 1.

[0042] The system retrieves the target asset file from the one-item-one-file asset database. Based on the slot data, it performs a precise match between "Dongmen PTZ Machine" and "Dongmen" to locate the unique target asset file (Dongmen No. 1 PTZ Machine). The calculated similarity Sa is 0.949, and the difference between it and the second highest score is greater than the set interval of 0.08, indicating that the asset is unique.

[0043] Based on intent information (C_int=0.92), slot data (ρ=1), and the similarity of the target asset file (Sa=0.949), the order creation confidence level C = 0.4×0.92 + 0.3×1 + 0.3×0.949 = 0.953 is calculated.

[0044] Since the order creation confidence level C (0.953) is greater than or equal to the preset confirmation order creation threshold of 0.85, and all required slots are filled and the target asset file is unique, the order creation confirmation process begins.

[0045] The system displays a draft work order to the reporting party, including the asset name "East Gate No. 1 PTZ Camera," location information "East Gate Guard Post," fault symptom "screen flickering," and urgency level "high." After receiving confirmation from the reporting party regarding the draft work order, the system writes it into a maintenance work order and displays the work order number. This embodiment does not involve a clarification dialogue process or transfer to a manual process.

[0046] Example 2 Receive fault reports from users in natural language, such as "The surveillance system is broken." The domain natural language understanding model was used to process the data and obtain intent information (intent label = repair, intent confidence C_int = 0.82); however, the asset reference in the slot data was too broad (unique asset not filled), resulting in a required slot completeness rate ρ = 0.8 (one required slot is missing).

[0047] The system retrieves the target asset file from the one-item-one-file asset database and performs precise matching based on slot data, recalling "East Gate No. 1 PTZ Camera" and "West Gate No. 2 PTZ Camera" as candidate assets. Because the difference between the highest and second-highest similarity scores is less than 0.08, the asset is deemed not unique, and a unique target asset file cannot be located.

[0048] Based on the intent information (C_int=0.82), slot data (ρ=0.8), and the target asset file similarity at this time (Sa=0.88), the order creation confidence level C = 0.4×0.82 + 0.3×0.8 + 0.3×0.88 = 0.832 is calculated.

[0049] Since the confidence level of the order creation C (0.832) falls between the preset lower threshold of 0.55 for manual transfer and the preset upper threshold of 0.85 for order confirmation, and the asset is not unique, the process of clarifying the dialogue begins.

[0050] In this embodiment of the invention, follow-up questions are generated based on the candidate asset list, such as: "Is it the PTZ camera at the East Gate No. 1 or the PTZ camera at the West Gate No. 2?", and the person reporting the fault replies "the one at the East Gate". The slot data is updated based on the reply, that is, the asset designation in the slot data is updated to "PTZ camera at the East Gate No. 1".

[0051] Subsequently, the search for the target asset file in the one-item-one-file asset database was performed again. The target asset file was uniquely matched, and the similarity Sa of the target asset file increased to 0.94, and the completeness rate ρ of the required slots became 1. The creation order confidence score C was calculated again to be 0.926. At this time, C≥0.85, and the creation order confirmation process was initiated. After the person reporting the problem confirmed it, it was written into the maintenance work order.

[0052] Example 3 Receive natural language reports of faults from the person reporting the fault, such as "That thing keeps making a loud noise, hurry up." The domain natural language understanding model is used to process the information to obtain intent information (intent swings between repair and complaint, intent confidence C_int=0.48); only the urgency level (high) and contact information are filled in the slot data (asset reference and fault phenomenon are empty), and the current required slot integrity rate ρ=0.4.

[0053] The system retrieves the target asset file from the one-item-one-file asset database and performs precise matching based on slot data. However, due to the inability to construct a valid asset designation, the recall set is empty, the candidate asset list is empty, Sa=0, and the target asset file cannot be located.

[0054] Based on intent information (C_int=0.48), slot data (ρ=0.4), and the similarity of the target asset file (Sa=0), the order creation confidence level C = 0.4×0.48 + 0.3×0.4 + 0.3×0 = 0.312 is calculated.

[0055] Since the confidence level of the order creation C (0.312) is less than or equal to the preset threshold for manual transfer (0.55), the order is transferred to manual transfer.

[0056] Generate a structured handover record (including dialogue summary, intent confidence 0.48, filled slots, empty candidate asset list, order creation confidence 0.312, and transfer reason "order creation confidence is below the lower threshold"), write it into the session handover table with the session identifier as the primary key, and send a pending notification to the human operations and maintenance agent; the human operations and maintenance agent completes the operations and maintenance work order based on the filled slots.

[0057] Compared with existing technologies, the natural language-based operation and maintenance fault reporting intent understanding and service request method described in this invention has the following advantages: 1. In this embodiment of the invention, a pre-trained domain natural language understanding model is used to parse the received natural language fault reports to obtain intent information and slot data. Based on the slot data, the target asset file is located, the confidence level for establishing a fault report is calculated, and the fault report establishment process is diverted according to the confidence level. This allows the system to start directly from the verbal expression of the person reporting the fault, and automatically and intelligently proceed to the fault report establishment process diversion after asset confirmation and confidence level assessment. This process can be either a confirmation process, a clarification dialogue process, or a transfer to a human process. This forms a complete automated and systematic closed loop from verbal input to the final fault report handling path, which greatly reduces the threshold for human intervention.

[0058] 2. This invention utilizes a domain-specific natural language understanding model fine-tuned from operational speech corpus to accurately extract slot data such as asset designation, fault phenomena, location of occurrence, and urgency level. When retrieving a one-item-one-file asset database based on slot data, it employs a combination of standardized precise matching and similarity between fault report statement vectors and file vectors. Furthermore, it sets a judgment rule that the difference between the highest and second-highest scores reaches a preset interval. This effectively overcomes the shortcomings of relying solely on text entity extraction, which cannot handle locations with the same name, aliases, or incomplete codes. It can accurately pinpoint ambiguous designations in spoken language as unique target asset files, fundamentally avoiding the problem of incorrect equipment or projects during order creation.

[0059] 3. The order creation confidence score is obtained by weighted summation and fusion of three factors: intent confidence score, required slot completion rate, and target asset file similarity score. The system uses preset upper threshold for confirmation and preset lower threshold for manual transfer to strictly separate the order creation process. This mechanism can accurately control the triggering conditions of the confirmation order creation process, clarification dialogue process, and manual transfer process according to the completeness and clarity of the data entered, avoiding the defects of general follow-up questions or low-quality automatic order creation in existing technologies.

[0060] 4. In the clarification dialogue process, this invention only generates follow-up questions for unfilled required slots or candidate asset lists, and adopts an incremental update mechanism to avoid repeating confirmed fields, significantly improving interaction efficiency. At the same time, in the confirmation order creation process, a draft work order containing the target asset file, fault phenomenon, intent type, and urgency level will be displayed before the person reporting the fault confirms. The work order will only be written after the person reporting the fault confirms. If the person reporting the fault rejects the work order, the process will accurately return to the clarification process for the corresponding field, greatly reducing the cost of subsequent modification after incorrect work order creation due to model misunderstanding.

[0061] 5. When the confidence level of the order creation is insufficient, this invention does not simply transfer the chat history to the manual process. Instead, it generates structured handover information that includes a dialogue summary, intent confidence level, filled slots, candidate asset list, order creation confidence level, and transfer reason. This handover information is written into the session handover table with the session identifier as the primary key, and a pending notification is sent to the human operations and maintenance agent. This allows the human operations and maintenance agent to complete and write the operation and maintenance work order based on the existing filled slots and candidate asset list without having to ask the person reporting the problem from the beginning. This truly achieves efficient human-machine collaboration and ensures the accuracy and processing efficiency of order creation in complex scenarios.

[0062] Based on the above embodiments, as a supplement to the above... Figure 1 The present invention provides an embodiment of a device for understanding and creating maintenance fault reports based on natural language, which implements the method shown. Figure 1 Corresponding to the method embodiments shown, this device can be specifically applied to various electronic devices, see reference. Figure 2 As shown, the natural language-based maintenance fault reporting intent understanding and order creation device includes: The information receiving module 100 is used to receive natural language fault reports after loading a pre-trained domain natural language understanding model. The natural language fault reports are text information or text information obtained through speech conversion. The information parsing module 200 is used to perform intent understanding and slot extraction on the natural language report through the domain natural language understanding model to obtain intent information and slot data; The similarity calculation module 300 is used to retrieve the one-item-one-file asset database based on the slot data, calculate the similarity of the target asset file, and locate the target asset file. The confidence calculation module 400 is used to calculate the order creation confidence based on the similarity between the intent information, the slot data, and the target asset file. The flow control module 500 is used to flow control of the order creation process according to the order creation confidence level, and enter the order creation process after flow control, including the order creation confirmation process, the clarification dialogue process, or the transfer to human agent process.

[0063] The modules in the aforementioned natural language-based maintenance fault reporting intent understanding and order creation device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of the computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0064] The natural language-based maintenance fault reporting intent understanding and order creation device described in this embodiment of the invention can execute the natural language-based maintenance fault reporting intent understanding and order creation method provided in the above embodiments. The natural language-based maintenance fault reporting intent understanding and order creation device has the corresponding functional steps and beneficial effects of the natural language-based maintenance fault reporting intent understanding and order creation method described in the above embodiments. For details, please refer to the embodiments of the natural language-based maintenance fault reporting intent understanding and order creation method described above. The embodiments of this invention will not be repeated here.

[0065] In one exemplary embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 3 As shown, the computer device includes a processor, memory, input / output interfaces, a communication interface, a display unit, and an input device. The processor, memory, and input / output interfaces are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interfaces. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The input / output interfaces are used for exchanging information between the processor and external devices. The communication interface, such as a network interface card, is used for wired or wireless communication with external terminals. Wireless communication can be achieved through Wi-Fi, mobile cellular networks, Near Field Communication (NFC), or other technologies. The display unit is used to create a visually visible image and can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0066] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0067] In one exemplary embodiment, a chip is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the method in any of the above embodiments.

[0068] In one exemplary embodiment, a network interface card is provided, including a chip as described in any of the above embodiments and multiple interfaces, wherein the chip communicates externally through the interfaces.

[0069] In one embodiment, a computer device is also provided, including a processor, a chip in any of the above embodiments, or a network interface card in any of the above embodiments, wherein the chip or the network interface card is used to schedule packets to the processor or the chip or the network interface card itself for processing, and the processor is used to process the packets scheduled by the chip or the network interface card.

[0070] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0071] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0072] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory may include read-only memory (Read-Only Memory). Only memory (ROM), magnetic tape, floppy disk, flash memory, optical storage, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory may include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc. The databases involved in the embodiments provided in this application may include at least one of relational databases and non-relational databases. Non-relational databases may include distributed databases based on blockchain, etc., and are not limited thereto. The processors involved in the various embodiments provided in this application may be general-purpose processors, central processing units, graphics processors, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited thereto.

[0073] Numerous specific details are set forth in the specification provided herein. However, it will be understood that embodiments of the invention may be practiced without these specific details. In some instances, well-known methods, structures, and techniques have not been shown in detail so as not to obscure the understanding of this specification.

[0074] Similarly, it should be understood that, for the purpose of simplification and aiding understanding of one or more aspects of the invention, various features of the invention are sometimes grouped together in a single embodiment, figure, or description thereof in the description of exemplary embodiments of the invention above. Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and it should be noted that the above embodiments are illustrative of the invention and not restrictive, and that alternative embodiments can be devised by those skilled in the art without departing from its scope.

Claims

1. A method for understanding and creating maintenance fault reports based on natural language, characterized in that, include: After loading a pre-trained domain natural language understanding model, a natural language error report is received, which is text information or text information obtained through speech conversion. The domain natural language understanding model is used to perform intent understanding and slot extraction on the natural language report to obtain intent information and slot data. Based on the slot data, a one-item-one-file asset database is retrieved, the similarity of the target asset file is calculated, and the target asset file is located. Based on the similarity of the intent information, the slot data, and the target asset file, the order creation confidence level is calculated; Based on the order creation confidence level, the order creation process is diverted and then enters the diverted order creation process, including the order confirmation process, the clarification dialogue process, or the transfer to a human agent process. The domain natural language understanding model includes at least an encoder, an intent classification head, and a slot labeling head; The step of performing intent understanding and slot extraction on the natural language report using the domain natural language understanding model to obtain intent information and slot data includes: The natural language error report is input into the encoder for encoding processing to obtain sentence vectors and character vector sequences; The sentence vector is classified using the intent classification head to obtain the intent information, which includes intent labels and intent confidence scores. The word vector sequence is sequence-annotated using a slot labeling header to extract the slot data, which includes at least the asset designation, fault phenomenon, location of occurrence, and urgency level.

2. The method for understanding and creating maintenance fault reports based on natural language according to claim 1, characterized in that, The step of retrieving a one-item-one-file asset database based on the slot data, calculating the similarity of the target asset file, and locating the target asset file includes: The asset designations in the slot data are standardized. The standardized asset designations are precisely matched with the asset files in the one-item-one-file asset database. When the asset designation is exactly matched with the asset file and is unique, the matched asset file will be used as the target asset file. If an exact match fails and no unique asset is found, the error reporting statement vector of the natural language error is obtained, and the cosine similarity between the error reporting statement vector and the file vector of each asset file in the one-item-one-file asset database is calculated. The asset file with the highest cosine similarity and whose similarity difference with other candidate asset files reaches a preset interval is selected as the target asset file.

3. The method for understanding and creating maintenance fault reports based on natural language according to claim 1, characterized in that, The calculation of order creation confidence based on the similarity of the intent information, the slot data, and the target asset file includes: The number of required slots that have been filled and passed the verification in the slot data is counted, and the completeness rate of required slots is calculated. Using preset intention confidence weights, slot completeness weights, and asset similarity weights, the intention confidence, the required slot completeness, and the similarity of the target asset file are weighted and summed to obtain the order creation confidence.

4. The method for understanding and creating maintenance fault reports based on natural language according to claim 1, characterized in that, The process of streamlining the order creation process based on the order creation confidence level includes: When the order creation confidence level is greater than or equal to the preset confirmation order creation threshold, and the number of filled mandatory slots in the slot data is equal to the total number of mandatory slots, and the target asset file is unique, the order is diverted to the confirmation order creation process. When the confidence level of the order is less than or equal to the preset threshold for transferring to human assistance, or when a clear instruction from the person reporting the fault is received to request human assistance, the order is diverted to the human assistance process. When the confidence level of the order creation is between the preset upper threshold for confirming the order creation and the preset lower threshold for transferring to manual processing, or when the asset file is not unique, or when the number of filled slots is not equal to the total number of filled slots, the user is diverted to the clarification dialogue process.

5. The method for understanding and creating maintenance fault reports based on natural language according to claim 4, characterized in that, The process of diverting traffic to the clarification dialogue includes: Based on the unfilled required slots or candidate asset list in the slot data, follow-up information is generated. The candidate asset list is a set of multiple asset files retrieved from the one-item-one-file asset database when the target asset file is not uniquely determined. Receive the response from the person reporting the fault based on the follow-up information, and update the response content to the slot data; If the updated slot data is still incomplete or the asset files in the candidate asset list are still not unique, the clarification dialogue process will be repeated. After the number of repeated rounds reaches the preset limit, the process will switch to manual intervention.

6. The method for understanding and creating maintenance fault reports based on natural language according to claim 5, characterized in that, The process of routing and confirming order creation includes: Show the fault reporter a draft work order, which includes the asset name and location information of the target asset file, the fault phenomenon and urgency level in the slot data, and the intent tag; Receive confirmation instructions from the person reporting the fault regarding the draft work order, and write it into the maintenance work order after confirmation; If a rejection instruction is received from the person reporting the fault regarding any field in the draft work order, the clarification dialogue process will only be executed for that field.

7. The method for understanding and creating maintenance fault reports based on natural language according to claim 6, characterized in that, The process of diverting patients to manual intervention includes: Generate a structured handover record, which includes at least a dialogue summary, intent confidence, filled slots, candidate asset list, order creation confidence, and transfer reason; The handover record is written into the session handover table with the session identifier as the primary key, and a pending notification is sent to the human operation and maintenance agent so that the human operation and maintenance agent can complete the operation and maintenance work order based on the filled slot.

8. A device for understanding and creating maintenance fault reports based on natural language, applied to the method described in any one of claims 1-7, characterized in that, The device includes: The information receiving module is used to receive natural language reports after loading a pre-trained domain natural language understanding model. The natural language reports are text information or text information obtained through speech conversion. The information parsing module is used to perform intent understanding and slot extraction on the natural language report using the domain natural language understanding model to obtain intent information and slot data. The similarity calculation module is used to retrieve the one-item-one-file asset database based on the slot data, calculate the similarity of the target asset file, and locate the target asset file. The confidence calculation module is used to calculate the order creation confidence based on the similarity between the intent information, the slot data, and the target asset file. The traffic diversion execution module is used to divert the order creation process according to the order creation confidence level, and enter the diverted order creation process, including the order confirmation process, the clarification dialogue process, or the transfer to human agent process.

9. A computer device comprising a processor and a memory, characterized in that, The memory stores computer program instructions that can be executed by the processor, and when the processor executes the computer program instructions, it implements the instructions of the method as described in any one of claims 1-7.

Images