A question-answering guided safety inspection system and method based on AI big model

Abstract

The present invention provides a question-and-answer guided safety inspection system and method based on a large AI model. The system includes: a model training module for processing safety inspection-related data using natural language processing technology and training a safety inspection guidance model based on the DeepSeek R1 large model; and an inspection module for implementing inspection interactions based on the safety inspection guidance model and the voice of inspectors. This system and method, based on a large AI model, uses natural language processing technology to process safety inspection-related data and uses the DeepSeek R1 large model to learn safety inspection inference rules to obtain a safety inspection guidance model. The model automatically recognizes the voice of inspectors to implement inspection interactions, providing real-time alerts for potential safety hazards and guiding inspectors to quickly identify them, thereby reducing omissions and misjudgments during manual inspections.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a question-and-answer guided safety inspection system and method based on an AI big model. Background Art

[0002] Traditional manual inspection systems rely primarily on observation and recording by inspectors. While this approach can meet basic inspection needs to a certain extent, its limitations are also very obvious. First, the professional knowledge and experience of inspectors directly affect the quality and effectiveness of inspections. When inspectors lack sufficient knowledge of inspection equipment, they may not be able to execute reasonable inspection steps, resulting in inappropriate inspection strategies and even the risk of omissions or misjudgments. Second, the efficiency and quality of manual inspections are easily affected by the external environment and the inspectors' own state. Inspection work usually requires long periods of time in complex environments. Inspectors are prone to fatigue, lack of concentration, and other factors, which can lead to reduced inspection efficiency and even missed inspections and false detections.

[0003] In view of this, there is an urgent need for a question-and-answer guided safety inspection system and method based on a large AI model to at least solve the above-mentioned shortcomings. Summary of the Invention

[0004] One of the purposes of the present invention is to provide a question-and-answer guided safety inspection system and method based on an AI large model, introduce natural language processing technology to process safety inspection related data, and use the deepseek-R1 large model to learn safety inspection inference rules to obtain a safety inspection guidance model. The model automatically recognizes the patrol personnel's voice to realize patrol interaction, prompts potential safety hazards in real time, and guides patrol personnel to quickly identify them, thereby reducing omissions and misjudgments in manual inspections.

[0005] An embodiment of the present invention provides a question-and-answer guided safety inspection system based on an AI large model, comprising:

[0006] The model training module is used to process safety inspection-related data through natural language processing technology and train the safety inspection guidance model based on the deepseek-R1 large model;

[0007] The inspection module is used to realize inspection interaction based on the security inspection guidance model and the inspection personnel's voice.

[0008] Preferably, the model training module processes the security inspection related data through natural language processing technology, and trains the security inspection guidance model based on the deepseek-R1 large model, including:

[0009] By using natural language processing technology, combined with expert scoring and industry standards, we can extract effective trigger words for safety inspections.

[0010] Based on the effective trigger words, obtain the safety inspection reasoning rules; the safety inspection reasoning rules include: description of the inspection phenomenon and the reasoned safety inspection items;

[0011] Based on the deepseek-R1 large model and according to the safety inspection inference rules, the safety inspection guidance model is trained.

[0012] Preferably, the inspection module implements inspection interaction based on the safety inspection guidance model and the inspection personnel's voice, including:

[0013] Determine the first speaker within the first area of the interactive drone;

[0014] Determine whether the first speaker has entered the inspection interaction process;

[0015] If so, the corresponding first speaker will be regarded as the second speaker;

[0016] If not, obtaining interaction triggering features of different interaction triggering feature types of the first speaker;

[0017] Among them, the interaction trigger features include:

[0018] The number of alternating views of the interactive drone and the inspection subject viewed by the first speaker is greater than or equal to the threshold of the number of alternating views corresponding to the viewing angle between the interactive drone and the inspection subject;

[0019] The rate at which the distance between the first speaker and the interactive drone decreases increases over time;

[0020] The semantics of the first speaker's speech contain at least one trigger word that indicates the start of interaction;

[0021] If at least two interaction trigger features of different interaction trigger feature types are obtained, the corresponding first speaker is used as the third speaker;

[0022] Control the interactive drone to conduct a question-and-answer guided safety inspection on the third speaker.

[0023] Preferably, the inspection module controls the interactive drone to conduct a question-and-answer guided safety inspection on the third speaker, including:

[0024] determining a first position of a second speaker and a second position of a third speaker;

[0025] Get the current listening position of the interactive drone;

[0026] Dynamically plan the planning points within the second area of the current listening position;

[0027] Calculating a first distance between the planned point and the first position, and simultaneously calculating a second distance between the planned point and the second position;

[0028] Summarize the first distance and the second distance to obtain a distance array;

[0029] If all distance values in the distance array are less than a preset first threshold, and the third distance between the current listening position and the planned point is less than a preset second threshold, calculate the standard deviation and the sum of the distance values in the corresponding distance array;

[0030] Assigning a first weight coefficient preset by the standard deviation to obtain a first count value, assigning a distance value and a preset second weight coefficient to obtain a second count value, summing the first count value and the second count value to obtain a target value, wherein the first weight coefficient is less than the second weight coefficient;

[0031] The planning point corresponding to the distance array with the smallest target value is used as the updated listening position.

[0032] Preferably, the inspection module controls the interactive drone to conduct a question-and-answer guided safety inspection on the third speaker, further comprising:

[0033] Based on the real-time questions and answers of the third speaker in the safety inspection guidance model, the first missing knowledge graph of the third speaker is constructed;

[0034] Retrieving graph knowledge of the first missing knowledge graph based on the first missing knowledge graph and the safety inspection guidance knowledge base;

[0035] After the third speaker views the graph knowledge, the security inspection guidance model outputs the second missing knowledge graph following the first missing knowledge graph;

[0036] Determine the overlapping graph of the first missing knowledge graph and the second missing knowledge graph;

[0037] Based on the historical questions and answers of the second speaker in the safety inspection guidance model, a standard knowledge graph of the second speaker is constructed;

[0038] Obtain the overlap relationship between the overlap graph and the standard knowledge graph, where the overlap relationship includes: the partial overlap between the standard knowledge graph and the overlap graph;

[0039] Get the connection relationship of the local overlapping parts;

[0040] Based on the connection relationship of the partially overlapping parts, an inspection knowledge practical auxiliary task is generated and sent to the second speaker corresponding to the standard knowledge graph to which the partially overlapping parts belong; the inspection knowledge practical auxiliary task includes: inspection knowledge practical auxiliary content and the position of the auxiliary personnel.

[0041] An embodiment of the present invention provides a question-and-answer guided security inspection method based on an AI large model, comprising:

[0042] Process safety inspection data using natural language processing technology, and train a safety inspection guidance model based on the deepseek-R1 large model;

[0043] Inspection interaction is achieved based on the safety inspection guidance model and the inspection personnel's voice.

[0044] Preferably, the security inspection related data is processed by natural language processing technology, and the security inspection guidance model is trained based on the deepseek-R1 large model, including:

[0045] By using natural language processing technology, combined with expert scoring and industry standards, we can extract effective trigger words for safety inspections.

[0046] Based on the effective trigger words, obtain the safety inspection reasoning rules; the safety inspection reasoning rules include: description of the inspection phenomenon and the reasoned safety inspection items;

[0047] Based on the deepseek-R1 large model and according to the safety inspection inference rules, the safety inspection guidance model is trained.

[0048] Preferably, the inspection interaction is realized according to the security inspection guidance model and the inspection personnel's voice, including:

[0049] Determine the first speaker within the first area of the interactive drone;

[0050] Determine whether the first speaker has entered the inspection interaction process;

[0051] If so, the corresponding first speaker will be regarded as the second speaker;

[0052] If not, obtaining interaction triggering features of different interaction triggering feature types of the first speaker;

[0053] Among them, the interaction trigger features include:

[0054] The number of alternating views of the interactive drone and the inspection subject viewed by the first speaker is greater than or equal to the threshold of the number of alternating views corresponding to the viewing angle between the interactive drone and the inspection subject;

[0055] The rate at which the distance between the first speaker and the interactive drone decreases increases over time;

[0056] The semantics of the first speaker's speech contain at least one trigger word that indicates the start of interaction;

[0057] If at least two interaction trigger features of different interaction trigger feature types are obtained, the corresponding first speaker is used as the third speaker;

[0058] Control the interactive drone to conduct a question-and-answer guided safety inspection on the third speaker.

[0059] Preferably, controlling the interactive drone to conduct a question-and-answer guided safety inspection on the third speaker includes:

[0060] determining a first position of a second speaker and a second position of a third speaker;

[0061] Get the current listening position of the interactive drone;

[0062] Dynamically plan the planning points within the second area of the current listening position;

[0063] Calculating a first distance between the planned point and the first position, and simultaneously calculating a second distance between the planned point and the second position;

[0064] Summarize the first distance and the second distance to obtain a distance array;

[0065] If all distance values in the distance array are less than a preset first threshold, and the third distance between the current listening position and the planned point is less than a preset second threshold, calculate the standard deviation and the sum of the distance values in the corresponding distance array;

[0066] Assigning a first weight coefficient preset by the standard deviation to obtain a first count value, assigning a distance value and a preset second weight coefficient to obtain a second count value, summing the first count value and the second count value to obtain a target value, wherein the first weight coefficient is less than the second weight coefficient;

[0067] The planning point corresponding to the distance array with the smallest target value is used as the updated listening position.

[0068] Preferably, controlling the interactive drone to conduct a question-and-answer guided safety inspection on the third speaker further includes:

[0069] Based on the real-time questions and answers of the third speaker in the safety inspection guidance model, the first missing knowledge graph of the third speaker is constructed;

[0070] Retrieving graph knowledge of the first missing knowledge graph based on the first missing knowledge graph and the safety inspection guidance knowledge base;

[0071] After the third speaker views the graph knowledge, the security inspection guidance model outputs the second missing knowledge graph following the first missing knowledge graph;

[0072] Determine the overlapping graph of the first missing knowledge graph and the second missing knowledge graph;

[0073] Based on the historical questions and answers of the second speaker in the safety inspection guidance model, a standard knowledge graph of the second speaker is constructed;

[0074] Obtain the overlap relationship between the overlap graph and the standard knowledge graph, where the overlap relationship includes: the partial overlap between the standard knowledge graph and the overlap graph;

[0075] Get the connection relationship of the local overlapping parts;

[0076] Based on the connection relationship of the partially overlapping parts, an inspection knowledge practical auxiliary task is generated and sent to the second speaker corresponding to the standard knowledge graph to which the partially overlapping parts belong; the inspection knowledge practical auxiliary task includes: inspection knowledge practical auxiliary content and the position of the auxiliary personnel.

[0077] The beneficial effects of the present invention are:

[0078] The present invention introduces natural language processing technology to process safety inspection related data, and uses the deepseek-R1 large model to learn safety inspection inference rules to obtain a safety inspection guidance model. The model automatically recognizes the voice of inspectors to realize inspection interaction, prompts potential safety hazards in real time and guides inspectors to quickly identify them, thereby reducing omissions and misjudgments in manual inspections.

[0079] Other features and advantages of the present invention will be described in the following description, and in part will become apparent from the description, or will be understood by practicing the present invention. The purpose and other advantages of the present invention can be achieved and obtained through the structures specifically pointed out in this application document.

[0080] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0082] Figure 1 Schematic diagram of a question-and-answer guided safety inspection system based on an AI big model in an embodiment of the present invention;

[0083] Figure 2 This is a schematic diagram of a question-and-answer guided safety inspection method based on an AI big model in an embodiment of the present invention. DETAILED DESCRIPTION

[0084] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0085] The embodiment of the present invention provides a question-answer guided safety inspection system based on AI big model, such as Figure 1 Shown, including:

[0086] Model training module 1 is used to process safety inspection related data through natural language processing technology and train the safety inspection guidance model based on the deepseek-R1 large model;

[0087] The model training module performs the following operations:

[0088] By using natural language processing technology, combined with expert scoring and industry standards, we can extract effective trigger words for safety inspections.

[0089] Based on the effective trigger words, obtain the safety inspection reasoning rules; the safety inspection reasoning rules include: description of the inspection phenomenon and the reasoned safety inspection items;

[0090] Based on the deepseek-R1 large model and the safety inspection reasoning rules, the safety inspection guidance model is trained;

[0091] Inspection module 2 is used to realize inspection interaction based on the safety inspection guidance model and the inspection personnel's voice.

[0092] The working principle and beneficial effects of the above technical solution are:

[0093] Safety inspection related data include: inspection object data, judgment standard data, legal and regulatory data, and inspection expert experience data. Taking the typical object of safety inspection in the chemical industry, "vacuum pump", as an example, the inspection object data at this time is the structure, characteristics, usage, emergency measures and other related data of the vacuum pump; the judgment standard data is the standard specifications, technical documents and other data related to the hidden danger judgment during the safety inspection of the vacuum pump; the legal and regulatory data is the laws, regulations, and rules related to vacuum pump problems and penalties; the inspection expert experience data is the vacuum pump inspection method and common hidden danger point judgment method provided by the experts. Combined with natural language processing technology, it can realize the safety inspection key points. Identification and mapping of the association relationship (safety inspection reasoning rules) between key factors (inspection phenomenon descriptions and their corresponding safety inspection items); after identifying the association relationship, the deepseek-R1 large model is used to learn the safety inspection reasoning rules to obtain the safety inspection guidance model. The safety inspection guidance model can automatically analyze the semantics of the inspectors based on the input voice signals. The inspector semantics contain information related to the inspection phenomena encountered by the inspectors. According to the relevant information of the inspection phenomena, the safety inspection items corresponding to the inspection phenomenon descriptions that match the inspection phenomena are obtained as the inspection intentions of the inspectors, and the preset standard inspection processes and inspection knowledge of the corresponding safety inspection items are automatically retrieved for the inspectors to review.

[0094] The present invention introduces natural language processing technology to process safety inspection related data, and uses the deepseek-R1 large model to learn safety inspection inference rules to obtain a safety inspection guidance model. The model automatically recognizes the voice of inspectors to realize inspection interaction, prompts potential safety hazards in real time and guides inspectors to quickly identify them, thereby reducing omissions and misjudgments in manual inspections.

[0095] In one embodiment, the inspection module implements inspection interaction based on the security inspection guidance model and the inspection personnel's voice, including:

[0096] Determine the first speaker within a first area of the interactive drone; the first area is the maximum distance range of sound that can be recognized by the built-in sound receiving device of the interactive drone, for example, 5 meters;

[0097] Determine whether the first speaker has entered the inspection interaction process; when determining whether to enter the inspection interaction process, read the conversation record of the model in the system to determine whether to access the interactive conversation of the corresponding first speaker;

[0098] If not, obtaining interaction trigger features of different interaction trigger feature types of the first speaker; interaction trigger feature types include: viewing action trigger, distance trigger, and trigger word trigger;

[0099] Among them, the interaction trigger features include:

[0100] The number of alternating views of the interactive drone and the inspection subject viewed by the first speaker is greater than or equal to the alternating view threshold corresponding to the viewing angle between the interactive drone and the inspection subject; the viewing angle between the interactive drone and the inspection subject is: the sight angle when the first speaker views the interactive drone and the inspection subject respectively;

[0101] The rate at which the distance between the first speaker and the interactive drone decreases increases over time. The rate at which the distance decreases over time is such that the speed at which the distance between the first speaker and the interactive drone decreases becomes faster and faster over time, indicating that the first speaker is tending to approach the interactive drone.

[0102] The semantics of the speech of the first speaker include at least one trigger word that indicates the start of interaction. The trigger word that indicates the start of interaction is, for example, "Please show me..." or "Please tell me...".

[0103] If at least two interaction triggering features of different interaction triggering feature types are obtained, the corresponding first speaker is used as the third speaker; at least two different interaction triggering feature types are obtained, and at least two types of interaction triggering feature types are constrained to mutually verify that the first speaker has an interaction tendency;

[0104] Get the first speaker who has already conducted the inspection interaction process and use him as the second speaker;

[0105] determining a first position of a second speaker and a second position of a third speaker;

[0106] Get the current listening position of the interactive drone; the current listening position is the position where the interactive drone and the first speaker are interacting in real-time voice;

[0107] Dynamically plan points within the second area of the current listening position; the second area is a spherical area within 3 meters from the current listening position; during planning, random planning is performed within the second area;

[0108] Calculating a first distance between the planned point and the first position, and simultaneously calculating a second distance between the planned point and the second position;

[0109] Summarize the first distance and the second distance to obtain a distance array;

[0110] If all distance values in the distance array are less than a preset first threshold, and the third distance between the current listening position and the planned point is less than a preset second threshold, the standard deviation and the sum of the distance values in the corresponding distance array are calculated; if all distance values are less than the preset first threshold (e.g., 5 meters), each first speaker in the interaction is constrained to be within the interaction range (receiving range) of the drone; the preset second threshold is, for example, 1 meter;

[0111] Assigning a first weight coefficient preset by the standard deviation to obtain a first count value, assigning a distance value and a preset second weight coefficient to obtain a second count value, summing the first count value and the second count value to obtain a target value, wherein the first weight coefficient is less than the second weight coefficient;

[0112] The planning point corresponding to the distance array with the smallest target value is used as the updated listening position.

[0113] The working principle and beneficial effects of the above technical solution are:

[0114] In the interactive drone deployment safety inspection guidance model, when interacting, the drone must first determine the person who needs to interact. Therefore, the first speaker within the longest distance range (first area range) that the interactive drone's built-in sound receiving device can recognize the sound is first determined. Not all first speakers have interaction needs, so they need to be further differentiated: first query the system to obtain the person who has already connected to the question-and-answer guidance (i.e., the second speaker), and then identify the interaction trigger features of the remaining first speakers. The interaction trigger feature types include three different trigger mechanisms: viewing action trigger, distance trigger, and trigger word trigger;

[0115] The interaction triggering feature of the standard viewing action trigger is: the sight angle between the first speaker and the interactive drone and the inspection subject is greater than the alternating viewing threshold corresponding to the angle, and the alternating viewing threshold corresponding to the viewing angle satisfies a negative correlation relationship. The specific relationship ratio can be set manually. The smaller the viewing angle, the lower the viewing cost. The more alternating viewing thresholds are set accordingly. Only when the trigger is triggered can it be indicated that the first speaker confirms whether the interactive drone recognizes his or her interaction request more frequently when the viewing cost is small. The alternating viewing threshold is dynamically set to improve the accuracy of viewing action triggering.

[0116] The distance-triggered interaction triggering feature is that the rate at which the distance between the first speaker and the interactive drone decreases increases over time. This forces the first speaker to move closer to the interactive drone faster, which better indicates their willingness to interact with the interactive drone.

[0117] The triggering characteristics of the interaction triggered by the trigger word are as follows: the speech semantics contain at least one trigger word that represents the start of the interaction, which constrains the first speaker to directly express the desire to interact;

[0118] When performing interaction trigger determination, if at least two interaction trigger feature types of different interaction trigger feature types are obtained and it is determined that the first speaker has an interaction tendency, at least two types of interaction trigger feature types are constrained to verify each other, thereby improving the accuracy of interaction trigger determination, and the corresponding first speaker is regarded as the third speaker;

[0119] After the third speaker is determined, there are two scenarios: the first is a one-to-one interaction between the interactive drone and the third speaker. In this case, a direct one-to-one interaction is sufficient. The second is a one-to-many interaction where the interactive drone already has a second speaker for Q&A guidance. The interactive drone's current listening position is no longer suitable for the newly added Q&A guide, so the current listening position needs to be updated.

[0120] During the update, dynamically plan the planned points within the second area of the current listening position, calculate the first distance between the planned point and the first position of the second speaker, calculate the second distance between the planned point and the second position of the third speaker, and sum up the first distance and the second distance to obtain a distance array;

[0121] The distance values in the restricted distance array are all less than a preset first threshold value, so as to constrain each first speaker in the interaction to be within the interaction range of the drone;

[0122] The third distance between the current listening position and the planned point is limited to less than the preset second threshold to prevent the listening position from changing too much. This prevents people who have already accessed the Q&A guidance from being unable to find the interactive drone in time when they initiate Q&A again.

[0123] The standard deviation represents the degree of difference in the distance between the connected Q&A guides and the interactive drones. The smaller the standard deviation, the more balanced the interactive drone takes care of each Q&A guide in interaction; the smaller the standard deviation, the more balanced the interactive drone takes care of each party when conducting multi-party interaction; the distance value and the sum represent the overall distance between the Q&A guides and the interactive drones. The first weight coefficient preset by the standard deviation is given to obtain the first count value, and the distance value and the preset second weight coefficient are given to obtain the second count value. The first count value and the second count value are summed to obtain the target value, and the planning point corresponding to the distance array with the smallest target value is selected as the updated listening position and the interactive drone is controlled to go there, thereby improving the collaborative Q&A guidance quality of the interactive drone.

[0124] In one embodiment, the inspection module controls the interactive drone to conduct a question-and-answer guided safety inspection on the third speaker, further comprising:

[0125] Based on the real-time questions and answers of the third speaker in the safety inspection guidance model, a first missing knowledge graph of the third speaker is constructed; the first missing knowledge graph is a personalized knowledge graph constructed based on the knowledge blind spots or deficiencies shown by the third speaker in the real-time questions and answers;

[0126] Based on the first missing knowledge graph and the safety inspection guidance knowledge base, the graph knowledge of the first missing knowledge graph is retrieved; the database of inspection-related knowledge stored in the safety inspection guidance knowledge base includes inspection processes, equipment information, safety specifications, etc.; the graph knowledge is related to the first missing knowledge graph and is obtained based on the matching of graph entities and entity relationships;

[0127] Obtain the second missing knowledge graph output by the security inspection guidance model immediately following the first missing knowledge graph after the third speaker views the graph knowledge; the second missing knowledge graph is: the knowledge graph immediately following the third speaker's second questioning after viewing the graph knowledge;

[0128] Determine an overlapping graph between the first missing knowledge graph and the second missing knowledge graph; the overlapping graph represents a knowledge graph of the inspection knowledge portion that the third speaker cannot understand;

[0129] Based on the historical questions and answers of the second speaker in the safety inspection guidance model, a standard knowledge graph of the second speaker is constructed; the standard knowledge graph is: the knowledge graph corresponding to the knowledge mastered by the second speaker determined based on historical questions and answers

[0130] Obtain the overlap relationship between the overlap graph and the standard knowledge graph, where the overlap relationship includes: the partial overlap between the standard knowledge graph and the overlap graph;

[0131] Obtaining the connection relationship of the local overlapping part; the connection relationship is: the atlas connection relationship of the local overlapping part in the overlapping atlas;

[0132] Based on the connection relationship of the partially overlapping parts, an inspection knowledge practical auxiliary task is generated and sent to the second speaker corresponding to the standard knowledge graph to which the partially overlapping parts belong; the inspection knowledge practical auxiliary task includes: inspection knowledge practical auxiliary content and the position of the auxiliary personnel.

[0133] The working principle and beneficial effects of the above technical solution are:

[0134] Under the guidance of questions and answers in the safety inspection guidance model, the third speaker will feedback knowledge that he is not familiar with. The model will deduce the first missing knowledge graph of the third speaker based on the questions and answers for the third speaker to review. After the third speaker reviews it, if the second missing knowledge graph output by the model immediately following the first missing knowledge graph still overlaps with the first missing knowledge graph, it means that the third speaker cannot understand the relevant knowledge of the overlapping graph in writing. The second speaker and the third speaker belong to the same inspection area, and their historical questions and answers are very likely to contain relevant knowledge of the overlapping graph, and their location is also near the third speaker. A standard knowledge graph is constructed based on historical questions and answers. The overlapping relationship between the overlapping graph and the standard knowledge graph is calculated. The partially overlapping part is the knowledge part that the second speaker can assist the third speaker in. However, the second speaker may have In multiple, their local overlapping parts may overlap or may not overlap (the connection relationship of the local overlapping parts is generated); based on the connection relationship of the local overlapping parts, the inspection knowledge practical auxiliary tasks are generated. Specifically, when the connection relationship is that the local overlapping parts do not overlap, the inspection knowledge practical auxiliary tasks are set for the corresponding second speakers respectively. When the connection relationship is that the local overlapping parts overlap, a communication link is established for the corresponding second speakers of the local overlapping parts. When the overlapping area is greater than the preset overlapping area threshold, the party with less important knowledge corresponding to the local overlapping part describes the practical knowledge part missing to the other party (the party with greater important knowledge corresponding to the local overlapping part), and the party with greater important knowledge corresponding to the local overlapping part receives the inspection knowledge practical auxiliary task to provide practical guidance to the third speaker, which greatly improves the efficiency of practical guidance.

[0135] In one embodiment, the inspection module generates an inspection knowledge practical auxiliary task based on the connection relationship of the partially overlapping parts, and sends it to the second speaker corresponding to the standard knowledge graph to which the partially overlapping parts belong, and further includes:

[0136] When the connection relationship is partially overlapping and partially non-overlapping, set inspection knowledge practical auxiliary tasks for the corresponding second speakers respectively;

[0137] When the connection relationship is partially overlapping, a communication link is established with the second speaker corresponding to the partially overlapping part. When the overlapping area is greater than a preset overlapping area threshold, the second speaker corresponding to the partially overlapping part, whose knowledge importance is less, describes the practical knowledge missing from the second speaker with greater knowledge importance to the second speaker corresponding to the partially overlapping part, and sets an inspection knowledge practical auxiliary task for the second speaker with greater knowledge importance. The preset overlapping area threshold is manually preset, for example, 50%;

[0138] Obtain the auxiliary knowledge corresponding to the executors of the inspection knowledge practical auxiliary tasks and divide them into graph blocks on the overlapping graph;

[0139] Obtain a standard graph block traversal sequence; the standard graph block traversal sequence includes multiple ordered graph blocks, and the order of the graph blocks is a reasonable understanding order on the overlapping graph, which is determined according to the graph structure;

[0140] Traverse the graph blocks in the standard graph block traversal sequence in sequence. When the nth graph block is traversed, dispatch the executive personnel corresponding to the nth graph block to the third speaker for inspection knowledge practical assistance. At the same time, the executive personnel corresponding to the n+1th to Nth graph blocks are used as preparatory executive personnel. n is a positive integer, and N is the total number of graph blocks.

[0141] Obtain the preparatory level of the inspection knowledge practical auxiliary task of the preparatory execution personnel; the preparatory level is: starting from the n+1th atlas block, the count value increases by 1 each time, and the preparatory level and the count value are equal in value;

[0142] Compare the preparatory auxiliary behavior feature sequence of the preparatory execution personnel whose preparatory level is less than or equal to the preset preparatory level threshold and the standard preparatory auxiliary behavior feature sequence of the preparatory level of the inspection knowledge practical auxiliary task corresponding to the corresponding preparatory execution personnel; the preset preparatory level threshold is set manually, for example: 5; the preparatory auxiliary behavior feature sequence is: the preparatory auxiliary behavior features of the preparatory executors whose preparatory level is less than or equal to the preset preparatory level threshold (for example: temporarily doing inspection work at hand, downloading inspection case videos for practical assistance, going to the third spokesperson, etc.) are sorted in order of the corresponding preparatory level from small to large; when determining the standard preparatory auxiliary behavior feature sequence of the preparatory level of the inspection knowledge practical auxiliary task corresponding to the corresponding preparatory execution personnel, the preparatory auxiliary behavior features of each inspection knowledge practical auxiliary task are different when corresponding to different preparatory levels, and the preparatory auxiliary behavior features of each inspection knowledge practical auxiliary task corresponding to its corresponding preparatory level are determined by looking up the table, and then sorting them in order of the preparatory level from small to large to obtain the standard preparatory auxiliary behavior feature sequence;

[0143] If there are differences in the characteristics of the preparatory auxiliary behaviors, the corresponding preparatory execution personnel will be reminded.

[0144] The working principle and beneficial effects of the above technical solution are:

[0145] The present invention determines the inspection knowledge practical operation auxiliary tasks of the second speaker respectively according to the different connection relationships. If there is a partial overlap, the second speaker whose knowledge importance is less corresponding to the partial overlap will describe the practical operation knowledge missing of the second speaker whose knowledge importance is greater to the second speaker whose knowledge importance is greater corresponding to the partial overlap, and the inspection knowledge practical operation auxiliary tasks are set for the second speaker whose knowledge importance is greater, thereby improving the rationality of setting the inspection knowledge practical operation auxiliary tasks.

[0146] According to the order of understanding of the map blocks on the overlapping map, the executive personnel corresponding to the corresponding map blocks are dispatched in sequence to the third speaker to provide inspection knowledge practical assistance. At the same time, for the executive personnel who are about to assist, their preparatory assistance behavior characteristics and the standard preparatory assistance behavior characteristics corresponding to the preparatory level of their inspection knowledge practical assistance tasks are matched one by one. According to the different inspection knowledge practical assistance tasks and preparatory levels, the preparatory assistance behaviors of the preparatory executive personnel are timely and dynamically monitored, thereby improving the timeliness of assistance practice.

[0147] The embodiment of the present invention provides a question-answer guided safety inspection method based on AI big model, such as Figure 2 As shown, including:

[0148] Step 1: Process the safety inspection data using natural language processing technology and train the safety inspection guidance model based on the DeepSeek-R1 large model;

[0149] Step 2: Implement inspection interaction based on the security inspection guidance model and the inspection personnel's voice.

[0150] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A question-answer guided safety inspection system based on AI big model, characterized by: include: The model training module is used to process safety inspection-related data through natural language processing technology and train the safety inspection guidance model based on the deepseek-R1 large model; Inspection module, used to implement inspection interaction based on the security inspection guidance model and the inspection personnel's voice; The inspection module implements inspection interaction based on the security inspection guidance model and the inspection personnel's voice, including: Determine the first speaker within the first area of the interactive drone; Determine whether the first speaker has entered the inspection interaction process; If yes, the first speaker is treated as the second speaker, and the interactive drone is controlled to continue the Q&A guided safety inspection of the second speaker. If not, obtaining interaction triggering features of different interaction triggering feature types of the first speaker; Among them, the interaction trigger features include: The number of alternating views of the interactive drone and the inspection subject viewed by the first speaker is greater than or equal to the threshold of the number of alternating views corresponding to the viewing angle between the interactive drone and the inspection subject; The rate at which the distance between the first speaker and the interactive drone decreases increases over time; The semantics of the first speaker's speech contain at least one trigger word that indicates the start of interaction; If at least two interaction trigger features of different interaction trigger feature types are obtained, the corresponding first speaker is used as the third speaker; Control the interactive drone to conduct a Q&A-guided safety inspection of the third speaker; Among them, the inspection module controls the interactive drone to conduct a question-and-answer guided safety inspection of the third speaker, including: determining a first position of a second speaker and a second position of a third speaker; Get the current listening position of the interactive drone; Dynamically plan the planning points within the second area of the current listening position; Calculating a first distance between the planned point and the first position, and simultaneously calculating a second distance between the planned point and the second position; Summarize the first distance and the second distance to obtain a distance array; If all distance values in the distance array are less than a preset first threshold, and the third distance between the current listening position and the planned point is less than a preset second threshold, calculate the standard deviation and the sum of the distance values in the corresponding distance array; Assigning a first weight coefficient preset by the standard deviation to obtain a first count value, assigning a distance value and a preset second weight coefficient to obtain a second count value, summing the first count value and the second count value to obtain a target value, wherein the first weight coefficient is less than the second weight coefficient; The planning point corresponding to the distance array with the smallest target value is used as the updated listening position.

2. The AI large model-based question-answer guided safety inspection system according to claim 1, characterized in that: The model training module processes security inspection-related data through natural language processing technology and trains a security inspection guidance model based on the DeepSeek-R1 large model, including: By using natural language processing technology, combined with expert scoring and industry standards, we can extract effective trigger words for safety inspections. Based on the effective trigger words, obtain the safety inspection reasoning rules; the safety inspection reasoning rules include: description of the inspection phenomenon and the reasoned safety inspection items; Based on the deepseek-R1 large model and according to the safety inspection inference rules, the safety inspection guidance model is trained.

3. The AI large model-based question-answer guided safety inspection system according to claim 1, characterized in that: The inspection module controls the interactive drone to conduct a Q&A-guided safety inspection of the third speaker, and also includes: Based on the real-time questions and answers of the third speaker in the safety inspection guidance model, the first missing knowledge graph of the third speaker is constructed; Retrieving graph knowledge of the first missing knowledge graph based on the first missing knowledge graph and the safety inspection guidance knowledge base; After the third speaker views the graph knowledge, the security inspection guidance model outputs the second missing knowledge graph following the first missing knowledge graph; Determine the overlapping graph of the first missing knowledge graph and the second missing knowledge graph; Based on the historical questions and answers of the second speaker in the safety inspection guidance model, a standard knowledge graph of the second speaker is constructed; Obtain the overlap relationship between the overlap graph and the standard knowledge graph, where the overlap relationship includes: the partial overlap between the standard knowledge graph and the overlap graph; Get the connection relationship of the local overlapping parts; Based on the connection relationship of the partially overlapping parts, an inspection knowledge practical auxiliary task is generated and sent to the second speaker corresponding to the standard knowledge graph to which the partially overlapping parts belong; the inspection knowledge practical auxiliary task includes: inspection knowledge practical auxiliary content and the position of the auxiliary personnel.

4. A question-answer guided safety inspection method based on AI big model, characterized by: include: Process safety inspection data using natural language processing technology, and train a safety inspection guidance model based on the deepseek-R1 large model; Realize inspection interaction based on the safety inspection guidance model and the inspection personnel's voice; Inspection interaction is achieved based on the security inspection guidance model and the inspection personnel's voice, including: Determine the first speaker within the first area of the interactive drone; Determine whether the first speaker has entered the inspection interaction process; If yes, the first speaker is treated as the second speaker, and the interactive drone is controlled to continue the Q&A guided safety inspection of the second speaker. If not, obtaining interaction triggering features of different interaction triggering feature types of the first speaker; Among them, the interaction trigger features include: The number of alternating views of the interactive drone and the inspection subject viewed by the first speaker is greater than or equal to the threshold of the number of alternating views corresponding to the viewing angle between the interactive drone and the inspection subject; The rate at which the distance between the first speaker and the interactive drone decreases increases over time; The semantics of the first speaker's speech contain at least one trigger word that indicates the start of interaction; If at least two interaction trigger features of different interaction trigger feature types are obtained, the corresponding first speaker is used as the third speaker; Control the interactive drone to conduct a Q&A-guided safety inspection of the third speaker; Among them, the interactive drone is controlled to conduct a Q&A-guided safety inspection on the third speaker, including: determining a first position of a second speaker and a second position of a third speaker; Get the current listening position of the interactive drone; Dynamically plan the planning points within the second area of the current listening position; Calculating a first distance between the planned point and the first position, and simultaneously calculating a second distance between the planned point and the second position; Summarize the first distance and the second distance to obtain a distance array; If all distance values in the distance array are less than a preset first threshold, and the third distance between the current listening position and the planned point is less than a preset second threshold, calculate the standard deviation and the sum of the distance values in the corresponding distance array; Assigning a first weight coefficient preset by the standard deviation to obtain a first count value, assigning a distance value and a preset second weight coefficient to obtain a second count value, summing the first count value and the second count value to obtain a target value, wherein the first weight coefficient is less than the second weight coefficient; The planning point corresponding to the distance array with the smallest target value is used as the updated listening position.

5. The question-answer guided safety inspection method based on AI big model as claimed in claim 4, characterized in that: We use natural language processing technology to process safety inspection data and train a safety inspection guidance model based on the DeepSeek-R1 large model, including: By using natural language processing technology, combined with expert scoring and industry standards, we can extract effective trigger words for safety inspections. Based on the effective trigger words, obtain the safety inspection reasoning rules; the safety inspection reasoning rules include: description of the inspection phenomenon and the reasoned safety inspection items; Based on the deepseek-R1 large model and according to the safety inspection inference rules, the safety inspection guidance model is trained.

6. The question-answer guided safety inspection method based on AI big model according to claim 4, characterized in that: Control the interactive drone to conduct a Q&A guided safety inspection of the third speaker, which also includes: Based on the real-time questions and answers of the third speaker in the safety inspection guidance model, the first missing knowledge graph of the third speaker is constructed; Retrieving graph knowledge of the first missing knowledge graph based on the first missing knowledge graph and the safety inspection guidance knowledge base; After the third speaker views the graph knowledge, the security inspection guidance model outputs the second missing knowledge graph following the first missing knowledge graph; Determine the overlapping graph of the first missing knowledge graph and the second missing knowledge graph; Based on the historical questions and answers of the second speaker in the safety inspection guidance model, a standard knowledge graph of the second speaker is constructed; Obtain the overlap relationship between the overlap graph and the standard knowledge graph, where the overlap relationship includes: the partial overlap between the standard knowledge graph and the overlap graph; Get the connection relationship of the local overlapping parts; Based on the connection relationship of the partially overlapping parts, an inspection knowledge practical auxiliary task is generated and sent to the second speaker corresponding to the standard knowledge graph to which the partially overlapping parts belong; the inspection knowledge practical auxiliary task includes: inspection knowledge practical auxiliary content and the position of the auxiliary personnel.

Images