Intelligent analysis method for special equipment inspection data based on natural language understanding

Abstract

The invention discloses a special equipment inspection data intelligent analysis method based on natural language understanding, and belongs to the technical field of special equipment safety supervision and big data analysis. The method comprises the following steps: receiving and analyzing a natural language request, and extracting a target use unit name and a risk analysis dimension; searching and constructing a device analysis group and matching a risk evaluation model; for each device in the group, extracting related data from the multi-source database; carrying out quantitative calculation by utilizing the matching model, and generating a single-equipment risk score and a risk factor label; carrying out aggregation analysis on a total-amount equipment result, and identifying an overall risk level, a dominant risk type and a high-risk equipment distribution characteristic of the unit; and finally, differential supervision suggestions are generated on the basis. According to the invention, full-process automation from natural language instructions to intelligent risk assessment and decision suggestions is realized.

Description

Technical Field

[0001] This invention relates to the field of special equipment safety supervision and big data analysis technology, specifically to a method for intelligent analysis of special equipment inspection data based on natural language understanding. Background Technology

[0002] The safe operation of special equipment (such as boilers, pressure vessels, elevators, and lifting machinery) is directly related to public safety and social stability. Regulatory authorities need to conduct regular risk assessments of all user units to determine the focus and frequency of supervision. Currently, risk assessment work mainly relies on the manual experience of regulatory personnel, which is cumbersome and inefficient.

[0003] Existing technologies suffer from the following limitations. First, risk assessment initiation relies on structured instructions or fixed menus, making it difficult to directly understand the complex and personalized analytical requests made by regulators in natural language. Second, the data required for risk analysis is scattered across multiple independent systems such as inspection records, monitoring platforms, and regulatory documents, resulting in severe data silos and making manual collection and integration time-consuming and labor-intensive. Furthermore, risk assessment models are relatively fixed and singular, unable to dynamically match the most suitable quantitative model based on different analytical dimensions such as fatigue damage, corrosion, and violations, leading to inaccurate analysis results. Finally, the generation of analytical conclusions depends on manually written reports, lacking standardized and visualized outputs, making it difficult to quickly support precise regulatory decisions based on hierarchical classification.

[0004] The aforementioned limitations restrict the level of intelligence in the safety supervision of special equipment. Therefore, a new method is needed that can understand natural language commands, automatically correlate multi-source data, intelligently match and analyze models, and generate quantitative risk reports to improve the effectiveness of supervision. Summary of the Invention

[0005] To overcome the shortcomings of existing special equipment risk assessment technologies, such as difficulties in understanding instructions, inefficient data integration, rigid model adaptation, and reliance on manual report generation, this invention provides an intelligent analysis method and system for special equipment inspection data based on natural language understanding. This method achieves rapid, accurate, and interpretable risk assessment for user units by constructing an automated analysis chain encompassing natural language parsing, simultaneous extraction of multi-source data, dynamic model matching, and risk quantification aggregation, forming a complete technical closed loop of language input, data aggregation, model calculation, and decision output.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides an intelligent analysis method for special equipment inspection data based on natural language understanding, comprising the following steps: S1: Receive the unit risk assessment request input by the user in natural language, and extract the target unit name and risk analysis dimensions from the unit risk assessment request; S2: Based on the target user's name, associate and retrieve all special equipment files under the target user's name, and construct an equipment analysis group; at the same time, based on the risk analysis dimensions, match the corresponding special equipment safety risk assessment model from the pre-set model library; S3: For each device in the device analysis group, extract multi-source data related to the risk analysis dimensions simultaneously from the historical inspection database, operation monitoring platform and regulatory record library, combined with the changes in the scenario; S4: Based on the special equipment safety risk assessment model, quantitative calculations are performed on the multi-source data of each piece of equipment to generate a single equipment risk score and risk factor label; S5: Statistically analyze and rank the individual risk scores of all devices in the equipment analysis group, and perform aggregate analysis on the risk factor labels to identify the overall risk level, dominant risk type, and distribution characteristics of high-risk devices of the target user unit. S6: Based on the overall risk level, dominant risk type, and distribution characteristics of high-risk equipment of the target user unit, output differentiated regulatory recommendations for the target user unit.

[0007] According to the above technical solution, the step of receiving a unit risk assessment request input by the user in natural language, and extracting the target unit name and risk analysis dimensions from the unit risk assessment request, includes: The system receives input natural language text as a risk assessment request from the intelligent monitoring terminal and web portal system used by regulatory personnel; based on a pre-trained natural language understanding model, it performs dependency parsing and semantic role labeling on the natural language text to obtain semantic labeling results; based on the semantic labeling results, it identifies and extracts the named entities that describe the institutions and enterprises as the target user unit names, and identifies and extracts the core phrases that describe the risk concerns and analysis focus as risk analysis dimensions.

[0008] According to the above technical solution, the step of associating and retrieving all special equipment files under the target user's name to construct an equipment analysis group; simultaneously, matching the corresponding special equipment safety risk assessment model from a pre-set model library according to the risk analysis dimensions, including: Using the extracted target user name as the query key, a fuzzy matching and exact matching retrieval is performed in the special equipment registration database to obtain a list of equipment registration codes under the target user name; based on the equipment registration code list, complete technical files for each piece of equipment are retrieved in batches from the equipment lifecycle archive to construct an equipment analysis group; the extracted risk analysis dimension text is converted into a feature vector, and cosine similarity is calculated with the descriptive keyword vector corresponding to each risk assessment model in the pre-set model library, selecting the model with the highest similarity as the matched special equipment safety risk assessment model.

[0009] According to the above technical solution, for each device in the equipment analysis group, multi-source data related to the risk analysis dimension is simultaneously extracted from the historical inspection database, the operation monitoring platform, and the regulatory record database, taking into account changes in the scenario. This includes: Based on the aforementioned risk analysis dimensions, the corresponding equipment state change scenarios are determined. This determination process follows a predefined scenario-time mapping rule. This scenario-time mapping rule library is constructed through inductive learning from historical risk cases, equipment operation logs, inspection specifications, and domain knowledge. It explicitly stores the mapping relationship from various risk analysis dimensions to specific equipment state change scenarios and defines the corresponding analysis time range calculation logic and key data feature set for each scenario. Then, based on this scenario-time mapping rule, the analysis time range and key data features corresponding to the equipment state change scenarios are extracted. For each device in the equipment analysis group, based on its device registration code, and... Based on the time range and the key features, three data extraction requests are initiated simultaneously: the first request is to the historical inspection database to extract historical inspection records within the time range that are related to the inspection items and the key features; the second request is to the operation monitoring platform to extract time-series data within the time range that matches the sensor type and the key features; and the third request is to the regulatory record database to extract supervisory inspection documents, orders for rectification documents, and administrative penalty documents within the time range that are related to the cause and conclusion and the key features. The received historical inspection records, sensor time-series data, and regulatory document data are integrated to form multi-source data related to the risk analysis dimension.

[0010] According to the above technical solution, the method of using a special equipment safety risk assessment model to quantify and calculate multi-source data for each piece of equipment, generating a single-equipment risk score and risk factor label, includes: The multi-source data of each integrated device is preprocessed, and the textual inspection conclusions and regulatory reasons are converted into numerical values ​​through encoding. The time-series data is feature extracted, and finally all data is converted into a standard feature vector. The standard feature vector is input into the matching special equipment safety risk assessment model to calculate the risk score of a single device. Based on the feature importance and contribution analysis results output by the special equipment safety risk assessment model, the key feature dimensions that contribute the most to the risk score of the single device are identified, and corresponding risk factor labels are generated.

[0011] According to the above technical solution, the step of statistically analyzing and ranking the individual risk scores of all devices in the equipment analysis group, and performing aggregate analysis on risk factor labels to identify the overall risk level, dominant risk type, and distribution characteristics of high-risk equipment of the target user unit includes: The average risk score of each device in the device analysis group and the proportion of high-risk devices are calculated. Based on a pre-set risk level threshold mapping table, and combined with the average risk score of each device in the device analysis group and the proportion of high-risk devices, the overall risk level of the target user unit is determined. The frequency and combination pattern of all risk factor tags are calculated, and the tag combination with the highest frequency is identified as the dominant risk type. The devices are sorted in descending order according to their individual risk scores, and the devices in the device analysis group whose individual risk scores exceed the high-risk device identification threshold are identified as high-risk devices. The distribution characteristics of the high-risk devices are then analyzed.

[0012] According to the above technical solution, the differentiated regulatory recommendations for the target user unit, based on the overall risk level, dominant risk type, and distribution characteristics of high-risk equipment, include: Based on the overall risk level of the target user unit, a corresponding overall regulatory intensity recommendation is generated; based on the dominant risk type of the target user unit, a corresponding specific risk investigation recommendation is generated; and based on the distribution characteristics of high-risk equipment of the target user unit, a corresponding precise regulatory recommendation for high-risk equipment is generated. These are combined to form a hierarchical regulatory recommendation for the target user unit.

[0013] Secondly, the present invention provides an intelligent analysis system for special equipment inspection data based on natural language understanding, used to implement the above method, the system comprising: The request parsing module is used to receive unit risk assessment requests input by users in natural language, and extract the target unit name and risk analysis dimensions from the unit risk assessment requests; The retrieval and matching module, based on the name of the target user unit, associates and retrieves all special equipment files under the name of the target user unit, and constructs equipment analysis groups; at the same time, based on the risk analysis dimensions, it matches the corresponding special equipment safety risk assessment models from the pre-set model library. The data extraction module is used to extract multi-source data related to risk analysis dimensions from historical inspection databases, operation monitoring platforms, and regulatory record databases for each device in the device analysis group, taking into account changes in the scenario. The risk calculation module is used to quantify the multi-source data of each piece of equipment based on the special equipment safety risk assessment model, and generate a single equipment risk score and risk factor label. The aggregation analysis module is used to statistically analyze and sort the individual risk scores of all devices in the device analysis group, and to perform aggregation analysis on risk factor labels to identify the overall risk level, dominant risk type, and distribution characteristics of high-risk devices of the target user unit. The suggestion generation module is used to output differentiated regulatory suggestions for the target user unit based on the overall risk level, dominant risk type, and distribution characteristics of high-risk equipment.

[0014] Thirdly, the present invention provides an electronic device, comprising: a processor and a memory, wherein the memory stores a computer program that can be called by the processor; the processor executes the intelligent analysis method for special equipment inspection data based on natural language understanding as described in the first aspect by calling the computer program stored in the memory.

[0015] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the intelligent analysis method for special equipment inspection data based on natural language understanding as described in the first aspect.

[0016] Compared with the prior art, the present invention has the following advantages and beneficial effects: This invention transforms vague natural language instructions from regulators into clear analytical objectives, reducing reliance on structured queries and improving user-friendliness. This invention solves the data silo problem and improves data preparation efficiency by automatically associating unit and equipment files and simultaneously extracting multi-source heterogeneous data based on analysis dimensions. This invention utilizes semantic similarity to dynamically match the most suitable risk assessment model, making quantitative analysis more targeted. This invention automatically generates structured risk assessment reports through multi-level statistical and visualization aggregation, providing intuitive and reliable decision support for implementing differentiated and precise supervision. Attached Figure Description

[0017] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a schematic diagram of the overall process of the intelligent analysis method for special equipment inspection data based on natural language understanding provided in the embodiments of this application; Figure 2 This is a detailed diagram of the request parsing process provided in the embodiments of this application; Figure 3 This is a detailed diagram of the retrieval and model matching process provided in the embodiments of this application; Figure 4 This is a detailed diagram of the multi-source data synchronous extraction process provided in the embodiments of this application; Figure 5 This is a detailed diagram of the risk quantification calculation process provided in the embodiments of this application; Figure 6 This is a detailed flowchart of the unit risk aggregation analysis process provided in the embodiments of this application; Figure 7 This is a pie chart showing the distribution of equipment risk levels for target users, provided in an embodiment of this application. Figure 8 This is a detailed diagram of the hierarchical supervision suggestion generation process provided in the embodiments of this application; Figure 9 This is a schematic diagram of the structure of the intelligent analysis system for special equipment inspection data based on natural language understanding provided in the embodiments of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be described in detail and completely below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of this invention.

[0019] Please see Figure 1 , Figure 1 This is a schematic diagram of the overall process of the intelligent analysis method for special equipment inspection data based on natural language understanding provided in the embodiments of this application, which specifically includes the following steps: S1: Receive the unit risk assessment request input by the user in natural language, and extract the target unit name and risk analysis dimensions from the unit risk assessment request; In this embodiment, step S1 includes the following specific contents, and the process can be found in the attached document. Figure 2 , Figure 2 Here is a detailed diagram of the request parsing process provided in this application embodiment: S110: Receive input natural language text as a unit risk assessment request from the intelligent monitoring terminal and web portal system used by regulatory personnel; First, regulatory personnel deploy the risk assessment request text input in natural language on mobile terminals and web browser interfaces of office computers at the special equipment supervision site. Then, they receive the risk assessment request text input in natural language via Hypertext Transfer Protocol (HTTP), an application layer protocol for distributed collaborative information systems. Next, the received text stream undergoes Unicode conversion format octet encoding standardization processing. Unicode conversion format octet encoding is a variable-length character encoding for Unicode, a current technology capable of representing the vast majority of written language characters in the world. During the processing, non-compliant control characters and redundant spaces are removed, thus forming a standardized text sequence that meets the input requirements of subsequent natural language processing models. S120: Based on a pre-trained natural language understanding model, it performs dependency parsing and semantic role labeling on natural language text to obtain semantic labeling results; In this embodiment, for example, a standardized natural language text sequence is input into the BERT Chinese pre-training model. The specific process of obtaining semantic annotation results using the BERT Chinese pre-training model is as follows: First, the input text sequence is segmented into sub-words. Based on the special tags defined by the BERT Chinese pre-training model, the input text sequence is converted into an input tensor acceptable to the BERT Chinese pre-training model. Subsequently, the BERT Chinese pre-training model performs forward computation on the input tensor through its built-in multi-layer Transformer encoder to obtain the context vector of each word in the sequence. The quantification is represented as follows: During this process, the BERT Chinese pre-trained model simultaneously calls its integrated syntactic analysis and semantic role labeling functions. Based on the language knowledge obtained from pre-training and the domain adaptation ability obtained from fine-tuning, it automatically identifies the dependency relationships between lexical units to construct a syntactic tree and identifies the core predicates to label the semantic roles of related lexical units. Finally, the BERT Chinese pre-trained model outputs a structured sequence in which each lexical unit has a dependency relationship parent node index and a semantic role label. This structured sequence is the semantic annotation result obtained from parsing the natural language request, which can be directly used for subsequent extraction of target unit names and risk analysis dimensions. In this embodiment, for example, assuming the natural language request text input by the regulatory personnel is "Please analyze the overall risk status of the pressure vessels of XX Chemical Co., Ltd."; the standardized text sequence is input into the BERT Chinese pre-trained model, and the specific process of using the model to obtain semantic annotation results is as follows: First, the input text sequence is segmented into sub-words. Using BERT's Chinese word segmenter, the sentence is segmented into sub-word sequences at the character level, and a special marker "[CLS]" is added at the beginning of the sequence and a special marker "[SEP]" is added at the end of the sequence. For the input text "Please analyze the overall risk status of the pressure vessels of XX Chemical Co., Ltd.", the segmented sub-word sequence is: "[CLS]", "please", "analyze", "X", "X", "chemical", "industry", "limited", "company", "of", "pressure", "capacity", "vessel", "whole", "body", "wind", "risk", "state", "situation", "[SEP]". Each Chinese character is segmented into a separate sub-word. The "XX" in the company name is segmented into two independent "X" characters because it is not in the common character list, but the model will recognize them as part of the same entity in subsequent processing through context. Next, each subword is mapped to its corresponding index in the BERT vocabulary, resulting in a series of numerically represented input vectors. Simultaneously, an attention mask is generated for each subword to identify the position of a valid subword. In this example, all subwords are valid, so the mask for each position is set to the valid identifier. Since this example uses a single sentence input, there is no need to distinguish between sentence pairs; therefore, the sentence type identifier for each subword is set to the same value. These preprocessed data collectively constitute the model's input representation. The input representation is then fed into the BERT model. The BERT model consists of multiple Transformer encoders. Each layer captures the contextual dependencies between words through a self-attention mechanism and performs non-linear transformations through a feedforward network. After multiple layers of encoding, the output is a context vector representation corresponding to each word. These vectors fuse the semantic and syntactic information of the word in the sentence. Based on the BERT pre-trained model, the model was fine-tuned using domain corpora annotated with dependency syntax and semantic role information for the field of special equipment safety supervision. The fine-tuned model added a classification task to the output layer, which can predict the parent node of the dependency relationship, the dependency relationship type, and the semantic role label for each subword. Specifically, based on the context vector of each subword, the model calculates the probability distribution of its parent node position, the probability distribution of the dependency relationship type, and the probability distribution of the semantic role label through linear transformation, and selects the result with the highest probability as the final label. For the input sentence in this example, the semantic annotation results output by the BERT Chinese pre-trained model after calculation are as follows: The special marker "[CLS]" serves as the root node of the syntax tree, with its dependent parent node pointing to itself, and it does not carry a specific semantic role. The sub-word "analysis" is identified as the core predicate, with its dependent parent node pointing to "[CLS]", and its semantic role label is "core predicate". The sub-word "please" is identified as a modifier, with its dependent parent node pointing to "analysis", and its semantic role label is "modifier". The sub-word sequence "X", "X", "chemical", "industry", "have", "limited", "company", and "company" together constitute the organization name "XX Chemical Co., Ltd." The model labels this continuous segment as an organization entity, with the last sub-word "company" serving as the representative of this entity, and its dependent parent node pointing to... The semantic role of “analysis” is “patient”, and its entity type is “organization name”. Similarly, the sub-word sequence “pressure”, “force”, “capacity”, and “vessel” together construct the equipment name “pressure vessel”. The model labels it as a whole as an equipment entity, and the dependent parent node of the last sub-word “vessel” points to “analysis”, with the semantic role label “patient” and the entity type label “equipment name”. The sub-words “whole”, “risk”, and “condition” constitute a risk description phrase, where “risk” and “condition” are core nouns, with their dependent parent nodes pointing to “analysis” and the semantic role label “patient”, while “whole” is a modifier, with its dependent parent node pointing to either “risk” or “condition”. Finally, BERT's Chinese pre-trained model outputs a structured semantic annotation sequence. This sequence annotates each subword with its parent node position in the syntax tree, the type of dependency relationship with the parent node, and possible semantic roles and entity types. This structured result is the semantic annotation result parsed from the natural language request, which can be directly used in subsequent steps to extract the target user unit name (e.g., combining "XX Chemical Co., Ltd." from consecutive segments labeled "organization name") and risk analysis dimensions (e.g., summarizing the overall risk assessment requirements for pressure vessels from the combination of the equipment entity "pressure vessel" and the description "overall risk status"). S130: Based on semantic annotation results, identify and extract the target user unit name and risk analysis dimensions; In this embodiment, the structured semantic annotation sequence output by the Chinese pre-trained model of BERT is parsed, for example: Traverse all terms and term combinations in the sequence that are labeled as organizational entity categories, query external enterprise information knowledge bases for entity linking and disambiguation, normalize different expressions referring to the same entity, and determine and extract the unique target unit name. In the semantically labeled sequence, the semantic roles of risk type and analysis focus are located, and the corresponding core noun phrases are extracted. The core noun phrases are then matched and mapped with a pre-built domain risk dimension dictionary, which stores standardized risk dimension terms and their synonyms and near-synonyms. The mapping process is completed by calculating the cosine similarity between the word embedding vector of the core phrase and the word embedding vector of each term in the dictionary. The cosine similarity measures the degree of proximity of the two vectors in the direction. The standardized terms with the highest similarity are used as the mapping results, and the descriptive core phrases are mapped to standardized risk analysis dimension terms to complete the extraction of risk analysis dimensions; S2: Based on the target user's name, associate and retrieve all special equipment files under the target user's name, and construct an equipment analysis group; at the same time, based on the risk analysis dimensions, match the corresponding special equipment safety risk assessment model from the pre-set model library; In this embodiment, step S2 includes the following specific details, which can be found in the flowchart below. Figure 3 , Figure 3 Here is a detailed diagram of the retrieval and model matching process provided in the embodiments of this application: S210: Using the extracted target user's name as the query key, perform a search in the special equipment use registration database to obtain a list of equipment registration codes under the target user's name; The extracted standardized targets are linked to the special equipment use registration database using the unit name as the primary query key. A precise matching query is performed in the normalized field of the unit name in the special equipment use registration database. If no completely matching record is found, the fuzzy matching module is further activated. The fuzzy matching module uses a composite algorithm based on edit distance and pinyin similarity to calculate the similarity score between the query name and the record name in the special equipment use registration database, and filters out records with scores exceeding a preset threshold as successfully matched records. For example, in this embodiment, all special equipment registration codes marked as "in use", "disabled" and "pending inspection" are extracted from the successfully matched records to form a list of equipment registration codes under the name of the target user unit; S220: Based on the equipment registration code list, retrieve the complete technical files of each device in batches from the equipment lifecycle archive and build equipment analysis groups; Based on the obtained list of device registration codes, all device registration codes and necessary technical file field descriptions under the target user's name are organized and encapsulated in strict accordance with the query statement format required by the device lifecycle archive, and a batch data query transaction is constructed. Send a request to the device full - life - cycle archive through a batch data query transaction. The request encapsulates all device registration codes and descriptions of the required technical archive fields. Exemplarily, in this embodiment, the device registration codes and descriptions of the required technical archive fields include, but are not limited to, device model specifications, design parameters, manufacturing information, installation information, summaries of previous inspection reports, repair and renovation records, and safety accessory configuration lists; When the device full - life - cycle archive receives the request, it retrieves each device archive in parallel and returns the retrieval results in a package; After receiving the returned data packet, the retrieval matching module preliminarily classifies and organizes the devices according to information such as device type and the affiliated process unit, forming a structured device analysis group data set; S230: Convert the extracted risk analysis dimension text into a feature vector, calculate the similarity with the model description vectors in the preset model library, and match the most suitable risk assessment model as the special equipment safety risk assessment model; First, input the extracted normalized risk analysis dimension technical term text into the Sentence - BERT model; through this Sentence - BERT model, perform sub - word segmentation and vectorization processing on the extracted normalized risk analysis dimension technical term text, perform context encoding through its internal Transformer network layer, and perform mean pooling operation on the vector representation of the output sequence, finally generating a high - dimensional dense feature vector representing the overall semantics of the extracted normalized risk analysis dimension technical term text as the request vector for this matching; In this embodiment, exemplarily, assume that the normalized risk analysis dimension term extracted from the natural - language request is "pressure vessel corrosion risk"; first, input this technical term text into the Sentence - BERT model. This model is a siamese network based on the BERT architecture, specifically used to generate sentence - level semantic vector representations; the specific processing process is as follows: First, the tokenizer of the Sentence - BERT model performs sub - word segmentation on this technical term text; according to the character - granularity segmentation rule of the BERT Chinese model, "pressure vessel corrosion risk" is segmented into the following sub - word sequence: "压", "力", "容", "器", "腐", "蚀", "风", "险"; at the same time, add a special marker "[CLS]" at the beginning of the sequence and a special marker "[SEP]" at the end of the sequence, obtaining the complete input sub - word sequence: "[CLS]", "压", "力", "容", "器", "腐", "蚀", "风", "险", "[SEP]"; Next, each subword is mapped to the corresponding index number in the Sentence-BERT model vocabulary, resulting in a series of numerical input identifiers; at the same time, attention masks are generated to identify the positions of valid subwords (in this example, all subwords are valid, so the mask for each position is set to valid); since it is a single sentence input, there is no need to generate identifiers to distinguish sentence pairs; these preprocessed data together constitute the input tensor of the model; The input tensor is then fed into the Sentence-BERT model. The Sentence-BERT model consists of multiple Transformer encoders (12 layers in this embodiment, with each hidden layer having a dimension of 768), which capture the contextual dependencies between words through a self-attention mechanism. After multiple layers of encoding, the model outputs a context vector representation corresponding to each word, with each vector having a dimension of 768. These vectors integrate the semantic information of the words in the terminology. To obtain the overall semantic representation of the entire term text, mean pooling is performed on the vectors of all sub-words in the output sequence. This involves calculating the average value of all sub-word vectors across all dimensions to obtain a 768-dimensional vector. This vector is the high-dimensional dense feature vector representing the overall semantics of the term "pressure vessel corrosion risk," and serves as the request vector for this matching. This vector captures the core semantics of the term and can be used for subsequent similarity calculations with the model description vector. Next, using the same Sentence-BERT model and following the exact same processing flow, the descriptive text associated with each special equipment safety risk assessment model in the pre-built model library is sequentially encoded into a descriptive feature vector of the same dimension. The pre-built model library contains multiple risk assessment models, such as "Pressure Vessel Corrosion Risk Assessment Model," "Boiler Fatigue Damage Risk Assessment Model," and "Elevator Operation Risk Assessment Model." Each model has a natural language descriptive text, such as "This model is applicable to the risk assessment of wall thickness reduction caused by corrosive media in pressure vessels." This descriptive text is processed in the same way as the request vector: first, word segmentation is performed, [CLS] and [SEP] tags are added, mapped to input identifiers, and fed into the Sentence-BERT model for encoding. Then, the output sequence is averaged and pooled to obtain a 768-dimensional vector, which is the descriptive feature vector of the model. This process is repeated for all models to obtain the descriptive feature vector for each model. After generating all the descriptive feature vectors for all models, the cosine similarity algorithm is used to calculate the similarity score between the request vector and each descriptive feature vector in turn. The cosine similarity measures the degree of similarity between two vectors by calculating the cosine of the angle between them. Specifically, the calculation method is to divide the dot product of the two vectors by the product of their magnitudes. The closer the result is to 1, the more similar the two vectors are. Finally, all the calculated similarity scores are compared, and the preset risk assessment model corresponding to the highest similarity score is selected. This model is determined to be the special equipment safety risk assessment model that best matches the current risk analysis dimension, and is output for subsequent risk quantification calculation. The formula for calculating the matching degree of the risk assessment model is: ; In the formula, represents the matching score of the i-th risk assessment model, with a value range of 0 to 1. The higher the value, the higher the semantic matching degree between the model and the current risk analysis dimension; exp represents the natural exponential function, which is used to perform exponential operations with the natural constant e as the base. In this formula, it is used to convert the distance-based difference measure into a similarity score between 0 and 1. The smaller the distance, the closer the exponential function value is to 1. The request feature vector is a mathematical representation obtained by semantically encoding the extracted normalized risk analysis dimension terms. In this embodiment, the process of generating the request feature vector is as follows: For example, suppose the normalized risk analysis dimension term extracted from the natural language request is "pressure vessel corrosion risk"; First, the terminology text is input into the pre-trained semantic encoding neural network ERNIE model. This model is based on the ERNIE general Chinese model and is further trained using professional corpora such as special equipment inspection procedures, accident reports and risk case documents. It is specifically designed to generate vector representations that can characterize the professional semantics of the special equipment risk domain. First, the terminology text is segmented into sub-words. According to the character-level segmentation rules of the ERNIE Chinese model, "pressure vessel corrosion risk" is segmented into the following sub-word sequence: pressure, force, capacity, vessel, corrosion, erosion, wind, danger. At the same time, a special marker [CLS] is added at the beginning of the sequence and a special marker [SEP] is added at the end of the sequence to obtain the complete input sub-word sequence: [CLS], pressure, force, capacity, vessel, corrosion, erosion, wind, danger, [SEP]; All input text is standardized to a fixed length. In this embodiment, the model's preset input length is 128. Since the length of the input sequence in this example is 10, which is less than 128, a special padding character [PAD] needs to be added to the end of the sequence to make it complete. Therefore, the complete input sequence is: [CLS], pressure, force, capacity, vessel, corrosion, wind, danger, [SEP], and the following 118 [PAD], for a total of 128 tokens. Convert each token into a numerical representation that the model can recognize; the ERNIE model internally maintains a vocabulary, and each token has a unique index number in the vocabulary; map each token to its corresponding index number to obtain a sequence of index numbers, which is called the input identification sequence; for example, [CLS] corresponds to index 101, each specific Chinese character corresponds to its respective index, [SEP] corresponds to index 102, and the padding symbol [PAD] corresponds to index 0; a numerical sequence with a length of 128 is obtained, and the first 10 positions are successively the indices corresponding to [CLS], pressure, vessel, corrosion, risk, [SEP], and the subsequent 118 positions are all the index 0 of the padding symbol; At the same time, in order to let the model know which numbers represent real text content and which numbers are padding parts, a corresponding identification sequence also needs to be generated, and the length of the sequence is also 128; among them, the first 10 positions are set to 1, indicating that these positions are real and valid tokens, and the subsequent 118 positions are set to 0, indicating that these positions are padding meaningless content; The above two sequences together constitute the input representation of the model; Subsequently, the input representation is fed into the pre-trained semantic encoding neural network ERNIE model; this model consists of multiple layers of Transformer encoders. In this embodiment, the ERNIE 3.0 architecture is adopted, which includes 12 layers of Transformer encoders, the dimension of each hidden layer is 768, and the number of self-attention heads is 12; The input representation first passes through the word embedding layer to convert the index number of each token into its corresponding initial vector representation, and the dimension of each vector is 768. Then, these vectors sequentially pass through each layer of the Transformer encoder. In each layer, the self-attention mechanism is used to capture the semantic dependencies between tokens, and a non-linear transformation is performed through the feed-forward network; after 12 layers of encoding, the model outputs the context vector representation corresponding to each token, and the dimension of each vector is still 768; these vectors integrate the deep semantic information of the tokens in the term, fully reflecting the professional understanding of the terms related to the risks of special equipment after domain adaptation training; In order to obtain the overall semantic representation of the entire term text, a pooling operation needs to be performed on the vectors of all tokens; in this embodiment, average pooling is adopted, that is, the average value of the vectors of all real tokens (excluding padding symbols) in the output sequence is calculated in each dimension, and finally a 768-dimensional vector is obtained; specifically, the model will perform dimension-by-dimension averaging on the 768-dimensional vectors of the first 10 real tokens to obtain a 768-dimensional real number vector; this vector is the high-dimensional dense feature vector representing the overall semantics of the term "pressure vessel corrosion risk" and serves as the request feature vector; this high-dimensional dense real number vector is the request feature vector ; The description feature vector of the i-th model is a mathematical representation obtained by semantically encoding the natural language description text associated with the i-th risk assessment model in the pre-set model library. In this embodiment, the generation process of the description feature vector is exactly the same as that of the request feature vector. The same pre-trained semantic coding neural network ERNIE model is used to encode the description text of the risk assessment model, thereby transforming the functional semantic connotation of the risk assessment model into a high-dimensional dense real number vector of the same dimension. Represents the request feature vector The descriptive feature vector of the i-th model The Mahalanobis distance between them; in this embodiment, the Mahalanobis distance is calculated as follows: First, the request feature vector is calculated. Feature vectors describing the target model The difference vector between them; secondly, obtain the inverse matrix of a pre-calculated covariance matrix, which is statistically calculated based on the set of descriptive feature vectors of all models in the model library. In this embodiment, the specific calculation method is as follows: calculate the mean of all values ​​in each feature dimension in the vector set, and then calculate the covariance between every two different feature dimensions to form a square matrix. This square matrix describes the degree of dispersion of these feature vectors in each semantic dimension and the degree of linear correlation between different semantic dimensions; then, the above-calculated request feature vectors are used. Feature vectors describing the target model The difference vector between them is multiplied on the left by the inverse matrix, and then multiplied on the right by the eigenvector requested for computation. Feature vectors describing the target model The transpose of the difference vector between the two values ​​yields an intermediate scalar value; finally, the square root of this intermediate scalar value is taken, and the result is the Mahalanobis distance. σ σ is a bandwidth parameter used to control the rate at which the similarity score decays as the Mahalanobis distance increases. Its value is determined through cross-validation. In this embodiment, the specific method is as follows: During the model development phase, a training dataset containing a large number of known correct matching relationships is collected. This dataset is randomly divided into k mutually exclusive subsets, and each subset is used as a validation set, while the remaining k-1 subsets are used as the training set. The covariance matrix is ​​obtained on the training set, and for a series of candidate σ values, the matching score of each sample in the validation set is calculated, and a model selection decision is made to evaluate the decision accuracy. Finally, the candidate σ value with the highest average decision accuracy in k rounds of validation is used as the fixed bandwidth parameter when deploying the model. After calculating the matching score of all models, the risk assessment model with the highest matching score is selected as the matching special equipment safety risk assessment model. S3: For each device in the device analysis group, extract multi-source data related to the risk analysis dimensions simultaneously from the historical inspection database, operation monitoring platform and regulatory record library, combined with the changes in the scenario; In this embodiment, step S3 includes the following specific details, which can be found in the flowchart below. Figure 4 , Figure 4 Here is a detailed flowchart of the multi-source data synchronous extraction process provided in the embodiments of this application: S310: Based on the risk analysis dimension, determine the corresponding equipment status change scenarios and extract the analysis time range and key data features; Based on the matched risk assessment model type and the input risk analysis dimension terms, a predefined scenario feature mapping knowledge base is invoked. The construction of this predefined scenario feature mapping knowledge base is based on the systematic analysis and summarization of massive historical special equipment inspection data, accident reports, safety technical specifications and operation and maintenance records. From this, the association rules between different risk dimensions and equipment status, time patterns and data characteristics are abstracted, and these rules are stored in a structured form in the database table. Each record clearly contains the risk analysis dimension, equipment status change scenario, time range calculation logic and key data feature identifier list. By retrieving this predefined scenario feature mapping knowledge base based on the input risk analysis dimension, the length of the historical time window required for this analysis can be directly obtained, and a targeted list of key data features can be generated. S320: For each device in the device analysis group, initiate three data extraction requests simultaneously; For each device in the device analysis group, the unique device registration code of each device is used as the core association identifier. Combined with the determined analysis time range and key data feature list, data extraction requests are sent concurrently to three independent data sources. The first extraction request is sent to the historical inspection database. The query conditions include the device registration code, time range, and explicitly specify the type and feature fields of the inspection items to be extracted. The second extraction request is sent to the operation monitoring platform, requesting to obtain the original time-series data or preprocessed statistical feature data of the sensors related to key features of the specified device within a given time range in a streaming query manner; The third extraction request is sent to the regulatory record database, requesting to retrieve regulatory records in the document database that are related to the device and involve key feature words in the cause within a specified time range; S330: Integrate the received multi-source data to form a structured multi-source data packet; Once responses are received from the three data sources, the data integration process is initiated. Establish a unified data framework with time as the baseline; Align discrete test records from the historical test database onto the timeline based on the test date; The continuous time-series data from the operation monitoring platform is interpolated with timestamps and then filled into the frame; the document records from the regulatory record database are extracted, and their issuance dates and key contents are used as event points and associated with corresponding dates; For situations where multiple data sources exist at the same point in time, establish references and relationships between the data; For each device, a structured multi-source data packet combining time series and event series is generated. This structured multi-source data packet fully contains all heterogeneous data related to the specified risk analysis dimension within the analysis time window. S4: Based on the special equipment safety risk assessment model, quantitative calculations are performed on the multi-source data of each piece of equipment to generate a single equipment risk score and risk factor label; In this embodiment, step S4 includes the following specific details, which can be found in the flowchart below. Figure 5 , Figure 5 Here is a detailed flowchart of the risk quantification calculation process provided in the embodiments of this application: S410: Preprocess the integrated multi-source data to transform it into a standard feature vector; Feature engineering is performed on the structured multi-source data packets for each device; these structured multi-source data packets aggregate heterogeneous data from historical inspection databases, operational monitoring platforms, and regulatory record databases; and different feature extraction and construction methods are used for different types of data. For continuous sensor time-series data from the operation monitoring platform, such as temperature, pressure, and flow time-series data, a set of statistics is calculated within a pre-defined time sliding window within the analysis time range determined based on the risk analysis dimension. In this example, taking temperature time-series data as an example, the set of statistics includes the arithmetic mean of temperature data points within the time sliding window, the standard deviation of temperature data points within the time sliding window, the skewness of the temperature data point distribution within the time sliding window, the kurtosis of the temperature data point distribution within the time sliding window, the maximum value of temperature data points within the time sliding window, the minimum value of temperature data points within the time sliding window, and the number of times temperature data points exceed a preset safety threshold and the total duration of exceeding the safety threshold within the time sliding window. The same calculation process is performed for other continuous sensor time-series data such as pressure and flow time-series data. For numerical inspection record data from a historical inspection database, such as wall thickness measurement record data and hardness measurement record data, extract specific statistics of the numerical inspection record data within the analysis time range, including the measurement value of the last record, the measurement value of the first record, the number of records, the average value of all recorded measurement values, the standard deviation of all recorded measurement values, and the slope of the trend of the measurement value changing with time obtained by linear regression fitting; For text record data from a regulatory record repository, such as supervision and inspection records, notice of order for rectification, or administrative penalty decision for pressure vessels, boilers, or elevators, use an existing lightweight pre-trained language model ERNIE-Tiny to convert these text records into low-dimensional dense real number vectors; in this embodiment, the specific conversion process is as follows: First, preprocess each regulatory text that records the registration code of special equipment, the facts of violations, and the disposal conclusions; exemplarily, assume there is a regulatory record text: "On March 15, 2025, it was found during inspection that there was local corrosion in pressure vessel R101 of XX Chemical Plant, and it was ordered to rectify within a time limit." Perform Chinese word segmentation on this text to obtain a sequence of words: "March 2025", "March", "15th", "inspection", "found", "XX", "Chemical Plant", "pressure vessel", "R101", "exists", "local", "corrosion", "ordered", "time limit", "rectify"; Remove stop words with less semantic contribution such as "exists" and "found", and retain key terms, obtaining the processed sequence of words: "inspection", "XX Chemical Plant", "pressure vessel", "R101", "local", "corrosion", "ordered", "time limit", "rectify"; According to the requirements of the ERNIE-Tiny model, all texts need to be unified to a fixed length, which is 128 tokens in this embodiment; since the length of the word sequence in this example is less than 128, special padding symbols need to be added at the end of the sequence to complete the length; if the length exceeds 128, truncation is performed; and a special marker [CLS] is added at the beginning of the sequence, and a special marker [SEP] is added at the end of the sequence to form a complete input sequence; Next, each token needs to be converted into a numerical representation that the model can recognize. The ERNIE-Tiny model maintains a vocabulary, and each token has a unique index number in the vocabulary. Therefore, mapping each token to its corresponding index number results in a numerical sequence composed of index numbers, which is called the input identifier sequence. For example, [CLS] corresponds to index 101, each specific word corresponds to its own index, [SEP] corresponds to index 102, and the padding character corresponds to index 0. Thus, a numerical sequence of length 128 is obtained. The first 11 positions are the indices corresponding to [CLS], inspection, XX chemical plant, pressure vessel, R101, local, corrosion, order, deadline, rectification, and [SEP], respectively. The subsequent 117 positions are all the index 0 of the padding character. Meanwhile, in order for the model to know which numbers represent real text content and which numbers are padding, a corresponding identifier sequence needs to be generated. This identifier sequence is also 128 in length, with the first 11 positions set to 1 to indicate that these positions are real and valid tokens, and the last 117 positions set to 0 to indicate that these positions are meaningless padding content. When processing, the model will only focus on the positions marked as 1 and ignore the padding. The two sequences mentioned above together constitute the input representation of the model; then, the input representation is fed into the ERNIE-Tiny model; this model contains 3 Transformer encoder layers, each with a hidden layer dimension of 256; The input representation is first converted into a vector through a word embedding layer, and then passed through each Transformer encoder layer in turn. In each layer, the semantic dependencies between tokens are captured through a self-attention mechanism, and then a non-linear transformation is performed through a feedforward network. After three layers of encoding, the model outputs a context vector representation for each token, with each vector having a dimension of 256. These vectors incorporate the semantic information of the token in the text. To obtain the overall semantic representation of the entire regulatory record, this embodiment takes the output vector corresponding to the special marker [CLS] to obtain a 256-dimensional real vector, which serves as the semantic representation vector for the regulatory record. This semantic representation vector incorporates key information directly related to equipment risk in the text, such as the equipment name "Pressure Vessel R101", the defect type "Local Corrosion", and the handling requirement "Order to Rectify within a Time Limit". It can be directly used as a feature to characterize management compliance risk and is concatenated with the inspection data features and operation monitoring data features from the same equipment. These features are then input into the subsequent special equipment safety risk assessment model for comprehensive risk calculation. All numerical statistics from the operation monitoring platform and historical inspection database are standardized so that the mean of each statistic across all device samples is 0 and the variance is 1. For missing values ​​in the data, different imputation strategies are used according to the data type: linear interpolation based on the time series trend is used for continuous time series data; imputation based on the mean of data from similar devices is used for discrete detection data; and a preset vector representing "no event" is used to imput the missing values ​​for text data. All processed sensor time-series statistical features, detection record statistical features, and text embedding vectors are concatenated according to a predetermined global feature order to form a unified, high-dimensional standard feature vector, which will characterize the overall status of the device under the specified risk analysis dimension in subsequent steps. S420: Input the standard feature vector into the matching special equipment safety risk assessment model to calculate the risk score of a single device; The standard feature vector is input into the matching special equipment safety risk assessment model to calculate the risk score of a single device; The formula for calculating the risk score of a single device is: ; In the formula, R represents the calculated risk score of a single device, which is a dimensionless scalar between 0 and 1, used to quantitatively characterize the relative risk level of the device; exp represents the natural exponential function, whose core function is to smoothly and monotonically map a real number input to the interval between 0 and 1, realizing the probabilistic and normalized output of the risk score; M represents the total number of nonlinear hidden nodes in the special equipment safety risk assessment model. The total number of nonlinear hidden nodes is pre-defined manually in the early stage of the construction of the special equipment safety risk assessment model as a structural hyperparameter. This represents the connection weight from the j-th hidden node to the output node, and is a real scalar. This connection weight is calculated during the training phase of the special equipment safety risk assessment model using a large amount of standard feature vector data from historical equipment, employing an error backpropagation algorithm to compute the loss function between the predicted risk score and the true label. The partial derivatives are the gradients, and the stochastic gradient descent optimization algorithm is used to iteratively update the gradient in the reverse direction. The values ​​are calculated until the overall loss function of the special equipment safety risk assessment model converges to its minimum value, and then finally determined. The backpropagation algorithm is an algorithm used for neural network training. It propagates the error of the final output layer layer by layer using the chain rule, calculating the contribution of each parameter in the network to the total error, i.e., the gradient. The stochastic gradient descent optimization algorithm is an iterative optimization method that updates parameter values ​​according to the calculated gradient direction with a preset learning rate step size, minimizing the loss function through repeated iterations. j is the index number of the hidden node, used to distinguish different hidden nodes in the special equipment safety risk assessment model; tanh represents the hyperbolic tangent activation function. The function of the number is to compress and map the linearly weighted input value of each hidden node to a range of negative one to positive one, introducing nonlinear transformation capability into the special equipment safety risk assessment model, enabling the special equipment safety risk assessment model to fit complex feature relationships; F represents the standard feature vector, which is obtained by feature engineering processing of multi-source heterogeneous data. In this embodiment, it specifically includes calculating window statistics for sensor time series data, extracting key indicators from inspection and testing records, semantically embedding and encoding regulatory text, and standardizing all feature components to make the mean 0 and the variance 1, forming a high-dimensional real number vector with dimensionless pure numbers in each dimension; It is a weight vector whose dimension is strictly equal to the dimension of the input standard feature vector F. It represents the specific weight connecting each feature dimension of the input standard feature vector F to the j-th hidden node. Each element of this weight vector, i.e., each weight value, in this embodiment, is... Each element value is obtained through a training method deeply integrated with the special equipment risk quantification analysis process. The specific process is as follows: First, a training dataset is prepared, which consists of a large number of historical equipment standard feature vectors generated according to the multi-source data fusion rules of this method. The real risk label associated with each feature vector is strictly labeled based on objective facts such as the defect level, safety fault records, and whether an accident has occurred as found in the subsequent inspection conclusions of the corresponding equipment. At the beginning of the training of the special equipment safety risk assessment model, each element of the weight vector is initialized to a random small value. Then, an iterative optimization process for special equipment risk assessment is entered. In each training batch, forward propagation is performed to calculate the equipment risk prediction score based on the current weights, and the mean square error loss between the equipment risk prediction score and the real risk label is calculated. This mean square error loss directly reflects the degree of misjudgment of the equipment safety status by the special equipment safety risk assessment model. Then, error inversion is used. The backpropagation algorithm, starting from the output layer, calculates the gradient of the loss function with respect to each special equipment risk feature dimension associated with the weight vector, such as corrosion rate and frequency of violation records, layer by layer using the chain rule. A stochastic gradient descent optimization algorithm is applied, with a preset learning rate step size, synchronously updating the weight values ​​corresponding to each of the aforementioned special equipment risk feature dimensions in the opposite direction of the gradient. This update operation directly affects the evaluation weights of the special equipment safety risk assessment model for each of the aforementioned specific risk features of special equipment. Through iteration, the special equipment safety risk assessment model gradually adjusts its quantitative assessment of the importance of different features such as corrosion rate, overpressure frequency, or violation records, thereby learning a risk judgment pattern closely related to special equipment safety. This iterative training process is repeated until the special equipment safety risk assessment model achieves a stable accuracy in risk assessment of various special equipment in the validation set, at which point training is complete. The values ​​of all elements of the weight vector are then determined; superscript... The transpose operation represents the transformation of a vector or matrix. This is a basic linear algebra operation used to convert a row vector into a column vector or vice versa. Represents the weight vector The inner product operation is performed with the standard eigenvector F, that is, the corresponding elements of the two vectors are multiplied and then summed to obtain a scalar value, which is used to calculate the linear combination of eigenvectors; This is the bias term of the j-th hidden node, a real scalar. The value of this bias term is calculated during the training of the special equipment safety risk assessment model using standard feature vector data of historical equipment and real risk labels, through the error backpropagation algorithm to calculate the loss function. The gradient is then used to iteratively update the bias term using the stochastic gradient descent optimization algorithm with a preset learning rate step size. The value of the bias term is determined after the special equipment safety risk assessment model has been trained and converged. Its function is to adjust the threshold for activation of hidden nodes. This is a correction term based on feature uncertainty. This correction term is obtained by calculating the product of the absolute value of the partial derivative of the intermediate output of the special equipment safety risk assessment model with respect to each input feature component and the uncertainty of each feature component itself, and then summing the products over all feature components. The uncertainty of each feature component originates from prior information such as the accuracy of the sensor used to acquire the feature component data, the inherent error of the testing method, and data gaps. The function of this correction term is to quantify the inherent measurement error of each feature data and transfer it to the risk calculation, so that the risk score reflects the reliability of the data. τ(F) is an uncertainty adjustment factor, with a value of a positive real number. In this embodiment, the calculation process of this uncertainty adjustment factor is as follows: First, obtain the uncertainty corresponding to each feature component in the standard feature vector F. This uncertainty is then used to calculate the correction term. The uncertainty values ​​of each characteristic component used are the same. Next, the uncertainty of each characteristic component is squared, and these squared values ​​are summed. Then, the square root of this sum is taken to obtain a scalar representing the overall unreliability of the input data, which is the total uncertainty. Subsequently, this total uncertainty is multiplied by a normal number pre-set based on the performance and historical data verification of the special equipment safety risk assessment model; this normal number is called the adjustment coefficient. Finally, the product is added to a constant, and the sum is the final value of the uncertainty adjustment factor τ(F). This calculation... The computational logic ensures that when the overall quality of the data represented by the standard eigenvector is high and the uncertainty of each eigencomponent is low, the value of τ(F) approaches one, which has no significant impact on the risk assessment calculation. When the overall quality of the data represented by the standard eigenvector is poor and the uncertainty of each eigencomponent is high, the value of τ(F) will be significantly greater than one, which increases the denominator of the input part of the natural exponential function in the calculation formula, resulting in a decrease in the overall score value. This drives the final risk score output to converge toward the probability median of 0.5, enabling the model to automatically provide a more prudent assessment result when the data credibility is low, thus improving the robustness of the risk assessment. For example, in this embodiment, the specific structure and training parameter settings of the special equipment safety risk assessment model are as follows: Model input layer: The standard feature vector F has a dimension of N=114, of which the sensor time-series statistical features are 30-dimensional, the detection record statistical features are 20-dimensional, and the regulatory text semantic embedding features are 64-dimensional. Hidden layer: A single hidden layer neural network structure is adopted, with the number of hidden nodes M=128; this value is optimally determined on the validation set through grid search, and the value range is generally from 64 to 256; the activation function of the hidden layer is the hyperbolic tangent function; Output layer: Single-node output, the weighted summation result is mapped to the interval between 0 and 1 through the natural exponential function exp, which is equivalent to the sigmoid activation function, and directly outputs the risk score R of a single device; Training data partitioning: Equipment samples that have completed their entire lifecycle and have clear post-event verification (such as accident records and serious defect reports) are selected from historical inspection databases, operation monitoring platforms, and regulatory record databases to construct training sets, validation sets, and test sets, with proportions of 70%, 15%, and 15%, respectively. Loss function: The mean squared error is used as the loss function to measure the deviation between the predicted risk score and the true label; Optimizer and learning rate: The Adam optimizer is used, with an initial learning rate of 0.001 and a learning rate decay strategy, where the learning rate is decayed to 0.9 every 10 training epochs. Regularization strategy: Add a Dropout layer after the hidden layers, with a Dropout ratio of 0.3, randomly discarding 30% of the neuron outputs to prevent overfitting; simultaneously, adjust the weight vector... and connection weights Apply L2 regularization with a regularization coefficient of 0.0001; Early stopping strategy: Monitor the validation set loss during training. If the validation loss does not decrease for 10 consecutive training rounds, terminate training early and restore the optimal model weights. Batch size: Set to 32; Model evaluation metrics: After the model is trained, its performance is evaluated on the test set. The coefficient of determination should be greater than 0.85 and the mean absolute error should be less than 0.05. Only models that meet the above metrics can be deployed for actual risk calculation to ensure the accuracy of risk quantification. S430: Based on the feature importance output by the special equipment safety risk assessment model, identify key risk factors and generate risk factor labels; While the risk assessment model calculates the score, exemplarily in this embodiment, an integrated interpretability analysis module calculates the contribution of each feature component in the standard feature vector to the final score. This integrated interpretability analysis module is based on the Shapley value framework in cooperative game theory, treating the set of all feature components as a cooperative alliance, and the predicted score of the special equipment safety risk assessment model as the total benefit generated by the alliance. Therefore, when calculating the contribution of a feature component to be evaluated in the standard feature vector, it is necessary to consider the marginal benefit change brought about by adding this feature component to all possible subsets composed of other feature components. Specifically, when calculating the contribution of each feature component in the standard feature vector to the final score... When assessing the contribution of a feature component, the integrated interpretability analysis module enumerates all possible combinations of the feature component to be evaluated with other feature components in the standard feature vector. For each combination, it calculates the change in the model's predicted score before and after the addition of this feature component, i.e., the marginal contribution. Finally, it performs a weighted average of the marginal contributions for all possible combinations, and the resulting mean is the Shapley value of this feature component, i.e., the contribution of this feature component. This calculation method ensures a fair distribution of the contribution. The absolute value of the contribution precisely quantifies the degree to which this feature component affects the risk score. A positive sign indicates that an increase in the value of this feature component will increase the risk score, while a negative sign indicates that an increase in the value of this feature component will decrease the risk score. Based on the calculated contribution, the absolute values ​​of the contribution of all feature components are sorted in descending order, and the top K feature components are selected as key risk features. To transform these key risk features into understandable risk semantics, they need to be mapped to standardized risk factor labels. In this embodiment, this mapping process is achieved by querying a preset feature-risk factor mapping rule base. The construction of this rule base involves a systematic analysis of a large amount of historical special equipment inspection data, accident cases, and safety technical specifications, automatically summarizing and solidifying the correspondence between various measurable features and standard risk factors, as well as the numerical triggering conditions. Then, the corresponding rules are retrieved based on the identifier of each key risk feature, and the specific value of the feature in the current equipment instance is used to determine whether the triggering conditions are met. For all key risk features that meet the triggering conditions, associated semantic risk factor labels are output, thereby converting these key risk feature sets into a set composed of risk factor labels, clearly representing the main risk composition of the equipment in the current risk assessment. S5: Statistically analyze and rank the individual risk scores of all devices in the equipment analysis group, and perform aggregate analysis on the risk factor labels to identify the overall risk level, dominant risk type, and distribution characteristics of high-risk devices of the target user unit. In this embodiment, step S5 includes the following specific details, which can be found in the flowchart below. Figure 6 , Figure 6 Here is a detailed flowchart of the unit risk aggregation analysis process provided in the embodiments of this application: S510: The average risk score of all individual devices in the statistical device analysis group and the proportion of high-risk devices; To assess the overall risk status of the target user, it is necessary to perform statistical analysis on the individual risk scores of all devices in the equipment analysis group; The arithmetic mean of the risk scores of all individual devices in the device analysis group is used to measure the average risk level of that unit device. Simultaneously, the standard deviation of the risk scores of all individual devices in the device analysis group is calculated to measure the dispersion of risk among devices. A dynamic high-risk equipment identification threshold is set. This high-risk equipment identification threshold is not a fixed value, but is dynamically determined based on the distribution characteristics of the risk scores of the current equipment group. In this embodiment, the specific method for determining the high-risk equipment identification threshold is to add the product of the standard deviation of the risk scores of all individual equipment in the equipment analysis group and a control coefficient to the arithmetic mean of the risk scores of all individual equipment in the equipment analysis group. The control coefficient is a positive number preset according to regulatory policy requirements and risk tolerance, used to adjust the strictness of high-risk equipment screening. The proportion of high-risk devices can be obtained by dividing the number of devices in the device analysis group whose risk scores exceed the high-risk device identification threshold by the total number of devices in the device analysis group. S520: Determine the overall risk level of the target user unit based on the risk level threshold mapping table; To determine the overall risk level of the target user unit, it is first necessary to calculate the comprehensive risk score of the target user unit. The formula for calculating the comprehensive risk score of the target user unit is as follows: ; In the formula, S represents the overall risk score of the target user unit, which is a dimensionless value between 0 and 1. The higher the value, the higher the overall risk. This represents the arithmetic mean of the risk scores of all individual devices in the device analysis group. Since the risk score of an individual device is itself a normalized dimensionless quantity, the arithmetic mean... Also a dimensionless quantity between 0 and 1; The high-risk equipment ratio represents the proportion of high-risk equipment, i.e., the ratio of the number of devices with risk scores exceeding the high-risk equipment identification threshold to the total number of devices in the target user unit. Its value ranges from 0 to 1 and is dimensionless. α is a weighting coefficient between 0 and 1, and its specific value is determined through a data-driven method. In this embodiment, the specific method is as follows: During the special equipment safety risk assessment model construction phase, a dataset containing a large number of historical user unit cases is collected. Each case includes the average risk score of the target user unit. Proportion of high-risk equipment And the known overall risk level determined by historical regulatory records and objective security incidents; using this data, to and Using the overall risk level as the output target and the input features as the input features, a multiple linear regression model is trained. The training process of this multiple linear regression model solves for two features by minimizing the error between the model's predicted risk level and the actual risk level. and Their respective regression coefficients; average risk score The corresponding regression coefficient is divided by the sum of the two regression coefficients, which is to normalize the coefficient. The final ratio is then determined as the value of the weight coefficient α. This method ensures that the value of α is based entirely on historical objective data, reflecting the actual statistical impact weight of the two factors, average risk and high-risk clustering, on the final level determination. After obtaining the comprehensive risk score S, the overall risk level of the target user unit is determined by querying a pre-set risk level threshold mapping table. The construction of this risk level threshold mapping table is based on historical objective data, and the specific process is as follows: Collect comprehensive risk scores from a large number of historical users and their verifiable security records from the same period; The K-means clustering algorithm was used to analyze these historical score data. The goal of the K-means clustering algorithm is to divide all data points into four clusters corresponding to four preset levels: "low risk", "medium risk", "high risk" and "high risk". The algorithm continuously optimizes the position of the cluster center and the affiliation of each data point through iterative calculation, so that the scores of data points within the same cluster are as close as possible, while the score differences between different clusters are as significant as possible. These boundary values, which are automatically optimized by the K-means clustering algorithm to distinguish different clusters, are determined as the segmentation thresholds of the mapping table. This process ensures that the threshold division is driven entirely by the data distribution pattern, thereby objectively and reasonably mapping the continuous comprehensive risk score to the preset discrete risk level. By comparing the current unit's comprehensive risk score S with the threshold range in the mapping table, the overall risk level of the target unit can be determined. S530: Statistical analysis of the frequency and combination patterns of risk factor labels to identify the dominant risk type; Summarize the risk factor labels generated by all devices in the device analysis group; count the frequency of each individual risk factor label. Simultaneously, the co-occurrence of multiple labels is analyzed, and the frequency of each pair or group of risk factor labels appearing simultaneously on the same device is statistically analyzed. In order to identify the dominant risk type, the pattern strength value of each risk factor label combination needs to be calculated. The pattern strength value is obtained by multiplying the co-occurrence frequency of the risk factor label combination by the geometric mean of the independent occurrence frequencies of each label in the risk factor label combination. The label combination with the highest pattern strength value is identified as the dominant risk type for the target user unit under the current risk analysis dimension; S540: Sort the equipment by risk, identify high-risk equipment and analyze the distribution characteristics of high-risk equipment; To identify high-risk equipment and analyze its distribution characteristics, all equipment needs to be sorted in descending order based on the risk score of each individual device to form an ordered sequence of equipment risks. Devices with risk scores exceeding the high-risk device identification threshold are identified as high-risk devices, and a list of high-risk devices is generated. A multidimensional feature analysis was conducted on the equipment in the high-risk equipment list, including the distribution of these equipment in terms of equipment type, process unit, geographical location, years of service, and previous inspection results. By calculating the ratio of the number of high-risk devices in each dimension to the total number of devices in that dimension within the group, significant clustering characteristics of high-risk devices can be identified. To more intuitively illustrate the overall distribution of risk levels for equipment used by the target user unit, please refer to [link / reference]. Figure 7 , Figure 7 This is a pie chart showing the risk level distribution of equipment in a target user unit, provided in this embodiment. The pie chart illustrates the risk level distribution of equipment in a target user unit as analyzed in this embodiment. Different risk level areas are visually distinguished using differentiated patterns: low-risk areas are filled with a dotted pattern, medium-risk areas with black fill, higher-risk areas with a diagonal line pattern, and high-risk areas with a grid pattern. The size and proportion of each sector visually reflect the relative distribution of equipment at each risk level within the equipment analysis group, providing a visual basis for differentiated supervision. S6: Based on the overall risk level, dominant risk type, and distribution characteristics of high-risk equipment of the target user unit, output differentiated regulatory recommendations for the target user unit; In this embodiment, step S6 includes the following specific details, which can be found in the flowchart below. Figure 8 , Figure 8 This is a detailed flowchart of the hierarchical regulatory suggestion generation process provided in the embodiments of this application: S610: Generate a corresponding overall regulatory intensity recommendation based on the overall risk level of the target user unit; Targeted access to a pre-built regulatory strategy knowledge base, which is constructed through systematic analysis of historical regulatory effectiveness data, legal and policy documents, and typical industry practice cases, clearly defines the basic regulatory resource allocation and action baseline corresponding to different overall risk levels; Using the overall risk level of the identified target user as the query condition, the corresponding rule entries are retrieved and matched; an overall regulatory intensity recommendation under the overall risk level of the target user is automatically generated; for example, in this embodiment, when the overall risk level is "higher risk", the generated recommendation is to clearly increase the frequency of annual statutory supervision and inspection of the target user, require an increase in the proportion of manual review of online monitoring data, and automatically include the target user in the key focus list for the next cycle of special governance, thereby determining the overall intensity of regulatory intervention and the priority of resource investment at the strategic level; S620: Generate corresponding specific risk assessment recommendations based on the primary risk type of the target user unit; Based on the identified dominant risk types, the governance experience base for specific risk patterns in the regulatory strategy knowledge base is queried again. The construction of this governance experience base is based on the structured sorting and summarization of a large number of historical special equipment inspection cases, failure handling reports and rectification plans after effectiveness verification. Based on the identified dominant risk type, a specific risk assessment suggestion list is automatically generated. This risk assessment suggestion list clearly indicates the equipment parts that need to be focused on during the next stage of on-site inspection, the recommended testing technologies, and the specific process parameter records that need to be reviewed, transforming the abstract "risk type" into technical actions that can be performed on-site. S630: Based on the distribution characteristics of high-risk equipment of the target user unit, generate corresponding precise supervision suggestions for high-risk equipment; Based on the analyzed distribution characteristics of high-risk equipment, precise guidance is provided at both the spatial and object levels. The generated specific recommendations are directly linked to the list of high-risk equipment and its location information. For example, in this embodiment, an enhanced patrol plan for a specific area is generated based on spatial clustering characteristics. Differentiated enhanced monitoring and inspection strategies are formulated for each high-risk equipment in the list, such as adjusting independent monitoring parameter thresholds and associating specific historical data as inspection references. Based on other significant characteristics of the equipment group, the preferred inspection strategy type is suggested. This achieves the precise allocation and deployment of regulatory resources and attention to the most specific risk targets. Please see Figure 9 , Figure 9This is a schematic diagram of the structure of the intelligent analysis system for special equipment inspection data based on natural language understanding provided in the embodiments of this application; This embodiment demonstrates the overall system architecture for implementing the above method. The system forms a complete technical chain from natural language instruction parsing to intelligent regulatory suggestion generation through the collaboration of six core modules. The request parsing module, as the system's input interface, is responsible for receiving natural language requests and performing deep semantic parsing through an integrated pre-trained language model to accurately extract the target user's name and risk analysis dimensions. The retrieval and matching module is responsible for resource scheduling. This module retrieves and constructs equipment analysis groups based on the extracted unit names. At the same time, based on the risk analysis dimensions, it selects the most suitable special equipment safety risk assessment model from the pre-set model library through semantic matching. The data extraction module is a multi-source data integration engine; for each device in the device analysis group, based on the data characteristics and time range determined by the risk analysis dimensions, it concurrently extracts relevant heterogeneous data from the historical inspection database, operation monitoring platform and regulatory record library, and aligns and integrates them. The risk calculation module is the core analysis unit; this module calls the matching risk assessment model to perform feature processing and quantitative calculation on the multi-source data integrated by each device, and outputs the risk score of a single device and its risk factor label. The aggregation analysis module is the intelligence generation center; this module performs statistical analysis on the risk scores of individual equipment across all equipment to determine the overall risk level and the proportion of high-risk equipment, performs frequency and co-occurrence analysis on risk factor tags to identify the dominant risk type, and performs characteristic analysis on the spatial, type, and service life distribution of high-risk equipment. The suggestion generation module is the decision output terminal. Based on the overall risk level, dominant risk type and high-risk equipment distribution characteristics provided by the aggregation analysis module, this module automatically generates and outputs a set of differentiated regulatory suggestions that include overall intensity, special investigation and precise monitoring measures by querying the pre-set regulatory strategy knowledge base. Embodiments of the present invention also provide an electronic device, including a memory, a processor, and a communication bus; the memory and the processor are connected via the communication bus. The memory stores a special equipment inspection data intelligent analysis method based on natural language understanding, which can be loaded by the processor and executed as provided in the above embodiments.

[0020] The memory can be used to store instructions, programs, code, code sets, or instruction sets. The memory may include a program storage area and a data storage area. The program storage area may store instructions for implementing an operating system, instructions for at least one function, and instructions for implementing the intelligent analysis method for special equipment inspection data based on natural language understanding provided in the above embodiments, etc. The data storage area may store data involved in the intelligent analysis method for special equipment inspection data based on natural language understanding provided in the above embodiments, etc.

[0021] The processor may include one or more processing cores. The processor executes instructions, programs, code sets, or instruction sets stored in memory, and calls data stored in memory to perform various functions and process data as described in this application. The processor may be at least one of the following: Application-Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Digital Signal Processing Device (DSPD), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), Central Processing Unit (CPU), controller, microcontroller, and microprocessor. It is understood that, for different devices, the electronic devices used to implement the above-described processor functions may also be other types, and the embodiments of this application do not specifically limit this.

[0022] A communication bus can include a pathway for transmitting information between the aforementioned components. The communication bus can be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. Communication buses can be categorized as address buses, data buses, control buses, etc.

[0023] This application provides a computer-readable storage medium storing a computer program that can be loaded by a processor and executed as described in the above embodiments, which is a method for intelligent analysis of special equipment inspection data based on natural language understanding.

[0024] In this embodiment, a computer-readable storage medium can be a tangible device that holds and stores instructions used by an instruction execution device. A computer-readable storage medium can be, but is not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any combination thereof. Specifically, a computer-readable storage medium can be a portable computer disk, a hard disk, a USB flash drive, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), spoofing random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory stick, floppy disk, optical disk, magnetic disk, mechanical encoding device, or any combination thereof.

[0025] The terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0026] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the foregoing application concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions claimed in this application.

Claims

1. A method for intelligent analysis of special equipment inspection data based on natural language understanding, characterized in that, Includes the following steps: S1: Receive the unit risk assessment request input by the user in natural language, and extract the target unit name and risk analysis dimensions from the unit risk assessment request; S2: Based on the target user's name, associate and retrieve all special equipment files under the target user's name, and construct an equipment analysis group; at the same time, based on the risk analysis dimensions, match the corresponding special equipment safety risk assessment model from the pre-set model library; S3: For each device in the device analysis group, extract multi-source data related to the risk analysis dimensions simultaneously from the historical inspection database, operation monitoring platform and regulatory record library, combined with the changes in the scenario; S4: Based on the special equipment safety risk assessment model, quantitative calculations are performed on the multi-source data of each piece of equipment to generate a single equipment risk score and risk factor label; S5: Statistically analyze and rank the individual risk scores of all devices in the equipment analysis group, and perform aggregate analysis on the risk factor labels to identify the overall risk level, dominant risk type, and distribution characteristics of high-risk devices of the target user unit. S6: Based on the overall risk level, dominant risk type, and distribution characteristics of high-risk equipment of the target user unit, output differentiated regulatory recommendations for the target user unit.

2. The intelligent analysis method for special equipment inspection data based on natural language understanding according to claim 1, characterized in that, The process of receiving a risk assessment request from a user in natural language, and extracting the target user's name and risk analysis dimensions from the request, includes: S110: Receive natural language text input by the user as a unit risk assessment request; S120: Perform syntactic analysis and semantic role labeling on the natural language text to obtain semantic labeling results; S130: Based on the semantic annotation results, identify and extract the target user unit name and risk analysis dimensions.

3. The intelligent analysis method for special equipment inspection data based on natural language understanding according to claim 2, characterized in that, The process involves associating and retrieving all special equipment files under the target user's name to construct an equipment analysis group; simultaneously, based on risk analysis dimensions, matching corresponding special equipment safety risk assessment models from a pre-set model library, including: S210: Using the extracted target user's name as the query key, search the special equipment use registration database to obtain a list of equipment registration codes under the target user's name; S220: Based on the device registration code list, retrieve the complete technical files of each device in batches from the device lifecycle archive and construct a device analysis group; S230: Perform semantic similarity calculation between the extracted risk analysis dimensions and the keyword vectors corresponding to each risk assessment model in the pre-set model library, and select the model with the highest similarity as the matching special equipment safety risk assessment model.

4. The intelligent analysis method for special equipment inspection data based on natural language understanding according to claim 3, characterized in that, For each device in the device analysis group, multi-source data related to the risk analysis dimensions are simultaneously extracted from the historical inspection database, operation monitoring platform, and regulatory record database, taking into account changes in the scenario. This includes: S310: Based on the risk analysis dimensions, determine the corresponding equipment status change scenarios, and extract the time range and key features corresponding to the equipment status change scenarios; S320: For each device in the device analysis group, based on the device registration code and the time range and key characteristics, simultaneously initiate three data extraction requests: the first request is to the historical inspection database to extract historical inspection records of each device in the device analysis group that match the key characteristics within the time range; the second request is to the operation monitoring platform to extract sensor time-series data of each device in the device analysis group that reflects the key characteristics within the time range; and the third request is to the regulatory record database to extract regulatory document data associated with the time range and key characteristics. S330: Integrate the received historical inspection records, sensor time-series data, and regulatory document data to form multi-source data related to the risk analysis dimension.

5. The intelligent analysis method for special equipment inspection data based on natural language understanding according to claim 4, characterized in that, The special equipment safety risk assessment model quantifies the multi-source data for each piece of equipment to generate a single-equipment risk score and risk factor label, including: S410: Preprocess the multi-source data of each integrated device to convert the historical inspection records, sensor time-series data and regulatory document data into standard feature vectors; S420: Input the standard feature vector into the matching special equipment safety risk assessment model to calculate the risk score of a single device; S430: Based on the feature weights output by the special equipment safety risk assessment model, identify the key feature dimensions that contribute the most to the risk score of the single equipment, and generate corresponding risk factor labels.

6. The intelligent analysis method for special equipment inspection data based on natural language understanding according to claim 5, characterized in that, The process involves statistically analyzing and ranking the individual risk scores of all devices in the equipment analysis group, and performing aggregated analysis on risk factor labels to identify the overall risk level, dominant risk type, and distribution characteristics of high-risk equipment for the target user unit. This includes: S510: Calculate the average risk score of each device in the device analysis group and the proportion of high-risk devices; S520: Based on a pre-set risk level threshold mapping table, and combining the average risk score of each device in the device analysis group with the proportion of high-risk devices, determine the overall risk level of the target user unit; S530: Statistically analyze the frequency and combination patterns of all the risk factor labels, identify the label combination with the highest frequency as the dominant risk type; sort the single device risk scores in descending order, identify the corresponding devices in the device analysis group whose single device risk scores exceed the high-risk device identification threshold as high-risk devices, and analyze the distribution characteristics of the high-risk devices.

7. The intelligent analysis method for special equipment inspection data based on natural language understanding according to claim 6, characterized in that, Based on the overall risk level, dominant risk types, and distribution characteristics of high-risk equipment of the target user unit, differentiated regulatory recommendations are output for the target user unit, including: S610: Generate a corresponding overall regulatory intensity recommendation based on the overall risk level of the target user unit; S620: Generate corresponding specific risk assessment recommendations based on the primary risk type of the target user unit; S630: Based on the distribution characteristics of high-risk equipment of the target user unit, generate corresponding precise regulatory suggestions for high-risk equipment.

8. A special equipment inspection data intelligent analysis system based on natural language understanding, used to implement the special equipment inspection data intelligent analysis method based on natural language understanding as described in any one of claims 1 to 7, characterized in that, include: The request parsing module is used to receive unit risk assessment requests input by users in natural language, and extract the target unit name and risk analysis dimensions from the unit risk assessment requests; The retrieval and matching module is used to associate and retrieve all special equipment files under the target user unit name, and construct equipment analysis groups; at the same time, it matches the corresponding special equipment safety risk assessment model from the pre-set model library according to the risk analysis dimensions. The data extraction module is used to extract multi-source data related to risk analysis dimensions from historical inspection databases, operation monitoring platforms, and regulatory record databases for each device in the device analysis group, taking into account changes in the scenario. The risk calculation module is used to quantify the multi-source data of each piece of equipment based on the special equipment safety risk assessment model, and generate a single equipment risk score and risk factor label. The aggregation analysis module is used to statistically analyze and sort the individual risk scores of all devices in the device analysis group, and to perform aggregation analysis on risk factor labels to identify the overall risk level, dominant risk type, and distribution characteristics of high-risk devices of the target user unit. The suggestion generation module is used to output differentiated regulatory suggestions for the target user unit based on the target user unit's overall risk level, dominant risk type, and distribution characteristics of high-risk equipment.

9. An electronic device comprising a processor and a memory, wherein, The memory stores a computer program that can be called by a processor; characterized in that the processor executes the intelligent analysis method for special equipment inspection data based on natural language understanding as described in any one of claims 1 to 7 by calling the computer program stored in the memory.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the intelligent analysis method for special equipment inspection data based on natural language understanding as described in any one of claims 1 to 7.

Images