A Dialogue Interaction Method and System for Building Equipment Operation and Maintenance Based on AIGC
By acquiring historical conversation texts during building equipment operation and maintenance, performing feature analysis and model generation, and combining this with knowledge base verification, the problem of inaccurate answers in existing technologies is solved, achieving more efficient and professional operation and maintenance guidance.
Patent Information
- Application Number
- CN202510915484.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-03
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2045-07-03
AI Technical Summary
In the operation and maintenance of building equipment, existing technologies, such as fixed templates and simple keyword matching, cannot provide accurate and targeted answers, and are difficult to meet the requirements of professionalism and accuracy.
By acquiring a collection of historical Q&A conversation texts from building equipment operation and maintenance scenarios, conversation features are parsed to generate semantic recognition features and operation and maintenance requirement association features. A pre-trained operation and maintenance conversation generation model is invoked, and compliance verification and semantic adaptation are performed in conjunction with a knowledge base in the building equipment operation and maintenance domain to generate target operation and maintenance guidance text.
It improves the accuracy and relevance of question-and-answer matching during the operation and maintenance of building equipment, and provides more efficient and high-quality operation and maintenance guidance services.
Smart Images

Figure CN120911593B_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of artificial intelligence technology, specifically relating to a method and system for interactive conversations in building equipment operation and maintenance based on AIGC. Background Technology
[0002] In the context of Artificial Intelligence Generated Content (AIGC), questions raised by users in the actual application of building equipment operation and maintenance are usually responded to using fixed templates or simple keyword matching. This approach involves finding a matching template based on a pre-set common question template after receiving a user's question and providing a corresponding answer; or simply selecting an answer containing the keyword from a pre-stored answer database based on the keyword in the question.
[0003] However, this technology has obvious problems: on the one hand, fixed templates cannot cover all possible user questions, and it is difficult to provide accurate and targeted answers to some more complex or personalized questions; on the other hand, simple keyword matching easily ignores the core semantics of the question and its relevance to the operation and maintenance tasks, resulting in answers that may be out of touch with actual operation and maintenance needs and fail to meet the requirements of professionalism and accuracy in building equipment operation and maintenance scenarios. Summary of the Invention
[0004] This application provides a method and system for interactive dialogue in building equipment operation and maintenance based on AIGC, which improves the accuracy and relevance of question-and-answer matching during equipment operation and maintenance interaction.
[0005] In a first aspect, embodiments of this application provide a building equipment operation and maintenance session interaction method based on AIGC, applied to a building equipment operation and maintenance session interaction system, the method comprising: Obtain a set of historical Q&A conversation texts in the operation and maintenance scenario of building equipment. The set of historical Q&A conversation texts includes multiple consecutive operation and maintenance Q&A dialogue units. Each operation and maintenance Q&A dialogue unit contains user question statements and corresponding operation and maintenance response statements. The historical question-and-answer dialogue text set is subjected to dialogue feature parsing processing to obtain the semantic recognition features and operation and maintenance requirement association features of each operation and maintenance question-and-answer dialogue unit. The semantic recognition features reflect the core semantic orientation of the user's question statement, and the operation and maintenance requirement association features reflect the degree of association between the user's question statement and the operation and maintenance tasks of building equipment. The pre-trained operation and maintenance conversation generation model is invoked to perform joint semantic generation processing on the semantic recognition features and the operation and maintenance requirement association features to generate a set of candidate operation and maintenance guidance texts corresponding to the operation and maintenance question and answer dialogue unit. Based on the knowledge base of building equipment operation and maintenance, the candidate operation and maintenance guidance text set is subjected to compliance verification and semantic adaptation processing to generate the target operation and maintenance guidance text. The target operation and maintenance guidance text is then embedded into the current operation and maintenance session interaction process to complete the session response operation.
[0006] Secondly, embodiments of this application provide a building equipment operation and maintenance session interaction system, which includes a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the above-described method.
[0007] Thirdly, embodiments of this application provide a computer-readable storage medium including a computer program. When the computer program is run on a building equipment operation and maintenance session interaction system, the computer program is used to cause the building equipment operation and maintenance session interaction system to perform the steps of the above-described method.
[0008] This application embodiment obtains a historical Q&A conversation text set of building equipment operation and maintenance scenarios, parses conversation features to obtain semantic recognition features and operation and maintenance requirement association features, jointly calls a pre-trained operation and maintenance conversation generation model to generate a candidate operation and maintenance guidance text set, and then performs compliance verification and semantic adaptation processing based on a building equipment operation and maintenance domain knowledge base to obtain the target operation and maintenance guidance text and embed it into the current operation and maintenance conversation interaction process. This allows for accurate identification of the core issues and operation and maintenance requirements of users in building equipment operation and maintenance scenarios, generating operation and maintenance guidance text that fits actual needs and conforms to domain standards, effectively improving the accuracy, professionalism, and compliance of building equipment operation and maintenance conversation interactions, and providing users with more efficient and high-quality operation and maintenance guidance services. Attached Figure Description
[0009] Figure 1 This is a flowchart illustrating a building equipment operation and maintenance session interaction method based on AIGC, provided in an embodiment of this application.
[0010] Figure 2 This is a schematic diagram of the structure of a building equipment operation and maintenance interactive system provided in an embodiment of this application. Detailed Implementation
[0011] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this application. Obviously, the described embodiments are only some embodiments of the technical solutions of this application, and not all embodiments. Based on the embodiments recorded in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the technical solutions of this application.
[0012] See Figure 1This is a building equipment operation and maintenance session interaction method based on AIGC provided in the embodiments of this application. This method can be applied to the building equipment operation and maintenance session interaction system. The specific process is as follows: steps 110-140.
[0013] Step 110: Obtain the historical Q&A conversation text set in the building equipment operation and maintenance scenario. The historical Q&A conversation text set includes multiple consecutive operation and maintenance Q&A dialogue units. Each operation and maintenance Q&A dialogue unit contains user question statements and corresponding operation and maintenance response statements.
[0014] In this embodiment of the application, regarding the acquisition of historical question-and-answer conversation text sets, firstly, relevant data can be extracted from the building equipment monitoring system. These monitoring systems record the communication information between users and maintenance personnel. For example, in a large commercial complex, for the maintenance of the air conditioning system, users can report the problem of poor air conditioning cooling effect through the interactive interface of the monitoring system, and the maintenance personnel will reply with corresponding solutions in the system. These question-and-answer records will be saved.
[0015] Secondly, data can also be obtained from the maintenance work order system. When a user submits a maintenance work order, they will describe the problem with the device in detail. Maintenance personnel will provide corresponding responses during the work order processing. These work order records constitute part of the historical Q&A conversation text collection.
[0016] In addition, online customer service platforms are also an important source of data. When users encounter maintenance issues with building equipment, they consult online customer service, and customer service personnel answer their questions. These dialogues (with the user's knowledge and authorization, and the processing flow of the dialogue can be queried and withdrawn at any time for that user) can also be collected into a historical Q&A conversation text set. The collected data is organized in chronological order to form multiple consecutive maintenance Q&A dialogue units, each containing the user's question and the corresponding maintenance response.
[0017] Step 120: Perform conversation feature parsing on the historical question-and-answer conversation text set to obtain the semantic recognition features and operation and maintenance requirement association features of each operation and maintenance question-and-answer dialogue unit. The semantic recognition features reflect the core semantic orientation of the user's question statement, and the operation and maintenance requirement association features reflect the degree of association between the user's question statement and the operation and maintenance tasks of building equipment.
[0018] The purpose of parsing historical Q&A conversation text sets is to extract valuable information from large amounts of text data, namely semantic recognition features and operation and maintenance (O&M) requirement association features. In practice, each O&M Q&A dialogue unit is first pre-screened and pre-processed to remove invalid information, such as repetitive expressions and irrelevant symbols. Then, natural language processing (NLP) techniques are used to analyze user questions in depth. For semantic recognition feature extraction, semantic understanding algorithms are used to determine the core semantic meaning of user questions. For example, when a user asks, "The elevator in the building has been frequently stuck lately, what's going on?", the semantic understanding algorithm will analyze that the core semantic meaning is the cause of the elevator's malfunction. For O&M requirement association feature extraction, the classification system of building equipment O&M tasks is used to analyze the degree of correlation between user questions and various O&M tasks. If the user's question involves equipment troubleshooting and repair, it can be determined that the question has a high degree of correlation with the corresponding O&M task.
[0019] As one implementation, the historical question-and-answer dialogue text set is subjected to dialogue feature parsing processing to obtain the semantic recognition features and operation and maintenance requirement association features of each operation and maintenance question-and-answer dialogue unit, including: Step 1211: Perform word segmentation on the user's question statements in the operation and maintenance Q&A dialogue unit to obtain a set of conversational lexical units containing operation and maintenance professional terms and general descriptive vocabulary.
[0020] When segmenting user queries, specialized word segmentation algorithms are used. Taking the previously mentioned example, "The elevator in the building has been frequently jamming lately, what's going on?", the algorithm breaks this sentence down into individual word units. "Elevator" is a technical term used in operations and maintenance, while "building," "recently," "frequently," and "jamming" are general descriptive terms. In practice, the algorithm segments words according to semantic and grammatical rules, ensuring that the segmented word units accurately reflect the meaning of the original sentence. To improve segmentation accuracy, a pre-built dictionary is used, containing a large number of operational and maintenance terms and common general vocabulary. When the algorithm encounters uncommon words, it consults the dictionary for judgment, resulting in a more accurate set of conversational word units.
[0021] Step 1212: Call the pre-built semantic pointing analysis model to perform context association analysis on the conversation lexical set, and extract semantic keywords related to building equipment type, fault phenomenon and operation requirements in the user's question statement as core semantic pointing identifiers.
[0022] The pre-built semantic targeting analysis model is trained on a large amount of historical data. This model can perform contextual analysis on the set of conversational lexical units to extract core semantic targeting identifiers. When processing the example "The elevator in the building has been frequently jamming lately, what's going on?", the model first identifies "elevator" as a type of building equipment and "jamming" as a malfunction. Since the user is asking for a reason, it can be inferred that the operational need is likely troubleshooting. The model analyzes the semantic relationships between lexical units to determine the relevance of these keywords to the core semantic meaning. For example, while words like "recently" and "frequently" are general descriptive terms, when combined with "jamming," they more accurately describe the occurrence of the malfunction, thus helping the model determine the core semantic targeting. During the model's training process, parameters are continuously adjusted to improve the accuracy of identifying different types of questions.
[0023] Step 1213: The core semantic pointer is input to the semantic dimension mapping branch to generate a multi-dimensional semantic vector containing the device type dimension, fault type dimension, and operation requirement dimension as semantic recognition features.
[0024] In this embodiment, the semantic dimension mapping branch can convert core semantic identifiers into multi-dimensional semantic vectors. When core semantic identifiers such as "elevator," "stuttering," and "troubleshooting" are input into this branch, they are mapped in the device type dimension, fault type dimension, and operational requirement dimension, respectively. In the device type dimension, "elevator" is mapped to the corresponding category; in the fault type dimension, "stuttering" is assigned a corresponding feature value; and in the operational requirement dimension, "troubleshooting" also has a corresponding representation. Through this mapping method, the generated multi-dimensional semantic vector can comprehensively and accurately reflect the semantic recognition features of the user's query. In actual implementation, the semantic dimension mapping branch searches a predefined mapping table based on the input core semantic identifiers, converts the identifiers into vector representations, and then concatenates and integrates them to finally obtain the multi-dimensional semantic vector.
[0025] Step 1214: Perform operation and maintenance task correlation analysis on the user's question statement to identify the various operation and maintenance task intentions implied in the user's question statement.
[0026] When performing correlation analysis on maintenance tasks, a pre-defined classification system for building equipment maintenance tasks is used. For the question, "The building's elevators have been frequently jamming lately, what's going on?", the implicit maintenance task intent is analyzed. Since the user mentioned elevator jamming, the implicit maintenance task intent may include troubleshooting, fault diagnosis, and repair. During the analysis, factors such as the user's tone and word choice are considered. If the user's question is urgent, a faster troubleshooting and diagnosis may be preferred; if the user's description is more detailed, there may also be a need for subsequent maintenance operations. Simultaneously, the handling methods for similar questions in historical data are referenced to further determine the implicit maintenance task intent.
[0027] Step 1215: Based on the matching results between the various types of operation and maintenance task intentions and the preset building equipment operation and maintenance task classification system, generate a correlation score vector that reflects the relevance between the user's question and the operation and maintenance task as an operation and maintenance requirement correlation feature.
[0028] The pre-defined building equipment maintenance task classification system in this embodiment contains detailed information on various maintenance tasks. When matching multiple maintenance task intentions with this classification system, the degree of matching between each intention and the corresponding task in the classification system is calculated. For intentions such as "troubleshooting," "fault diagnosis," and "repair," the corresponding task categories in the classification system are found, and the similarity of the matching is evaluated. Based on the matching results, a relevance score is assigned to each intention. For example, if "troubleshooting" highly matches a certain task category in the classification system, a higher score is given; if the matching degree is low, a lower score is given. Finally, these scores are combined into a relevance score vector, which accurately reflects the relevance between the user's question and the maintenance task, serving as a feature for the association of maintenance needs.
[0029] In another implementation, the historical question-and-answer dialogue text set is subjected to dialogue feature parsing to obtain the semantic recognition features and operation and maintenance requirement association features of each operation and maintenance question-and-answer dialogue unit, including: Step 1221: Perform dialogue round alignment processing on the user question statements and corresponding operation and maintenance response statements in the operation and maintenance Q&A dialogue unit to generate the semantic connection relationship between user questions and historical responses.
[0030] During dialogue round alignment, user questions and corresponding maintenance responses are organized according to the chronological order of the dialogue. For example, in a maintenance Q&A session with multiple dialogue rounds, in the first round, the user asks, "The elevator is making abnormal noises," and the maintenance personnel respond, "Check the elevator's mechanical components first." In the second round, the user asks, "I checked, the mechanical components are fine, what should I do next?" Dialogue round alignment clearly shows the semantic connection between user questions and historical responses. In this example, the second round's question is based on the first round's response; that is, after checking the mechanical components as suggested by the maintenance personnel, the user found the problem persisted and thus inquired about the next steps.
[0031] Step 1222: Extract the keyword repetition rate and semantic topic consistency parameter from the semantic continuation relationship as conversation coherence features.
[0032] For the semantic connection relationships in the above example, keyword repetition rate and semantic topic consistency parameters are extracted. Regarding keyword repetition rate, the keyword "elevator" appears repeatedly in the user's two questions, indicating that this keyword plays an important role in semantic connection. Regarding the semantic topic consistency parameter, both questions revolve around elevator troubleshooting, showing a high degree of consistency. In practice, these parameters are calculated through keyword statistics and semantic analysis. For example, the frequency of the same keyword appearing in different rounds of dialogue is counted as the basis for keyword repetition rate; semantic understanding algorithms are used to determine whether the dialogue topics are consistent, thereby determining the semantic topic consistency parameter.
[0033] Step 1223: The conversation coherence features are fused with the core semantic orientation of the user's question to generate semantic recognition features that include semantic depth and contextual relevance.
[0034] For example, the fusion process comprehensively considers conversational coherence features and core semantic focus. Taking the example of "the elevator is making abnormal noises," the core semantic focus is the problem of the elevator making abnormal noises, and the conversational coherence features reflect the continuity between the user's question and historical responses. During fusion, based on these two aspects, the semantic depth and contextual relevance are further explored. For instance, the high coherence of the dialogue indicates that the user's description of the problem is progressively more in-depth, facilitating a more accurate understanding of the core semantics. Through fusion processing, the generated semantic recognition features can more comprehensively reflect the meaning of the user's question, including its contextual relevance and semantic depth.
[0035] Step 1224: Perform task scenario adaptation analysis on the operation and maintenance requirement association features to identify the matching relationship between the building equipment type corresponding to the user's question and the target operation and maintenance scenario.
[0036] When performing task scenario adaptation analysis, the specific target operation and maintenance scenario is considered. For example, if the target operation and maintenance scenario is elevator operation and maintenance in a commercial building, and a user asks, "The elevators in my building have been frequently stuck lately, what's going on?", the matching relationship between the building equipment type "elevator" and this scenario will be analyzed. Since the question explicitly mentions the building's elevators, it highly matches the elevator operation and maintenance scenario of a commercial building. In practice, the building equipment type will be evaluated based on the characteristics and requirements of the target operation and maintenance scenario to determine if it meets the scenario's needs. If the building equipment type does not match the scenario, further adjustments to the operation and maintenance strategy may be necessary.
[0037] Step 1225: Generate operation and maintenance requirement association features containing scenario adaptation tags and task priority identifiers based on the matching relationship.
[0038] Based on the previously identified matching relationships, maintenance requirement association features containing scenario adaptation tags and task priority identifiers will be generated. If the building equipment type highly matches the target maintenance scenario, an "adapted" scenario adaptation tag will be added. For task priority identifiers, since elevator malfunctions may affect the normal use of the building, the task priority will be set to a higher level. In the actual generation process, these identifiers will be determined according to preset rules and standards, combined with the matching relationships. For example, if the matching relationship is only partially adapted, the scenario adaptation tags and task priority identifiers can be adjusted according to the specific situation.
[0039] Step 130: Call the pre-trained operation and maintenance conversation generation model to perform joint semantic generation processing on the semantic recognition features and the operation and maintenance requirement association features to generate a set of candidate operation and maintenance guidance texts corresponding to the operation and maintenance question and answer dialogue unit.
[0040] In this embodiment, the pre-trained operation and maintenance (O&M) conversation generation model is trained on a large amount of historical data and is capable of joint semantic generation processing of semantic recognition features and O&M requirement-related features. After inputting semantic recognition features and O&M requirement-related features, the model processes these features through its internal neural network structure. Taking the previous elevator lag issue as an example, the model combines semantic recognition features such as "elevator," "lag," and "troubleshooting" with corresponding O&M requirement-related features to generate a set of candidate O&M guidance texts corresponding to the O&M question-and-answer dialogue unit. Based on the input features, the model predicts possible O&M guidance content and generates multiple candidate texts.
[0041] In a preferred embodiment, the invocation of the pre-trained operation and maintenance conversation generation model performs joint semantic generation processing on the semantic recognition features and the operation and maintenance requirement-related features to generate a set of candidate operation and maintenance guidance texts corresponding to the operation and maintenance question-and-answer dialogue unit, including: Step 131: Input the semantic recognition features and the operation and maintenance requirement association features into the feature fusion layer of the operation and maintenance session generation model, and establish the association mapping relationship between the semantic dimension and the requirement dimension through the feature cross attention mechanism to generate a fused feature vector.
[0042] The feature fusion layer in this embodiment can fuse semantic recognition features and operation and maintenance requirement-related features through a feature cross-attention mechanism. After these features are input, the feature cross-attention mechanism maps the semantic and requirement dimensions together. Taking the elevator jamming problem as an example, the information such as "elevator" and "jamming" in the semantic recognition features and the "troubleshooting" requirement in the operation and maintenance requirement-related features will be associated under the action of the feature cross-attention mechanism. The mechanism calculates the correlation between different features to determine their weights in the fusion process. In this way, the generated fused feature vector can comprehensively reflect information from both semantic and requirement aspects.
[0043] Step 132: Perform sequence decoding processing on the fused feature vector through the text generation layer of the operation and maintenance session generation model to generate initial guidance response text.
[0044] The text generation layer is built on a neural network architecture and performs sequence decoding on the fused feature vectors. During processing, the input fused feature vectors flow between the layers of the text generation layer. First, the embedding layer transforms the fused feature vectors into a representation suitable for subsequent processing. Then, in a recurrent neural network or transformer network, these representations are processed to gradually generate a text sequence. During generation, appropriate words are selected based on the model's predicted probabilities, ultimately forming the initial guidance response text. For example, in the case of an elevator jamming problem, the generated initial guidance response text might be "It is recommended to check if there is a fault in the elevator's electrical system."
[0045] Specifically, the step of performing sequence decoding processing on the fused feature vector through the text generation layer of the operation and maintenance session generation model to generate initial guidance response text includes: Step 1321: Input the fused feature vector into the position encoding branch of the text generation layer to add position information encoding that reflects the session turn order to the fused feature vector.
[0046] Optionally, the positional encoding branch adds positional information encoding to the fused feature vector. In practice, each fused feature vector corresponds to a conversation round. Taking the elevator lag problem as an example, if the problem is raised in the second conversation round, the positional encoding branch will add positional information encoding representing the second round to the corresponding fused feature vector. The purpose of this is to enable the model to better understand the order and context of the text, thereby generating more accurate response text.
[0047] Step 1322: Perform global dependency modeling on the fused feature vector with positional information encoding through the multi-head self-attention branch of the text generation layer to generate an intermediate feature representation containing contextual semantic associations.
[0048] Furthermore, multi-head self-attention branches can model global dependencies in fused feature vectors encoded with location information. During processing, the multi-head self-attention branch calculates attention weights between different features to determine their degree of association. Taking the elevator jamming problem as an example, the multi-head self-attention branch considers the relationships between features such as "elevator," "jamming," and "troubleshooting," as well as their association with location information encoding. In this way, global dependencies in the text can be captured, generating intermediate feature representations that include contextual semantic relationships.
[0049] Step 1323: Perform nonlinear transformation on the intermediate feature representation through the feedforward neural network branch of the text generation layer to generate a decoded feature vector containing the probability distribution of words.
[0050] The feedforward neural network branch performs a non-linear transformation on the intermediate feature representations. In this process, the feedforward neural network branch calculates and transforms the intermediate feature representations to generate a decoded feature vector containing the probability distribution of words. Taking the elevator jamming problem as an example, this vector contains the probability of different words appearing, such as the probability of words like "electrical system," "inspection," and "fault." These probabilities reflect the likelihood that the model will select these words when generating the response text.
[0051] Step 1324: Perform autoregressive decoding based on the decoded feature vector to generate a word-by-word guiding response text sequence that semantically matches the fused feature vector.
[0052] Optionally, the autoregressive decoding process generates a word-by-word sequence of guiding response text based on the decoded feature vector. At each step, the model selects the word with the highest probability from the current decoded feature vector as the next word to be generated. Taking the elevator malfunction problem as an example, the model might first select the word "suggest," then select "check" based on the updated decoded feature vector, and so on, progressively generating a complete sequence of response text. During the generation process, the decoded feature vector is continuously updated to reflect the information of the words already generated.
[0053] Step 1325: Perform repeated word filtering and sentence fluency correction on the guidance response text sequence to generate the initial guidance response text.
[0054] After obtaining the sequence of guiding response text, repeated word filtering and sentence fluency correction are performed. For repeated word filtering, if consecutive words appear in the sequence, the redundant words are removed. For example, if the sequence contains "check check," only one "check" will be retained. Sentence fluency correction checks the syntax and semantics of the sentences in the sequence, adjusting the word order to make the sentences more fluent. Through these processes, the initial guiding response text is finally generated, such as "It is recommended to check whether there is a fault in the elevator's electrical system."
[0055] Step 133: Perform equipment terminology compliance checks on the initial guidance response text, replace common terms in the initial guidance response text that do not conform to the building equipment operation and maintenance field standards with field-specific terms, and generate the first candidate operation and maintenance guidance text.
[0056] When conducting equipment terminology compliance checks, the initial guidance response text is compared with standards in the field of building equipment operation and maintenance. Taking "It is recommended to check whether there is a fault in the elevator's electrical system" as an example, the text is checked for common terms that do not conform to the standards. If so, appropriate domain-specific terms are searched from a pre-built template terminology library and replaced.
[0057] Specifically, the initial guidance response text undergoes equipment terminology compliance checks, replacing generic terms in the initial guidance response text that do not conform to building equipment operation and maintenance standards with domain-specific terms to generate a first candidate operation and maintenance guidance text, including: Step 1331: Extract the set of device-related terms from the initial guidance response text.
[0058] For the initial guidance response text "It is recommended to check whether there is a fault in the elevator's electrical system," the set of equipment-related words, namely "elevator" and "electrical system," will be extracted. In practice, these words are extracted by performing syntactic analysis and semantic understanding on the initial guidance response text. This process identifies words related to building equipment, excludes some general descriptive words, and ensures that the extracted word set accurately reflects equipment-related information.
[0059] Step 1332: Perform a matching query process between the set of equipment-related terms and the template terminology database in the knowledge base of building equipment operation and maintenance, and identify non-standard terms in the set of equipment-related terms that do not conform to the standard terminology specifications.
[0060] Furthermore, the extracted set of equipment-related terms is compared with a template terminology database in the building equipment operation and maintenance knowledge base. The template terminology database contains all standard professional terms in the building equipment operation and maintenance field. For "elevator" and "electrical system," the corresponding standard expressions are searched in the template terminology database. If a term completely matches the standard term in the template terminology database, it is considered compliant; otherwise, it is identified as a non-standard term. In this example, if the template terminology database has a more precise description of "electrical system," and the current "electrical system" differs from it, then "electrical system" will be identified as a non-standard term. The matching query process compares the semantics and spelling of the terms to ensure accurate identification of non-standard terms.
[0061] Step 1333: For each of the non-standard words, extract the domain-specific term with the smallest semantic distance from the template terminology library as a replacement candidate word.
[0062] Once non-standard terms are identified, the template terminology library is searched for domain-specific terms with the smallest semantic distance as replacement candidates. The semantic distance calculation considers the meaning, usage, and relevance of the term in the field of building equipment operation and maintenance. Taking the identified non-standard term "electrical system" as an example, the template terminology library is searched for the standard term with the closest semantics. There might be a more precise expression in the template terminology library, such as "elevator electrical control system," which, because it is semantically closest to "electrical system," is selected as a replacement candidate. During the search for replacement candidates, semantic analysis is performed on each term in the template terminology library, comparing it with non-standard terms to determine the term with the smallest semantic distance.
[0063] Step 1334: Based on the contextual semantic relationship of the initial guidance response text, select the replacement candidate word with the highest semantic fit with the context for word replacement processing.
[0064] After identifying candidate replacement words, the most suitable replacement word is selected based on the contextual semantic relationships of the initial guidance response text. Taking "It is recommended to check if there is a fault in the elevator's electrical system" as an example, if there are candidate replacement words such as "elevator electrical control system" and another similar term that doesn't quite fit the context, the context will be considered comprehensively. Since the text is about troubleshooting elevator malfunctions, "elevator electrical control system" more accurately expresses the object to be checked and has the highest semantic fit with the context, so it will be selected for replacement. Further analysis of the grammatical structure and semantic coherence of the candidate replacement words within the sentence ensures that the replaced sentence remains semantically fluent and reasonable.
[0065] Step 1335: Perform semantic coherence adjustment on the replaced guidance response text to generate the first candidate operation and maintenance guidance text.
[0066] After word replacement, the semantic coherence of the resulting guidance response text is adjusted. Taking the replaced sentence "It is recommended to check if there is a fault in the elevator's electrical control system" as an example, redundancy is found. The sentence is then optimized by adjusting word order and removing unnecessary repetitions to make it more fluent and grammatically correct. After adjustment, the first candidate maintenance guidance text is generated, such as "It is recommended to check if there is a fault in the elevator's electrical control system." During the adjustment process, the overall semantics and logical relationships of the sentence are considered to ensure that the generated text is semantically coherent and accurate.
[0067] Step 134: Perform an operation logic integrity check on the initial guidance response text, supplement the missing key operation steps in the initial guidance response text, and generate a second candidate operation and maintenance guidance text.
[0068] The initial guidance response text, "It is recommended to check whether there is a fault in the elevator's electrical system," undergoes an operational logic integrity check. This involves analyzing whether the text contains all the key steps required to complete the inspection. In this example, only the text mentions checking whether there is a fault in the electrical system, but it doesn't explain how to perform the inspection. It may be necessary to add some key operational steps, such as checking whether electrical wiring connections are loose or checking the working status of electrical components. Based on standard procedures and experience in building equipment operation and maintenance, the missing steps in the initial text are identified and added. During the addition process, it is ensured that the added steps match the semantics and operational logic of the initial text, ultimately generating a second candidate operation and maintenance guidance text, such as "It is recommended to check whether the electrical wiring connections in the elevator's electrical system are loose and to check whether the working status of electrical components is normal to determine whether there is a fault."
[0069] Step 135: Perform a language simplification conversion on the initial guidance response text to generate a third candidate operation and maintenance guidance text containing explanatory statements.
[0070] The steps aim to simplify the initial guidance response text, making it easier to understand. For example, the technical terminology in the sentence "It is recommended to check if there is a fault in the elevator's electrical system" will be explained, with "electrical system" being interpreted as "the circuits and equipment in the elevator that control various functions." The sentence structure will also be adjusted to be more conversational. After the transformation, a third candidate maintenance guidance text containing explanatory statements will be generated, such as "It is recommended that you check if there are any problems with the circuits and equipment in the elevator that control various functions." During the transformation process, appropriate explanations and expressions will be selected based on the target user's knowledge level and language habits.
[0071] Step 136: Merge the first candidate operation and maintenance guidance text, the second candidate operation and maintenance guidance text, and the third candidate operation and maintenance guidance text to form a set of candidate operation and maintenance guidance texts.
[0072] Optionally, the generated first, second, and third candidate maintenance guidance texts can be merged. These three candidate texts optimize the initial guidance response text from different perspectives, and merging them yields a more comprehensive and richer set of candidate maintenance guidance texts. For example, the first candidate maintenance guidance text emphasizes the compliance of equipment terminology, the second candidate maintenance guidance text supplements the completeness of operational logic, and the third candidate maintenance guidance text achieves simplified language. The merged set of candidate maintenance guidance texts incorporates the advantages of these three aspects.
[0073] Step 140: Based on the knowledge base of building equipment operation and maintenance domain, perform compliance verification and semantic adaptation processing on the candidate operation and maintenance guidance text set, generate target operation and maintenance guidance text, and embed the target operation and maintenance guidance text into the current operation and maintenance session interaction process to complete the session response operation.
[0074] Based on a set of candidate maintenance guidance texts, compliance verification and semantic adaptation are performed using a knowledge base in the building equipment maintenance domain. The knowledge base contains various specifications, standards, and cases related to building equipment maintenance. For each candidate maintenance guidance text in the set, its compliance with equipment safety operating procedures and whether it falls within the equipment's performance limitations are checked. Simultaneously, the semantic matching degree between the text and the user's question, its adaptability to the current maintenance scenario, and its coherence with the historical conversation context are evaluated. After screening and processing, the text that best meets the requirements is selected as the target maintenance guidance text. The target maintenance guidance text is then embedded into the current maintenance conversation interaction flow, for example, by sending it to the user through a chat interface to complete the conversation response.
[0075] In an alternative embodiment, the step of performing compliance verification and semantic adaptation processing on the candidate maintenance guidance text set based on the building equipment maintenance domain knowledge base to generate the target maintenance guidance text includes: Step 141: Extract the core operation instruction set for each candidate operation and maintenance guidance text from the candidate operation and maintenance guidance text set.
[0076] Each text in the candidate maintenance guidance text set is analyzed to extract its core operation instruction set. Taking the text "It is recommended to check whether the electrical wiring connections in the elevator electrical system are loose and whether the electrical components are working properly to determine if there is a fault" as an example, the core operation instruction set includes "check whether the electrical wiring connections are loose" and "check whether the electrical components are working properly". Through semantic understanding and grammatical analysis of the text, the parts that directly guide the operation are identified, while some descriptive and explanatory content is excluded to ensure that the extracted core operation instructions accurately reflect the key operation information of the text.
[0077] Step 142: Perform matching and verification processing between the core operation instruction set and the operation specification library in the building equipment operation and maintenance knowledge base to identify illegal operation instructions in the core operation instruction set that violate equipment safety operation specifications or exceed equipment performance limits.
[0078] The extracted set of core operation instructions is compared with the operation specification library in the knowledge base for building equipment operation and maintenance. The operation specification library contains safety specifications and performance limitations for various operations of building equipment. For the two core operation instructions, "check if electrical wiring connections are loose" and "check if electrical components are functioning properly," their compliance with the requirements in the operation specification library is checked. If the operation specification library stipulates that the power must be disconnected before checking electrical wiring, and the current instruction does not mention this, it may pose a safety hazard and will be identified as a violation. Then, each core operation instruction undergoes a detailed check to ensure it complies with the equipment's safety operation specifications and performance limitations.
[0079] Under a preferred design approach, the process of matching and verifying the core operation instruction set with the operation specification library in the building equipment operation and maintenance knowledge base to identify non-compliant operation instructions in the core operation instruction set that violate equipment safety operation specifications or exceed equipment performance limits includes: Step 1420: Parse the device type identifier and operation type identifier in the core operation instruction set. The device type identifier is used to locate building equipment, and the operation type identifier is used to locate operation behavior. Based on the device type identifier and the operation type identifier, extract the corresponding safety operation constraints and equipment performance limitations from the operation specification library. Perform constraint matching checks on each operation instruction in the core operation instruction set, and identify instructions that do not meet the safety operation constraints as safety violation instructions. Perform performance limitation matching checks on each operation instruction in the core operation instruction set, and identify instructions that exceed the equipment performance limitations as performance violation instructions. Merge the safety violation instructions and the performance violation instructions to form a set of violation operation instructions.
[0080] For the core operation instruction sets "Check if electrical wiring connections are loose" and "Check if electrical components are functioning normally," the equipment type identifier is identified as "elevator," and the operation type identifiers are "Check electrical wiring connections" and "Check electrical component functioning status," respectively. Based on these identifiers, corresponding safety operation constraints and equipment performance limitations are extracted from the operation specification library. For elevator electrical system inspections, safety operation constraints may include disconnecting the power supply first and using insulated tools; equipment performance limitations may include the operating voltage and current range of electrical components. Each operation instruction undergoes a constraint matching check to verify compliance with safety operation constraints. If the "Check if electrical wiring connections are loose" instruction does not mention disconnecting the power supply first, it will be identified as a safety violation instruction. Each operation instruction undergoes a performance limitation matching check; if the required inspection method exceeds the performance limitations of the electrical components, it will be identified as a performance violation instruction. Finally, the safety violation instructions and performance violation instructions are merged to form a set of violation operation instructions.
[0081] Step 143: Perform instruction correction processing on candidate operation and maintenance guidance texts containing non-compliant operation instructions, replace the non-compliant operation instructions with safe operation instructions that meet the requirements of the operation specification library, and generate compliance correction text.
[0082] For candidate maintenance guidance texts containing instructions for non-compliant operations, the instructions are corrected. For example, a candidate maintenance guidance text containing the safety violation instruction "Check if electrical wiring connections are loose" will be corrected to "First disconnect the elevator power supply, then use insulated tools to check if electrical wiring connections are loose," according to the requirements of the operation specification library. This ensures that the corrected instruction complies with the equipment's safety operation specifications and performance limitations. During the correction process, standard operating procedures in the operation specification library are referenced to adjust and supplement the non-compliant instruction, generating a compliant corrected text, such as "It is recommended to first disconnect the elevator power supply, then use insulated tools to check if the electrical wiring connections in the elevator electrical system are loose, and check if the electrical components are functioning normally to determine if a fault exists."
[0083] Step 144: Perform semantic adaptability evaluation processing on the compliance correction text, calculate the semantic matching degree between the compliance correction text and the user's question, the scenario adaptability degree with the current operation and maintenance scenario, and the coherence index with the historical session context, and obtain the semantic adaptability evaluation processing result.
[0084] The compliance correction text undergoes semantic fit evaluation, which involves calculating the semantic matching degree between the text and the user's question, the scenario fit degree with the current operation and maintenance scenario, and the coherence index with the historical conversation context. For semantic matching degree, it analyzes whether the core semantics of the compliance correction text are consistent with the core semantics of the user's question, "The elevator has been frequently stuck lately, what's going on?" If the correction text revolves around elevator troubleshooting, and the troubleshooting content is related to the stuck elevator problem, then the semantic matching degree is high. For scenario fit degree, it considers the current operation and maintenance scenario as elevator maintenance in a commercial building, checking whether the operations in the correction text are applicable to this scenario. If the operations in the correction text are for troubleshooting common elevator problems in commercial buildings, then the scenario fit degree is high. For the coherence index, it analyzes the semantic continuity between the correction text and the historical conversation context. If the relevant issues of elevator stuck elevators have already been discussed in the historical conversation, and the correction text can continue to address this issue, then the coherence index is high. The semantic fit evaluation result is obtained through these three aspects of evaluation.
[0085] Optionally, the semantic fit evaluation processing of the compliance correction text, which calculates the semantic matching degree between the compliance correction text and the user's question, the scenario fit degree with the current operation and maintenance scenario, and the coherence index with the historical session context, to obtain the semantic fit evaluation processing result, includes: Step 1440: Input the compliance correction text and the user question into the semantic matching model, and generate a semantic matching score reflecting the degree of overlap of the core semantics of the two through the cosine similarity calculation method; extract the set of scenario keywords in the compliance correction text, and perform matching and counting processing on the set of scenario keywords and the set of scenario tags in the current operation and maintenance scenario to generate a scenario fit score; input the compliance correction text and the selected subset of operation and maintenance question-and-answer dialogue units in the historical conversation context into the coherence evaluation model, and generate a coherence index reflecting the semantic continuity between the text through the sequence similarity calculation method; perform weighted summation processing on the semantic matching score, scenario fit score and coherence index to generate a comprehensive fit score as the semantic fit evaluation result.
[0086] The compliance correction text and user questions are input into a semantic matching model. The model performs semantic analysis, calculating the overlap of their core semantics using cosine similarity to generate a semantic matching score. A higher score indicates greater similarity in their core semantics. A set of scenario keywords, such as "elevator" and "electrical system," is extracted from the compliance correction text. These keywords are then matched against the current maintenance scenario's tag set to generate a scenario fit score. A higher score indicates better fit between the correction text and the current maintenance scenario. Finally, the compliance correction text and a selected subset of maintenance Q&A dialogue units from historical conversations are input into a coherence evaluation model. This model uses sequence similarity to analyze the semantic continuity between the preceding and following texts, generating a coherence metric. A higher metric indicates better coherence between the correction text and the historical conversation context. Finally, the semantic matching score, scene adaptability score, and coherence index are weighted and summed. Based on the preset weights, a comprehensive adaptability score is calculated as the semantic adaptability evaluation result.
[0087] Step 145: Select the compliance correction text that meets the corresponding preset threshold requirements in terms of semantic matching degree, scenario adaptability and coherence as the target operation and maintenance guidance text.
[0088] The compliance correction texts are screened based on preset threshold requirements. If a compliance correction text's semantic matching score, scenario adaptability score, and coherence metric all reach the corresponding preset thresholds, then that text is selected as the target operation and maintenance guidance text. For example, the preset semantic matching threshold is a certain similarity level, the scenario adaptability threshold is a certain number of matching keywords, and the coherence metric threshold is a certain sequence similarity level. Only when a compliance correction text meets the threshold requirements in all three aspects will it be selected as the target operation and maintenance guidance text.
[0089] Step 146: Obtain the latest user question in the current operation and maintenance session interaction process, and extract the latest core semantic reference and the latest operation and maintenance requirement association features of the latest user question.
[0090] In the current operations and maintenance (O&M) session interaction flow, the latest user question is retrieved. For example, if the latest user question is "I checked as you instructed, but found no problem. What should I do next?", this question will be analyzed to extract its latest core semantic focus and the latest O&M requirement association features. Through semantic understanding of the question, it is determined that the core semantic focus is asking for the next step after no problem was found, and the latest O&M requirement association feature is continuing troubleshooting and resolution. The previously mentioned semantic focus analysis model and O&M task association analysis method will be used to process the latest question to obtain accurate core semantic focus and O&M requirement association features.
[0091] Step 147: Perform semantic alignment processing between the target maintenance guidance text and the latest core semantic reference of the user's latest question statement to obtain semantically aligned guidance text; perform requirement matching processing between the target maintenance guidance text and the latest maintenance requirement association features of the user's latest question statement to ensure that the response content covers the user's implicit maintenance task requirements, resulting in requirement-matched guidance text; perform language style adaptation processing on the target maintenance guidance text, adjusting the text's expression according to the user's language habits in historical conversations to obtain language style-adapted guidance text.
[0092] The target maintenance guidance text concerns elevator troubleshooting, while the latest core semantic focus is on inquiring about the next steps if no problems are found during the inspection. The target maintenance guidance text will be adjusted to address this latest core semantic focus, resulting in semantically aligned guidance text. The target maintenance guidance text will then be matched with the latest maintenance requirement's associated features for requirement matching. Since the latest maintenance requirement's associated features are continued troubleshooting and resolution, the response content in the target maintenance guidance text will be ensured to cover these requirements, generating requirement-matched guidance text. Finally, the target maintenance guidance text will undergo language style adaptation processing, analyzing the user's language habits in historical conversations, such as vocabulary and expression. If the user prefers to use easily understandable language in historical conversations, the target maintenance guidance text will be adjusted to use more colloquial expressions, resulting in language style-adapted guidance text.
[0093] Step 148: Combine the semantic alignment guidance text, the requirement matching guidance text, and the language style adaptation guidance text to generate the current session response content, and output the current session response content to the user terminal device.
[0094] By combining semantic alignment guidance text, requirement matching guidance text, and language style adaptation guidance text, the advantages of these three texts are taken into account, and their content is integrated to generate the current session response content. For example, the semantic alignment part, the requirement-matching part, and the language style-matching part are merged to form a complete and accurate response content. The generated current session response content is then output to the user's terminal device, such as through chat software, SMS, etc., to complete the session response operation.
[0095] As a non-limiting embodiment, after embedding the target maintenance guidance text into the current maintenance session interaction flow to complete the session response operation, the method further includes: Obtain the user's interactive feedback text in response to the target operation and maintenance guidance text, wherein the interactive feedback text includes user follow-up questions, confirmation statements, or operation result description statements; The interactive feedback text is parsed at the intent level to extract semantic satisfaction markers, demand coverage markers, and operation guidance clarity markers from the user feedback as a set of feedback features. The feedback feature set is subjected to cross-modal correlation analysis with the semantic recognition features and operation and maintenance requirement association features of the target operation and maintenance guidance text to generate an interaction quality evaluation feature vector that reflects the session response effect. The pre-trained operation and maintenance session generation model is calibrated based on the interaction quality assessment feature vector, and the feature fusion weight parameters of the semantic generation layer in the operation and maintenance session generation model are adjusted to form an adaptively optimized operation and maintenance session generation model.
[0096] In this embodiment of the application, after completing the session response operation, the user's interactive feedback text for the target maintenance guidance text will be obtained. The user's interactive feedback text may be a follow-up question, such as "I followed the instructions but the problem was still not solved. Are there any other reasons?"; it may also be a confirmation statement, such as "I have completed the inspection according to the instructions, and there is indeed no problem"; or it may be a description of the operation result, such as "The inspection found that a certain electrical component has minor damage".
[0097] When performing intent-level parsing on these interactive feedback texts, natural language processing techniques and pre-trained intent recognition models are employed. For follow-up questions, the system analyzes whether the user is dissatisfied with the current guidance and needs further troubleshooting suggestions, thus determining that the semantic satisfaction rating may be unsatisfactory. For confirmation statements, it can be determined that the user considers the current guidance effective, and the semantic satisfaction rating is satisfactory. For statements describing the operation results, the system can determine whether the problem has been solved based on the description, thereby determining the semantic satisfaction rating. The requirement coverage rating is determined based on user feedback to see if the target maintenance guidance text covers all of the user's maintenance needs. If the user raises new requirements in the feedback, it indicates insufficient requirement coverage. The operation guidance clarity rating considers whether the user can successfully follow the instructions. If the user reports confusion during the operation, it indicates that the operation guidance is not clear enough. Through the above analysis, the semantic satisfaction rating, requirement coverage rating, and operation guidance clarity rating are extracted from the user feedback to form a feedback feature set.
[0098] Next, cross-modal correlation analysis is performed on the feedback feature set and the semantic recognition features and operation and maintenance requirement association features of the target operation and maintenance guidance text. Cross-modal correlation analysis comprehensively considers the semantic relationships and logical connections between different features. Semantic recognition features reflect the core semantic direction of the target operation and maintenance guidance text, operation and maintenance requirement association features reflect the relevance to the building equipment operation and maintenance tasks, and the feedback feature set contains the user's actual feedback to the guidance text. By analyzing the interactions between these features, an interaction quality assessment feature vector reflecting the effectiveness of the session response can be generated. For example, if the semantic satisfaction level is marked as unsatisfied, but the semantic recognition features show a deviation between the core semantics of the target operation and maintenance guidance text and the user's question, this inconsistency will be reflected in the interaction quality assessment feature vector.
[0099] Based on the generated interaction quality assessment feature vector, the pre-trained operation and maintenance (O&M) session generation model is calibrated. The calibration process adjusts the feature fusion weight parameters of the semantic generation layer within the O&M session generation model. The semantic generation layer is responsible for fusing semantic recognition features and O&M requirement-related features to generate guidance text. When the interaction quality assessment feature vector shows that certain features are not performing reasonably in generating guidance text, the fusion weights of these features are adjusted accordingly. For example, if insufficient coverage of user feedback is found, and O&M requirement-related features have low weights in generated text, the fusion weight of these features is appropriately increased, enabling the model to better meet users' O&M needs in subsequent guidance text generation. After these adjustments, an adaptively optimized O&M session generation model is formed, which continuously improves the quality of the generated guidance text based on user feedback.
[0100] As a non-limiting embodiment, the step of performing conversation feature parsing processing on the historical question-and-answer conversation text set to obtain the semantic recognition features and operation and maintenance requirement association features of each operation and maintenance question-and-answer dialogue unit is implemented through a conversation feature parsing processing network. After embedding the target operation and maintenance guidance text into the current operation and maintenance conversation interaction process to complete the conversation response operation, the method further includes: Extract the user's questions and the target maintenance guidance text from this conversation as new maintenance Q&A dialogue units; The newly added operation and maintenance Q&A dialogue unit is subjected to semantic overlap detection processing with the historical Q&A conversation text set, and the newly added dialogue unit whose semantic difference with the historical dialogue unit exceeds a preset threshold is selected as valid derived data. The effective derived data is subjected to feature annotation processing to generate a derived feature set containing newly added semantic recognition features and newly added operation and maintenance requirement related features; Based on the aforementioned derivation, the parameters of the session feature parsing and processing network are iteratively updated to adjust the keyword extraction strategy and intent matching rules for the operation and maintenance task correlation analysis of the session feature parsing and processing network.
[0101] In this embodiment of the application, after completing the session response operation, the user's question and the target maintenance guidance text from the current session interaction are extracted to form a new maintenance Q&A dialogue unit. For example, if a user asks, "The elevator has been frequently stuck recently, and I can't find the cause, what should I do?", and the target maintenance guidance text is "It is recommended to check whether the elevator's control system parameters are abnormal," these two constitute a new maintenance Q&A dialogue unit.
[0102] The newly added operations and maintenance Q&A dialogue unit is subjected to semantic overlap detection with the historical Q&A conversation text set. Semantic overlap detection uses semantic similarity calculation methods to analyze the degree of similarity between the new and historical dialogue units in core semantics. Similarity is calculated by comparing the consistency of keywords, semantic topics, etc. If the similarity between the new and historical dialogue units is low, it indicates a large semantic difference. A preset threshold is set; when the semantic difference exceeds this threshold, the new dialogue unit is filtered out as valid derived data. This is because dialogue units with large semantic differences may contain new question types or operations and maintenance requirements, which are of significant value for model updates.
[0103] The selected valid derived data undergoes feature annotation processing. The previously mentioned conversation feature parsing method is then applied again to perform semantic recognition and maintenance requirement correlation analysis on user queries. For example, for the query "The elevator has been frequently stuck lately, and I can't find the cause, what should I do?", the building equipment type is identified as an elevator, the symptom is frequent stuckness with no identifiable cause, and the operational requirement is likely in-depth troubleshooting, thus yielding new semantic recognition features. Simultaneously, based on the association between the query and building equipment maintenance tasks, new maintenance requirement correlation features are determined. Through the above annotation, a derived feature set containing the new semantic recognition features and the new maintenance requirement correlation features is generated.
[0104] Based on these derived feature sets, the conversation feature parsing processing network undergoes iterative parameter updates. During network training, the keyword extraction strategy is adjusted according to the derived feature sets. If new keywords closely related to maintenance tasks are found in the new data, these keywords are included in the keyword extraction scope, improving the accuracy of keyword extraction. The intent matching rules for maintenance task correlation analysis are also adjusted based on the derived feature sets. For example, if a new fault condition appears in the new data, the original intent matching rules may not accurately determine its correlation with maintenance tasks. In this case, the rules are updated based on the new data, enabling the rules to more accurately identify the relationship between different problems and maintenance tasks. Through these iterative parameter updates, the conversation feature parsing processing network can adapt to new problems and requirements, improving its feature parsing capabilities for historical question-and-answer conversation text.
[0105] As a non-limiting embodiment, after embedding the target maintenance guidance text into the current maintenance session interaction flow to complete the session response operation, the method further includes: Obtain the complete dialogue round record for this session, which includes the user's question sequence and the corresponding operation and maintenance response sequence; Cross-round semantic dependency analysis is performed on the dialogue round records to extract the topic continuity features, semantic progression features, and demand evolution features between user questions and historical responses as long-term context association features; The long-term context association features are input into the context memory module of the operation and maintenance session generation model for feature fusion and storage processing, forming a context memory vector containing deep semantic relationships of historical sessions; In response to subsequent session interaction tasks, the context memory vector is invoked to perform joint generation processing with the semantic recognition features and operation and maintenance requirement association features of the current session.
[0106] In this embodiment of the application, after completing the session response operation, a complete dialogue round record of this session interaction is obtained. This record will contain the user's question sequence and the corresponding maintenance response sequence according to the order of the dialogue. For example, in the first round, the user asks "There is an abnormal noise when the elevator is running," and the maintenance response is "Check the mechanical parts of the elevator first"; in the second round, the user asks "The mechanical parts are checked and there is no problem. What should we do next?" and the maintenance response is "Check the electrical system next."
[0107] Cross-round semantic dependency analysis was performed on this dialogue round record. When analyzing topic continuity, it was determined whether different rounds of dialogue revolved around the same topic. In the example above, the entire dialogue revolved around the topic of elevator troubleshooting, indicating strong topic continuity. Semantic progression focused on whether subsequent rounds of dialogue offered deeper semantic expression based on the previous round. The second round of user questions was raised after checking the mechanical components according to the first round of maintenance response, further inquiring about the next steps, demonstrating semantic progression. Demand evolution considered whether the user's needs changed during the dialogue. In this example, the user's needs evolved from initially investigating the cause of the abnormal noise to further investigating other causes after the mechanical components were found to be fine. Through the above analysis, topic continuity, semantic progression, and demand evolution characteristics between user questions and historical responses were extracted, forming long-term contextual association characteristics.
[0108] These long-term contextual features are input into the context memory module of the operation and maintenance session generation model. The context memory module performs fusion and storage processing on the input features. It integrates these features with previously stored historical session features, and through a certain fusion algorithm, considers the importance and correlation of each feature to form a context memory vector containing deep semantic relationships of historical sessions. This vector can reflect the semantic connections between different rounds in historical sessions and the changes in user needs.
[0109] When subsequent conversational tasks arise, this context memory vector is used in conjunction with the semantic recognition features and maintenance requirement association features of the current conversation for joint generation. For example, when a user asks again, "I checked the electrical system as you instructed, but there's still an abnormal noise. What should I do?", the semantic recognition features of the current question (elevator, abnormal noise, continue troubleshooting) and the maintenance requirement association features are input into the maintenance conversation generation model along with the context memory vector. The model comprehensively considers the deep semantic relationships of historical conversations and the current problem to generate maintenance guidance text that better matches the user's needs and contextual logic, such as, "Based on the previous inspection, it is recommended to check if there is a problem with the elevator's vibration damping device." In this way, information from historical conversations is fully utilized to improve the accuracy and relevance of maintenance guidance text.
[0110] In this embodiment, for further implementation of the operation and maintenance session generation model, a feature cross-attention mechanism can be implemented based on the encoder-decoder structure of the Transformer architecture (such as the BERT-GPT fusion model). Feature fusion weights are dynamically allocated by calculating the mutual information entropy value between semantic recognition features and operation and maintenance requirement-related features. For the mapping rules of the semantic dimension mapping branch, a graph neural network embedding algorithm based on the operation and maintenance knowledge graph can be deployed to automatically generate multi-dimensional semantic vectors using entity relationships such as device type and fault phenomena to optimize the static mapping table query logic. In the global modeling stage of the multi-head self-attention branch, relative position encoding and domain adaptive pre-training techniques are introduced. Gradient backpropagation optimizes the attention head parameters in the operation and maintenance scenario to refine the parameter adjustment mechanism.
[0111] For the matching method of operation specification verification, a rule engine based on dependency parsing and semantic role labeling can be adopted. The core operation instructions are parsed into <subject-action-object> triples and then used for subclass reasoning with the OWL ontology of the knowledge base. This clarifies the confidence threshold for violating security constraints and improves the verification process in the text description. Regarding the dimensions of the generated scoring vector, Z-score standardization can be performed at the feature layer: when outputting the cosine similarity score through the dual-tower Siamese network of the semantic matching model, the synchronous constraint scenario fit index uses the Jaccard coefficient to calculate the label overlap rate, and the coherence index is based on the sequence relationship probability output of the BERT-NSP model. Weighted fusion is then performed after ensuring that all three dimensions are dimensionless values in the [0, 1] interval.
[0112] In the model iteration and optimization stage, the weights of the semantic generation layer can be adjusted by maximizing the mutual information of features from positive feedback samples in a multimodal contrastive learning framework (such as the operation and maintenance domain variant of the CLIP architecture) that inputs the feedback feature set and the model output features.
[0113] This application embodiment obtains a historical Q&A conversation text set of building equipment operation and maintenance scenarios, parses conversation features to obtain semantic recognition features and operation and maintenance requirement association features, jointly calls a pre-trained operation and maintenance conversation generation model to generate a candidate operation and maintenance guidance text set, and then performs compliance verification and semantic adaptation processing based on a building equipment operation and maintenance domain knowledge base to obtain the target operation and maintenance guidance text and embed it into the current operation and maintenance conversation interaction process. This allows for accurate identification of the core issues and operation and maintenance requirements of users in building equipment operation and maintenance scenarios, generating operation and maintenance guidance text that fits actual needs and conforms to domain standards, effectively improving the accuracy, professionalism, and compliance of building equipment operation and maintenance conversation interactions, and providing users with more efficient and high-quality operation and maintenance guidance services.
[0114] Based on the same inventive concept, embodiments of this application also provide a building equipment operation and maintenance conversational interaction system. (See also...) Figure 2As shown, this is a schematic diagram of the structure of a possible building equipment operation and maintenance session interaction system provided in an embodiment of this application. Figure 2 The building equipment operation and maintenance session interaction system 200 includes a processor 210 and a memory 220. The memory 220 stores computer programs that can be executed by the processor 210. By executing the instructions stored in the memory 220, the processor 210 can perform the steps of the aforementioned AIGC-based building equipment operation and maintenance session interaction method.
[0115] Based on the same inventive concept, embodiments of this application provide a computer-readable storage medium including a computer program. When the computer program runs on a building equipment operation and maintenance session interaction system, the computer program causes the building equipment operation and maintenance session interaction system to execute the steps of the aforementioned AIGC-based building equipment operation and maintenance session interaction method. In some possible implementations, various aspects of the AIGC-based building equipment operation and maintenance session interaction method provided in this application can also be implemented as a program product, including a computer program. When the program product runs on a building equipment operation and maintenance session interaction system, the computer program causes the building equipment operation and maintenance session interaction system to execute the steps in the aforementioned AIGC-based building equipment operation and maintenance session interaction method. For example, the building equipment operation and maintenance session interaction system can execute... Figure 1 The steps are shown in the figure.
[0116] In the technical solutions involved in the above embodiments of this application, whether it is performing comparison calculations of multi-dimensional features or constructing composite parameters, if there are problems caused by significant differences in the number of dimensions, units of measurement, and semantic meanings of different features, those skilled in the art, based on their professional knowledge and past practical experience, can fully understand that these differences need to be properly handled so that the calculation results are accurate and comparable, and to avoid situations such as logical confusion and unclear mathematical meaning.
[0117] In detail, when faced with features of different numbers of dimensions, those skilled in the art can employ various strategies, including but not limited to feature selection, feature extraction, and kernel function processing, in order to accurately calculate the similarity, matching degree, or feature distance between different features.
[0118] In order to achieve comparability alignment of feature spaces when comparing multidimensional features, those skilled in the art can use a variety of existing and common technical means, including but not limited to standardization preprocessing, mapping transformation, and spatial projection.
[0119] In the process of constructing composite parameters (such as loss function values), different parameter terms often have different dimensions. Those skilled in the art can use normalization processing or an adaptive weight allocation mechanism based on distribution characteristics.
[0120] The aforementioned general techniques for solving feature matching and loss balance problems are all common knowledge in this field. These techniques have been fully verified and widely used in numerous practical applications, and those skilled in the art can skillfully and flexibly apply these methods to handle similar problems involving differences in dimensions.
[0121] The formulas and calculation processes involved in the embodiments of this application, whether used for multidimensional feature comparison or composite loss function construction, strictly adhere to the principle of dimensional correspondence. Each variable in the formula has a clear and explicit physical meaning, and its operational logic fully conforms to basic mathematical and physical logic. The calculation results are necessarily the reasonable results expected by this application. Those skilled in the art are capable of effectively solving various problems arising from the number of dimensions, dimensional differences, etc., in the multidimensional feature comparison calculation and composite loss function construction in the embodiments, based on specific data conditions and business needs, by comprehensively utilizing the aforementioned general technical means, thus ensuring the accuracy, reliability, and feasibility of the technical solution of this application.
Claims
1. A method for interactive dialogue in building equipment operation and maintenance based on AIGC, characterized in that, The method includes: Obtain a set of historical Q&A conversation texts in the operation and maintenance scenario of building equipment. The set of historical Q&A conversation texts includes multiple consecutive operation and maintenance Q&A dialogue units. Each operation and maintenance Q&A dialogue unit contains user question statements and corresponding operation and maintenance response statements. The historical question-and-answer dialogue text set is subjected to dialogue feature parsing processing to obtain the semantic recognition features and operation and maintenance requirement association features of each operation and maintenance question-and-answer dialogue unit. The semantic recognition features reflect the core semantic orientation of the user's question statement, and the operation and maintenance requirement association features reflect the degree of association between the user's question statement and the operation and maintenance tasks of building equipment. The pre-trained operation and maintenance conversation generation model is invoked to perform joint semantic generation processing on the semantic recognition features and the operation and maintenance requirement association features to generate a set of candidate operation and maintenance guidance texts corresponding to the operation and maintenance question and answer dialogue unit. Based on a knowledge base in the building equipment operation and maintenance domain, the candidate operation and maintenance guidance text set is subjected to compliance verification and semantic adaptation processing to generate a target operation and maintenance guidance text. The target operation and maintenance guidance text is then embedded into the current operation and maintenance session interaction process to complete the session response operation. This includes: extracting the core operation instruction set of each candidate operation and maintenance guidance text set from the candidate operation and maintenance guidance text set; performing matching and verification processing of the core operation instruction set with the operation specification library in the building equipment operation and maintenance domain knowledge base; using a rule engine based on dependency parsing and semantic role labeling, parsing the core operation instructions into <subject-action-object> triples and performing subclass reasoning with the OWL ontology of the knowledge base to quantify the confidence threshold for violating security constraints. After embedding the target maintenance guidance text into the current maintenance session interaction flow to complete the session response operation, the method further includes: Obtain the complete dialogue round record for this session, which includes the user's question sequence and the corresponding operation and maintenance response sequence; Cross-round semantic dependency analysis is performed on the dialogue round records to extract the topic continuity features, semantic progression features, and demand evolution features between user questions and historical responses as long-term context association features; The long-term context association features are input into the context memory module of the operation and maintenance session generation model for feature fusion and storage processing, forming a context memory vector containing deep semantic relationships of historical sessions; In response to subsequent session interaction tasks, the context memory vector is invoked to perform joint generation processing with the semantic recognition features and operation and maintenance requirement association features of the current session.
2. The AIGC-based building equipment operation and maintenance session interaction method according to claim 1, characterized in that, The process of parsing the historical question-and-answer dialogue text set yields the semantic recognition features and operation and maintenance requirement association features for each operation and maintenance question-and-answer dialogue unit, including: The user's question statements in the operation and maintenance Q&A dialogue unit are segmented into words to obtain a set of conversational lexical units containing operation and maintenance professional terms and general descriptive words; The pre-built semantic pointing analysis model is invoked to perform context association analysis on the conversation word set, and semantic keywords related to building equipment type, fault phenomenon and operation requirements in the user's question statement are extracted as core semantic pointing identifiers; The core semantic pointer is used as the input semantic dimension mapping branch to generate a multi-dimensional semantic vector containing the device type dimension, fault type dimension, and operation requirement dimension as semantic recognition features. The operation and maintenance task correlation analysis is performed on the user's question statement to identify the various operation and maintenance task intentions implied in the user's question statement; Based on the matching results between the various operation and maintenance task intentions and the preset building equipment operation and maintenance task classification system, a correlation score vector reflecting the relevance between user questions and operation and maintenance tasks is generated as an operation and maintenance requirement correlation feature.
3. The AIGC-based building equipment operation and maintenance session interaction method according to claim 1, characterized in that, The process of parsing the historical question-and-answer dialogue text set yields the semantic recognition features and operation and maintenance requirement association features for each operation and maintenance question-and-answer dialogue unit, including: The user questions and corresponding operation and maintenance response statements in the operation and maintenance Q&A dialogue unit are aligned by dialogue rounds to generate a semantic connection between user questions and historical responses. Keyword repetition rate and semantic topic consistency parameter are extracted from the semantic sequence relationship as conversation coherence features; The conversation coherence features are fused with the core semantic orientation of the user's question to generate semantic recognition features that include semantic depth and contextual relevance. The associated features of the operation and maintenance requirements are analyzed for task scenario adaptation to identify the matching relationship between the building equipment type corresponding to the user's question and the target operation and maintenance scenario. Based on the matching relationship, operation and maintenance requirement association features containing scenario adaptation tags and task priority identifiers are generated.
4. The AIGC-based building equipment operation and maintenance session interaction method according to claim 1, characterized in that, The pre-trained operation and maintenance conversation generation model is invoked to perform joint semantic generation processing on the semantic recognition features and the operation and maintenance requirement association features, generating a set of candidate operation and maintenance guidance texts corresponding to the operation and maintenance question-and-answer dialogue unit, including: The semantic recognition features and the operation and maintenance requirement association features are input into the feature fusion layer of the operation and maintenance session generation model. The semantic dimension and the requirement dimension are established through the feature cross-attention mechanism to generate a fused feature vector. The fused feature vector is sequence-decoded through the text generation layer of the operation and maintenance session generation model to generate initial guidance response text; The initial guidance response text is subjected to equipment terminology compliance checks, and common terms that do not conform to the building equipment operation and maintenance field standards in the initial guidance response text are replaced with field-specific terms to generate the first candidate operation and maintenance guidance text; The initial guidance response text is subjected to an operational logic integrity check, and key operational steps missing in the initial guidance response text are supplemented to generate a second candidate operation and maintenance guidance text. The initial guidance response text is converted into a more accessible language to generate a third candidate operation and maintenance guidance text containing explanatory statements; The first candidate operation and maintenance guidance text, the second candidate operation and maintenance guidance text, and the third candidate operation and maintenance guidance text are merged to form a set of candidate operation and maintenance guidance texts.
5. The AIGC-based building equipment operation and maintenance session interaction method according to claim 4, characterized in that, The step of performing sequence decoding processing on the fused feature vector through the text generation layer of the operation and maintenance session generation model to generate initial guidance response text includes: The fused feature vector is input into the position encoding branch of the text generation layer to add position information encoding that reflects the session turn order to the fused feature vector; The multi-head self-attention branch of the text generation layer is used to perform global dependency modeling on the fused feature vector with positional information encoding to generate an intermediate feature representation containing contextual semantic associations. The intermediate feature representation is subjected to nonlinear transformation processing through the feedforward neural network branch of the text generation layer to generate a decoded feature vector containing the probability distribution of words. Autoregressive decoding is performed based on the decoded feature vector to generate a word-by-word guiding response text sequence that semantically matches the fused feature vector; The sequence of guiding response texts is subjected to repeated word filtering and sentence fluency correction to generate the initial guiding response text.
6. The AIGC-based building equipment operation and maintenance session interaction method according to claim 4, characterized in that, The initial guidance response text undergoes a device terminology compliance check, replacing generic terms in the initial guidance response text that do not conform to the building equipment operation and maintenance field standards with domain-specific terms to generate a first candidate operation and maintenance guidance text, including: Extract the set of device-related terms from the initial guidance response text; The set of equipment-related terms is matched and queried with the template terminology library in the knowledge base of building equipment operation and maintenance to identify non-standard terms in the set of equipment-related terms that do not conform to the standard terminology specification. For each of the non-standard words, the domain-specific term with the smallest semantic distance is extracted from the template terminology library as a replacement candidate word; Based on the contextual semantic relationship of the initial guidance response text, the replacement candidate word with the highest semantic fit with the context is selected for word replacement processing; The semantic coherence of the replaced guidance response text is adjusted to generate the first candidate operation and maintenance guidance text.
7. The AIGC-based building equipment operation and maintenance session interaction method according to claim 1, characterized in that, The candidate maintenance guidance text set is subjected to compliance verification and semantic adaptation processing based on a knowledge base in the building equipment maintenance domain to generate the target maintenance guidance text, including: Extract the core operation instruction set of each candidate operation and maintenance guidance text from the candidate operation and maintenance guidance text set; The core operation instruction set is matched and verified with the operation specification library in the knowledge base of building equipment operation and maintenance to identify illegal operation instructions in the core operation instruction set that violate equipment safety operation specifications or exceed the equipment performance limits. Candidate operation and maintenance guidance texts containing non-compliant operation instructions are processed to correct the non-compliant operation instructions, replacing them with safe operation instructions that comply with the operation specification library requirements, and generating compliance correction texts; The compliance correction text is subjected to semantic adaptability evaluation processing. The semantic matching degree between the compliance correction text and the user's question statement, the scenario adaptability degree with the current operation and maintenance scenario, and the coherence index with the historical session context are calculated to obtain the semantic adaptability evaluation processing result. Select compliance correction texts that meet the corresponding preset threshold requirements in terms of semantic matching degree, scenario adaptability, and coherence as target operation and maintenance guidance texts.
8. The AIGC-based building equipment operation and maintenance session interaction method according to claim 7, characterized in that, The process of matching and verifying the core operation instruction set with the operation specification library in the building equipment operation and maintenance knowledge base to identify non-compliant operation instructions in the core operation instruction set that violate equipment safety operation specifications or exceed equipment performance limits includes: The device type identifier and operation type identifier in the core operation instruction set are parsed. The device type identifier is used to locate building equipment, and the operation type identifier is used to locate operation behavior. Based on the equipment type identifier and the operation type identifier, extract the corresponding safe operation constraints and equipment performance limitations from the operation specification library; Each operation instruction in the core operation instruction set undergoes constraint matching checks to identify instructions that do not meet safe operation constraints as safety violation instructions. Each operation instruction in the core operation instruction set is subjected to a performance constraint matching check, and instructions that exceed the device performance constraints are identified as performance violation instructions. The security violation instructions and the performance violation instructions are combined to form a set of violation operation instructions.
9. The AIGC-based building equipment operation and maintenance session interaction method according to claim 7, characterized in that, The semantic fit evaluation process for the compliance correction text involves calculating the semantic matching degree between the compliance correction text and the user's question, the scenario fit degree with the current operation and maintenance scenario, and the coherence index with the historical session context, to obtain the semantic fit evaluation result, including: The compliance correction text and the user's question are input into the semantic matching model, and a semantic matching score reflecting the degree of overlap of the core semantics of the two is generated by the cosine similarity calculation method. Extract the set of scenario keywords from the compliance correction text, match and count the set of scenario keywords with the set of scenario tags of the current operation and maintenance scenario, and generate a scenario adaptability score. The compliance correction text and a subset of selected operation and maintenance Q&A dialogue units in the historical session context are input into the coherence evaluation model, and a coherence index reflecting the semantic connection between the text is generated by the sequence similarity calculation method. The semantic matching score, scene adaptability score, and coherence index are weighted and summed to generate a comprehensive adaptability score as the semantic adaptability evaluation result.
10. A building equipment operation and maintenance interactive system, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program that, when executed by the processor, causes the processor to perform the steps of any one of the methods described in claims 1 to 9.
Citation Information
Patent Citations
Knowledge base sensitive question test method, system, equipment and medium
CN119807683A
Multi-round dialogue interaction method and system based on context reconstruction and multi-library retrieval
CN120123485A
Intelligent question answering method and system for low-altitude economic education for interactive learning platform
CN120234398A
