Industrial product trusted operation method based on crowd sensing

Abstract

In order to solve the problem that the single source of existing industrial product operation and maintenance information leads to the lack of flexibility and comprehensiveness of product operation and maintenance knowledge base, the application provides a kind of operation and maintenance knowledge graph construction method of industrial product based on crowd sensing, the reliability evaluation process is used to screen user experience, the reliable user experience obtained after screening includes specific problems encountered by users in actual production process and corresponding solution strategies, these reliable user experience is integrated into the basic knowledge graph of industrial product, the knowledge graph is further improved and the efficiency of crowd sensing in industrial product operation and maintenance is improved.Further, in order to solve the problem that the existing industrial product operation and maintenance method has the problem that each link in the operation and maintenance process is difficult to effectively cooperate and share information, the application provides a kind of reliable operation and maintenance method of industrial product based on crowd sensing, the operation and maintenance method involves the whole process of real-time monitoring, fault early warning and maintenance decision support, each link can effectively cooperate and share information, and closed-loop management is formed.

Description

Technical Field

[0001] The present invention relates to the field of industrial product operation and maintenance. The industrial product described in the present invention refers to production and processing equipment. Background Art

[0002] In recent years, with the rapid development of technologies such as microelectronics, the Industrial Internet of Things, and artificial intelligence, the manufacturing industry has rapidly developed toward automation, digitization, and intelligence. During the manufacturing process of industrial parts, the parameters of many key components change dynamically, significantly impacting the quality of the finished parts. Therefore, to ensure part quality, it is necessary to monitor these performance parameters during the manufacturing process, provide fault warnings, and provide accurate and efficient repair decisions when a fault occurs.

[0003] As a structured, semantic and interconnected information representation method, knowledge graph can efficiently organize and manage the experience and knowledge accumulated in the operation and maintenance of industrial products, and provide accurate query functions, which is of great significance for providing maintenance decision support in the operation and maintenance of industrial products.

[0004] Existing industrial product operation and maintenance methods, such as those in Chinese invention patent application number 202311870779.8, construct a knowledge graph of power equipment defects based on collected official data. This graph then uses the relationships and rules within it, combined with relevant algorithms, to predict potential defects in power equipment. Another example is Chinese invention patent application number 202311581518.4, which uses machine learning algorithms to predict the health status of equipment by inputting operating parameters into a constructed equipment knowledge graph. This predicts whether the equipment is abnormal and generates warning information.

[0005] While these studies have advanced industrial product operation and maintenance methods, they only consider industrial product information from manufacturers within the product operation and maintenance knowledge base, resulting in a lack of flexibility and comprehensiveness. Furthermore, the methods disclosed in these patent applications lack an integrated, full-process industrial product operation and maintenance process, hindering effective collaboration and information sharing across all aspects of the operation and maintenance process. Summary of the Invention

[0006] In order to solve the problem that the existing industrial product operation and maintenance information source is single, resulting in the lack of flexibility and comprehensiveness of the product operation and maintenance knowledge base, the present invention provides a method for constructing an operation and maintenance knowledge graph of industrial products based on crowd intelligence perception.

[0007] Furthermore, in order to solve the technical problem of the existing industrial product operation and maintenance methods that various links in the operation and maintenance process are difficult to effectively coordinate and share information, the present invention provides an industrial product trusted operation and maintenance method based on crowd intelligence perception.

[0008] The technical solution of the present invention is:

[0009] The method for constructing the operation and maintenance knowledge graph of industrial products based on crowd intelligence perception is special in that it includes the following steps:

[0010] Step 1: Based on official data of industrial products, build a basic knowledge graph of industrial products based on deep learning;

[0011] Step 2: Based on crowd intelligence perception, improve the basic knowledge graph of industrial products constructed in step 1 to obtain the operation and maintenance knowledge graph of industrial products;

[0012] Step 2.1: Obtain the industrial product triple P: P = {M, C, F}; M is the module that constitutes the industrial product, C is the component of the module, and F is the attribute of the component;

[0013] Step 2.2: Based on the mutual influence relationships between components belonging to different modules in the industrial product, construct an influence matrix E. The element values ​​in the influence matrix E are 0 or 1, determined according to the following rules:

[0014]

[0015] Step 2.3: Collect user experiences and represent each user experience with keywords as U_e = {Q, A}, where Q is the keyword set of the question and A is the keyword set of the answer. Based on the keyword set Q of the question and the industrial product triple P, the user experience can be assigned to the corresponding module of the industrial product.

[0016] Step 2.4: Filter out reliable user experiences from the collected user experiences;

[0017] Step 2.4.1: For each user experience, score the presence of the key words in its answer;

[0018] Step 2.4.2: For each user experience, rate the match between the keywords in the answer and the product module mentioned in the question;

[0019] Step 2.4.3: For each user experience, score the matching between the answer keywords and the product modules other than the question based on the influence matrix E;

[0020] Step 2.4.4: Based on the scores from steps 2.4.1-2.4.3, use the fuzzy evaluation method to calculate the column vector B consisting of the comprehensive evaluation results of all user experiences:

[0021] Step 2.4.5: Set a threshold to filter out reliable user experiences. If the value of an element in column vector B is greater than or equal to the threshold, the user experience corresponding to the element is considered reliable. Otherwise, the user experience corresponding to the element is unreliable.

[0022] Step 2.5: Add the screened reliable user experience to the basic knowledge graph of industrial products constructed in step 1 to obtain the operation and maintenance knowledge graph of industrial products.

[0023] Furthermore, step 1 is specifically as follows:

[0024] Step 1.1: Collect official data on industrial products and remove duplicates, missing values, and standardize the data format;

[0025] Step 1.2: Ontology modeling, including defining entities and relationships between them;

[0026] Step 1.3: Entity annotation;

[0027] For each category of official data processed in step 1.1, randomly select some data and annotate them with BIOE based on the entities defined in step 1.2, so that they correspond one-to-one with the entities defined in step 1.2;

[0028] Step 1.4: Knowledge extraction;

[0029] Step 1.4.1: Use the official data labeled with entities from step 1.3 to train the BERT+Bi-LSTM+CRF neural network model. Use the trained BERT+Bi-LSTM+CRF model to extract entities from the remaining unlabeled official data to obtain industrial product entities.

[0030] Step 1.4.2: For the types of industrial product entities extracted in step 1.4.1, form triples of product knowledge based on the relationships between the defined entities;

[0031] Step 1.4.3: Modify or delete the triplet information obtained in step 1.4.2, and retain the valid triplet information;

[0032] Step 1.5: Knowledge storage;

[0033] The Neo4j graph database is used to store the valid triple information obtained in step 1.4 to obtain the basic knowledge graph of industrial products.

[0034] Furthermore, step 2.4 is specifically as follows:

[0035] Step 2.4.1: For each user experience, score the presence of the key words in its answer;

[0036] For each user experience U_e, it is judged whether the answer keyword set A is empty. If it is empty, the score is 0, steps 2.4.2 and 2.4.3 are skipped, and the score of the two evaluation indexes of the user experience in steps 2.4.2 and 2.4.3 is 0; if it is not empty, the score is 1, and step 2.4.2 is entered;

[0037] Step 2.4.2: For each user experience, the matching of the answer keyword and the product module in the question is scored;

[0038] For each user experience U_e, the user experience is first assigned to the corresponding product module according to the question keyword Q, and then the answer keyword A in the user experience is matched with the components and component attributes in the product module assigned according to the question keyword. If the matching is successful, the score is 0.5, and step 2.4.3 is entered; if the matching is unsuccessful, the score is 0, and step 2.4.3 is entered;

[0039] Step 2.4.3: For each user experience, the matching of the answer keyword and other product modules except the question is scored based on the influence matrix E;

[0040] For each user experience U_e, if the answer keyword A in a user experience refers to the components of other modules except the product module assigned according to the question keyword in step 2.4.2, and it is judged according to the influence matrix E that the components of the other modules and the components referred to in the question keyword of the user experience influence each other, the score is 0.5, and step 2.4.4 is entered; otherwise, the score is 0, and step 2.4.4 is entered;

[0041] Step 2.4.4: The column vector B composed of the comprehensive evaluation results of all user experiences is calculated by the following formula:

[0042] B = R·W T

[0043] Wherein, W is the weight vector of the evaluation indexes, W = [ω1ω2ω3], ω1ω2ω3 are the weights of the existence score of the answer keyword, the matching score of the answer keyword and the product module in the question, and the matching score of the answer keyword and the product module except the question, respectively, and R is the fuzzy evaluation matrix, The elements in the matrix R are the scores of the three evaluation indexes obtained in steps 2.4.1-2.4.3, the rows in the matrix R represent different user experiences, and the columns represent the three evaluation indexes of the user experiences;

[0044] Step 2.4.5: A threshold is set to screen out reliable user experiences. If the value of an element in the column vector B is greater than or equal to the threshold, the user experience corresponding to the element is considered reliable, otherwise, the user experience corresponding to the element is considered unreliable.

[0045] Furthermore, in step 2.4.4, the values ​​of ω1ω2ω3 are 0.4, 0.4, and 0.2, respectively, and the threshold in step 2.4.5 is set to 0.5.

[0046] Furthermore, step 2.5 is specifically as follows:

[0047] Step 2.5.1: De-duplicate reliable user experiences;

[0048] Step 2.5.2: Based on the ontology model constructed in step 1.2, divide the reliable user experience after deduplication into entities and add relationships to form triple information of user experience;

[0049] Step 2.5.3: Use the Neo4j graph database to add and save the triple information of user experience obtained in step 2.5.2 to the basic knowledge graph of industrial products obtained in step 1 to obtain the improved operation and maintenance knowledge graph of industrial products.

[0050] The present invention also provides a method for trusted operation and maintenance of industrial products based on crowd intelligence perception, which is special in that it includes the following steps:

[0051] Step 1: Use physical sensors to build industrial products with sensing capabilities;

[0052] Step 2: Industrial product status monitoring and fault warning;

[0053] Step 2.1: Set corresponding fault warning threshold ranges for each key component of the industrial product;

[0054] Step 2.2: Industrial products with sensing capabilities monitor the operating status data of their key components in real time;

[0055] Step 2.3: Display the operating status data of each key component monitored in step 2.2 in real time and regularly compare it with its corresponding fault warning threshold range. If all are within the fault warning threshold range, it indicates that the industrial product is operating normally, and return to step 2.2. If the operating status data of a key component is not within its corresponding fault warning threshold range, it indicates that the industrial product is operating abnormally, and a fault description is generated and sent to the front end to notify the user, and then proceed to step 3.

[0056] Step 3: Generate industrial product maintenance decisions;

[0057] Step 3.1: Input the fault description obtained in step 2.2 into a pre-trained TextCNN model to obtain the fault type to which the fault description data belongs. The TextCNN model is trained using a fault description dataset constructed based on common fault types of industrial products. Each fault type in the fault description dataset is corresponding to a Neo4j query statement.

[0058] Step 3.2: Determine the corresponding Neo4j query language based on the fault type obtained in step 3.1;

[0059] Step 3.3: Input the Neo4j query language obtained in step 3.2 into the operation and maintenance knowledge graph constructed by the above-mentioned method for constructing the operation and maintenance knowledge graph of industrial products based on crowd intelligence perception to obtain the maintenance decision.

[0060] Furthermore, in step 2.1, based on the historical operating data and statistical laws of each key component, the 3σ principle is adopted to set the fault warning threshold range of each key component to (μ-3σ, μ+3σ), where μ and σ are the mean and standard deviation of the historical operating data of each key component, respectively.

[0061] Furthermore, in step 2.3, basic information of industrial products is also displayed in real time, including product model, product size, product processing principle, product processing accuracy, product processing size, product components and their component parameters.

[0062] The advantages of the present invention are:

[0063] 1. In the process of constructing the knowledge graph of industrial product operation and maintenance, in order to improve the comprehensiveness and flexibility of the knowledge graph, the present invention designs a reliability assessment process to evaluate and screen the reliability of user experience. The reliable user experience obtained after screening includes the specific problems encountered by users in the actual production process and the corresponding solution strategies. These reliable user experiences are of great value in product operation and maintenance. The present invention integrates these reliable user experiences into the basic knowledge graph of industrial products, thereby further improving the knowledge graph and improving the efficiency of crowd intelligence perception in industrial product operation and maintenance.

[0064] 2. The operation and maintenance method of the present invention achieves real-time monitoring and fault warning of industrial products by combining static and dynamic data about industrial products. By applying microelectronics technology, industrial products can sense the operating status of their internal key components and efficiently utilize them. The collected dynamic sensing data is combined with the static basic information of the industrial products to achieve real-time monitoring of industrial products. The collected operating parameters of key components are compared with fault warning thresholds set based on statistical laws. If an anomaly is found, a fault description is generated based on a template, thereby achieving precise location of industrial product faults and providing effective support for operation and maintenance decision-making.

[0065] 3. The present invention constructs an operation and maintenance knowledge graph for industrial products based on deep learning. The operation and maintenance knowledge graph contains text information such as expert technical manuals and reliable user experiences, and is stored in the Neo4j graph database, so that the operation and maintenance knowledge base has efficient management and data retrieval functions. The TextCNN model is used to classify industrial product fault descriptions, and a corresponding Neo4j query language is set for each type of fault description. When an industrial product fault description is obtained, the trained TextCNN model and the improved industrial product operation and maintenance knowledge graph can quickly help users find fault repair solutions, which helps to improve repair efficiency and repair results.

[0066] 4. The operation and maintenance method of the present invention involves a full-process industrial product operation and maintenance process of real-time monitoring, fault warning, and maintenance decision support. Each link can effectively collaborate and share information to form a closed-loop management, ultimately providing a reliable operation and maintenance method for industrial products.

[0067] 5. The crowd intelligence perception of the present invention is embodied in:

[0068] 1) Multi-source information collection: Collect user experiences from the exchanges and discussions of multiple users in industrial product forums. These user experiences are diverse and rich.

[0069] 2) Design processing and screening evaluation process: By screening, evaluating and filtering the information posted by users, reliable user experiences can be extracted from them, and these reliable user experiences contain useful knowledge information.

[0070] 3) Improve the knowledge graph of industrial product operation and maintenance: By enriching the knowledge base with reliable user experience obtained after screening, the basic knowledge graph of industrial products is enriched. The obtained operation and maintenance knowledge graph can provide specific scenarios and solutions during the actual use of industrial products. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 It is an overall flow chart of the operation and maintenance method of the present invention.

[0072] Figure 2 It is a flow chart for constructing the operation and maintenance knowledge graph of industrial products in the present invention.

[0073] Figure 3 It is a flow chart of reliability evaluation of user experience in the present invention. DETAILED DESCRIPTION

[0074] The present invention will be described in further detail below with reference to the accompanying drawings.

[0075] like Figures 1-3As shown, the industrial product trusted operation and maintenance method based on crowd intelligence perception provided by the present invention includes the following steps:

[0076] Step 1: Build a basic knowledge graph of industrial products based on deep learning;

[0077] Step 1.1: Data collection and data processing;

[0078] Collect official data about industrial products. This data includes basic information about the products (such as product specifications, components, and parameter information for each component), user manuals, test documents, and official maintenance plans. The collected official data types include structured data (product specifications and parameter data for each component) and unstructured data (manuals, usage logs, test documents, and maintenance reports). Data processing is required (deduplication, handling missing values, and standardization of data formats).

[0079] Step 1.2: ontology modeling;

[0080] The essence of ontology modeling is to describe the entities, attributes, relationships, and their hierarchical structure within a specific domain. In layman's terms, this involves refining industrial product information data into a conceptual representation of that information. Therefore, various class concepts must be defined, and specific knowledge content is represented as instantiations of these classes. For example, a 3D printer is an instance of a product. Entity and relationship types are crucial components of an ontology model, and this step of ontology modeling primarily involves defining these entities and relationships.

[0081] The specific process is as follows:

[0082] Step 1.2.1: Define entities;

[0083] Entities include product entities, attribute entities, and operation and maintenance entities. Product entities include products (i.e., conceptual representations of instances (e.g., 3D printers)), product modules, and product components, where product components generally represent specific objects that require operation and maintenance. Attribute entities include the working principles of industrial products, the specifications of industrial products and the dimensions of their components, the processing accuracy of industrial products, the normal operating parameters of product components, and the material properties of processed parts. Operation and maintenance entities include fault phenomena, fault causes, and fault solutions.

[0084] Step 1.2.2: Define relationships;

[0085] Relationships include relationships between entity elements within a class and relationships between entities of different classes. Predefined relationship types include "contains", "has", "occurs", "results in", and "resolves".

[0086] Step 1.3: Entity annotation;

[0087] For each type of official data obtained after processing in step 1.1, a portion of the data is randomly selected and annotated with BIOE based on the various entities defined in step 1.2.1. B, I, and E represent the beginning, middle, and end of the entity, respectively, and O represents the non-entity part of the sentence. BIOE annotation needs to be performed based on the entities defined in step 1.2.1, so that the official data is matched one-to-one with the various entities defined in step 1.2.1.

[0088] Step 1.4: Knowledge extraction;

[0089] Step 1.4.1: Use the official data with entity annotations from step 1.3 above to train the BERT+Bi-LSTM+CRF neural network model. Use the trained BERT+Bi-LSTM+CRF model to extract entities from the remaining unlabeled official data to obtain industrial product entities.

[0090] Step 1.4.2: Based on the types of industrial product entities extracted in step 1.4.1, form triples of product knowledge according to the predefined relationship types between entities. The triple format is <entity 1, category 1>-relationship-<entity 2, category 2>, such as <3D printer, product>-include-<printing platform module, product module>;

[0091] Step 1.4.3: The triple information obtained in step 1.4.2 is screened and whether there is erroneous triple information is determined based on the entity type and the relationship between entities. The erroneous triple information is corrected (including deduplication and missing information supplementation) or deleted, and finally the valid triple information is retained for subsequent operations.

[0092] Step 1.5: Knowledge storage;

[0093] After knowledge extraction, the valid triples obtained in step 1.4 are stored in a Neo4j graph database. The Neo4j graph database stores data in the form of nodes and relationships. The extracted entities and relationships are defined as node and relationship types, and the attributes of the relevant nodes are stored and defined. This completes the construction of the basic knowledge graph for industrial products.

[0094] Step 2: Based on crowd intelligence perception, improve the basic knowledge graph of industrial products constructed in step 1 to obtain the operation and maintenance knowledge graph of industrial products;

[0095] The basic knowledge graph for industrial products obtained in step 1 is supplemented with user experience. This experience can typically be obtained from industrial product community forums and consists of questions and answers. The reliability of the collected user experience is determined before it can be added to the knowledge graph. Reliable user experience is then added to the basic knowledge graph for industrial products.

[0096] The specific method is as follows:

[0097] Step 2.1: Obtain the industrial product triples represented by the structure and attributes of the industrial products;

[0098] Based on the functional units of industrial products, user experience needs to be allocated to different modules of industrial products. Therefore, the module set of industrial products is first defined as follows:

[0099] P M ={M1, M2...M X}

[0100] Among them, M1, M2...M X They are the 1st, 2nd, …, Xth modules in the industrial product respectively;

[0101] The components and component attributes of each module are defined as follows:

[0102] M i ={C1, C2...C Y}

[0103] C j ={F1, F2...F Z}

[0104] Among them, C1, C2...C Y Module M i The first, second, ..., Yth components in F1, F2...F Z Component C j The first, second, ..., Zth attributes of M; i takes 1, 2, ..., X, and j takes 1, 2, ..., Y respectively; M i 、C j 、F k It can be obtained from the basic knowledge graph of industrial products constructed in step 3, where k is 1, 2, …, Z respectively.

[0105] Combining M, C, and F, we can get the industrial product triple P that represents the structure and attributes of the industrial product:

[0106] P = {M, C, F}.

[0107] Step 2.2: Construct the impact matrix;

[0108] Considering that there can be influence relationship between components of different modules, a two-dimensional influence matrix E is constructed to represent the influence relationship, each element in the influence matrix E takes value 0 or 1, which is determined according to the following rules:

[0109]

[0110] For example, if an industrial product has two modules, module 1 and module 2; module 1 includes components A and B, and module 2 includes components C and D; if only component A in module 1 and component C in module 2 influence each other, component A in module 1 and component D in module 2 do not influence each other, and component B in module 1 and components C and D in module 2 do not influence each other, only the influence relationship between components in different modules is considered, and the influence matrix and the corresponding module components of its rows and columns are constructed as follows:

[0111]

[0112] Step 2.3: Collect user experiences and represent the user experiences with keywords;

[0113] The collected user experiences include questions and answers. The keywords are extracted from the collected user experiences. The keywords extracted from the questions include the module components and component attributes of the industrial product that appear problems, and the keywords extracted from the answers include the module components and component attributes in the questions and the components of different modules that influence each other. Finally, based on the extracted keywords, each piece of collected user experience can be represented as U_e:

[0114] U_e = {Q, A}

[0115] Wherein, Q represents the keyword set of the question, and A represents the keyword set of the answer. The keywords of the question and the triplets P obtained in step 2.1 can be used to assign the collected user experience to the corresponding product module.

[0116] Step 2.4: Reliability evaluation of user experience based on fuzzy evaluation, and screening of reliable user experience;

[0117] Step 2.4.1: For each user experience, score the existence of the answer keywords;

[0118] For each user experience U_e, it is judged whether the answer keyword set A is empty. If the answer keyword set A is empty, the score is 0, and steps 2.4.2 and 2.4.3 are skipped, and the score of the two evaluation indexes of the user experience in steps 2.4.2 and 2.4.3 is 0; if the answer keyword set A is not empty, the score of the evaluation index of the user experience in this step is 1, and step 2.4.2 is entered;

[0119] Step 2.4.2: For each user experience, score the match between the answer keywords and the product module in the question;

[0120] For each user experience U_e, first assign the user experience to the corresponding product module based on its question keyword Q. Then, match the answer keyword A in the user experience with the components and component attributes in the product module assigned based on the question keyword. If the match is successful, the evaluation index score of this step is 0.5, and the process proceeds to step 2.4.3. If the match is unsuccessful, the evaluation index score of this step is 0, and the process proceeds to step 2.4.3.

[0121] Step 2.4.3: For each user experience, score the matching between the answer keywords and other product modules besides the question based on the influence matrix E;

[0122] For each user experience U_e, if the answer keyword A in a user experience mentions a component of a module other than the product module assigned to the question keyword in step 2.4.2, and the component of the other module influences the component mentioned in the question keyword of the user experience (this can be determined based on the influence matrix E, where the value of the element at the corresponding position in the influence matrix E is 1), the evaluation index score of this step is 0.5, and the process proceeds to step 2.4.4. Otherwise, the evaluation index score of this step is 0, and the process proceeds to step 2.4.4.

[0123] Step 2.4.4: Based on the scores from steps 2.4.1-2.4.3, use the fuzzy evaluation method to calculate a column vector B consisting of the comprehensive evaluation results of all user experiences;

[0124] B=R·W T

[0125] Among them, W is the weight vector of the evaluation index, W = [ω1ω2ω3], ω1ω2ω3 are the weights of the existence score of the answer keyword, the matching score between the answer keyword and the product module in the question, and the matching score between the answer keyword and the product module outside the question, respectively. R is the fuzzy evaluation matrix. The elements in the matrix R are the scores of the three evaluation indicators obtained in steps 2.4.1-2.4.3. The rows in the matrix R represent different user experiences, and the columns represent the three evaluation indicators of user experience.

[0126] Step 2.4.5: Set a threshold and judge the reliability of each user experience based on the column vector B. If the value of an element in the column vector B is greater than or equal to the threshold, the user experience corresponding to the element is considered reliable. Otherwise, the user experience corresponding to the element is unreliable.

[0127] Step 2.5: Add the reliable user experience selected in step 2.4.5 to the basic knowledge graph of industrial products obtained in step 1;

[0128] Step 2.5.1: De-duplicate reliable user experiences;

[0129] Step 2.5.2: Based on the ontology model constructed in step 1.2, divide the reliable user experience after deduplication into entities and add relationships to form triple information of user experience;

[0130] Step 2.5.3: Use the Neo4j graph database to add and save the triplet information of user experience obtained in step 2.5.2 to the basic knowledge graph of industrial products obtained in step 1, and obtain a complete operation and maintenance knowledge graph of industrial products to improve the practicality, flexibility, and comprehensiveness of maintenance decision support.

[0131] Step 3: Use physical sensors to build industrial products with sensing capabilities;

[0132] Step 3.1: Analyze the structure and working principle of the industrial product to determine the key components and their operating parameters that need to be monitored during the operation of the industrial product. These components and their operating parameters are closely related to the performance and stable operation of the industrial product.

[0133] Step 3.2: Select corresponding physical sensors based on the key components and operating parameters that need to be monitored during the operation of industrial products, and deploy the selected physical sensors on the industrial products so that the industrial products can easily obtain the operating parameter data of the key components of the industrial products (such as the temperature of the processing platform, the speed of the motor, and the position of the processed parts), thereby having the ability to perceive their operating status.

[0134] Step 4: Real-time operation status monitoring and fault warning of industrial products;

[0135] Step 4.1: Set corresponding fault warning threshold ranges for each key component of the industrial product;

[0136] To implement industrial product failure warnings, it's necessary to set fault warning thresholds for each key component. These thresholds should be set based on historical operating data and statistical patterns. Furthermore, these thresholds should be adjusted based on changes in the product's performance and recent operating data to ensure sensitivity and accuracy.

[0137] Step 4.2: Industrial products with sensing capabilities monitor the operating status data of their key components in real time;

[0138] During the manufacturing process, physical sensors installed on industrial products collect real-time processing data from their key components to reflect the real-time operating status of the industrial products and display it in real time for monitoring purposes. First, a database is used on the server backend to store each sensor's information, including sensor number, sensor type, installation location, fault warning threshold range, current status value, and whether any parameters are abnormal. The raw data collected by the physical sensors is then de-noised, filtered, calibrated, and packaged. The packaged sensor data is then sent to the backend server using a suitable data transmission protocol (such as HTTP or MQTT). Upon receiving the sensor data, the backend server verifies, merges, and stores it. The backend server then updates the current status value of each sensor in real time based on the sensor number. Communication between the frontend and the server backend (such as WebSocket) is then established to obtain real-time sensor data updates. A frontend interface is designed to visualize the real-time operating data of the key components of the industrial products. This interface also displays basic information about the industrial products, including product model, dimensions, processing principle, processing accuracy, processing dimensions, and components and their parameters. By combining static and dynamic data on industrial products, real-time monitoring of industrial products is achieved.

[0139] Step 4.3: Implement real-time monitoring and early warning logic on the server backend;

[0140] The sensor information is used to regularly check whether the current state value of each sensor is within the set fault warning threshold range. If not, an abnormality is detected. The sensor information with abnormal data is read according to the sensor number, and a fault description is generated according to the template and notified to the user through the notification mechanism, and step 5 is entered at the same time. If it is, it indicates that the industrial product is operating normally, and regular checks are continued to check whether the data of each sensor is within the set fault warning threshold. The fault description template is as follows:

[0141] The [Parameter] of the current [Position] is abnormal, the current value is [Status Value], and the threshold range is [Threshold Range].

[0142] Step 5: Generate industrial product maintenance decision;

[0143] Step 5.1: Train the TextCNN model (the model can be trained in advance and used directly during fault monitoring);

[0144] Step 5.1.1: Based on the common fault types of industrial products, generate fault description data for each common fault type using the same fault description template as in Step 4.3. From this fault description data, construct a fault description dataset that includes different fault descriptions of industrial products and their corresponding fault types. At the same time, set a Neo4j query statement for each fault type in the fault dataset.

[0145] Step 5.1.2: Use the constructed fault description dataset and the existing conventional method to train the TextCNN model so that the TextCNN model can classify the fault description data in the fault description dataset into fault types.

[0146] Step 5.2: Obtain maintenance decision plan;

[0147] Step 5.2.1: Use the trained TextCNN model to classify the industrial product fault description data generated in step 4.3 to obtain the fault type to which the fault description data belongs;

[0148] Step 5.2.2: Determine the corresponding Neo4j query language based on the fault type obtained in step 5.2.1;

[0149] Step 5.2.3: Input the Neo4j query language obtained in step 5.2.2 into the industrial product operation and maintenance knowledge graph constructed in step 2.5. The industrial product operation and maintenance knowledge graph outputs maintenance decisions, including maintenance steps, replacement parts recommendations, and actions that maintenance personnel can take.

[0150] From the above process, it can be seen that the maintenance decision obtained by the present invention is obtained by combining the inference of the machine learning model and the information in the industrial product operation and maintenance knowledge graph, which can provide users with practical and feasible solutions and ultimately realize the trusted operation and maintenance process of industrial products.

[0151] Example:

[0152] This embodiment uses an FDM 3D printer as an example to illustrate the present invention.

[0153] See also Figure 1 , the process of this embodiment is as follows:

[0154] Step 1: Build a basic knowledge graph of FDM 3D printers based on deep learning;

[0155] Step 1.1: Data collection and data processing;

[0156] Collect official data about FDM 3D printers, including user manuals, expert technical manuals, usage logs, and official repair plans. Most of this data is text and unstructured, requiring deduplication and missing value processing before use.

[0157] Step 1.2: ontology modeling;

[0158] For the operation and maintenance of FDM 3D printers, ontology modeling includes the definition of entities and relationships.

[0159] Step 1.2.1: Define entities;

[0160] Entities include product entities, attribute entities, and operation and maintenance entities. Product entities include products, product modules, and product components, represented by Product, Product_Module, and Product_Component, respectively. Attribute entities include the attributes of products and product components, and attributes are uniformly represented by Feature. Operation and maintenance entities include fault symptoms, fault causes, and fault solutions, represented by Fault, Cause, and Solution, respectively.

[0161] Step 1.2.2: Define relationships;

[0162] Relationships include those between entity elements of the same type and those between entities of different types. Predefined relationship types include Include, Has, Occur, Lead to, and Solution, represented by Include, Has_feature, Happen, Lead to, and Overcome, respectively. The relationships between predefined entities are as follows: Product contains Product_Module, Product_Module contains Product_Component, Product and Product_Component have Feature, Product_Component has Fault, Cause causes Fault, and Solution resolves Cause.

[0163] Step 1.3: Entity annotation;

[0164] Based on the entities defined in step 1.2.1, the BIOE annotation method is used to annotate the official data of the 3D printer obtained after processing in step 1.1. The entities in the official data are annotated according to the entities defined in step 1.2.1, with B, I, and E representing the start, middle, and end of the entity, respectively. Non-entities in the official data are annotated with O. BIOE annotation is performed based on the entities defined in step 1.2.1, and the official data is mapped one-to-one to the entities defined in step 1.2.1. For example, if the printing platform is a product module, the printing platform is annotated using BIOE as follows: "printing" is annotated with "B-Product_Module", "printing" is annotated with "I-Product_Module", "flat" is annotated with "I-Product_Module", and "platform" is annotated with "E-Product_Module".

[0165] Step 1.4: Knowledge extraction;

[0166] The BERT+Bi-LSTM+CRF neural network model is used for entity extraction (the model can be built based on the PyTorch framework using the existing method).

[0167] Using the official 3D printer data annotated in step 1.3, we trained a BERT+Bi-LSTM+CRF neural network model using conventional methods. The first layer of this model uses a pre-trained BERT to receive the official data and convert each word in the data into a corresponding embedding vector. The second layer uses a Bi-LSTM layer to process the input sequence (i.e., the BERT output) in both forward and reverse directions, capturing contextual features at each position in the input sequence. The third layer, a CRF layer, constrains and decodes the output of the Bi-LSTM layer, outputting the entity type corresponding to each word in the official data.

[0168] For the remaining unlabeled official 3D printer data, the trained BERT+Bi-LSTM+CRF neural network model is used to extract entities and obtain the entities of FDM 3D printers.

[0169] For the extracted entities, add relationships to the extracted entities according to the relationships predefined in step 1.2.2 to form triple information of 3D printer knowledge in the format of <entity 1, type>-relationship-<entity 2, type>.

[0170] Finally, the triple information of 3D printer knowledge needs to be deduplicated, supplemented, and deleted.

[0171] Step 1.5: Knowledge storage;

[0172] After knowledge extraction, the Neo4j graph database is used to store the triple information of the 3D printer knowledge obtained in step 1.4.

[0173] The Neo4j graph database stores data in the form of nodes and relationships. Each node in a Neo4j graph database represents an entity, and the node label is the entity type. Nodes can be assigned attributes corresponding to the entity (e.g., Attribute 1: Attribute 1 Content...). Nodes are connected by edges, which represent relationships between entities and can also be assigned attributes. By defining the extracted entities and relationships as node and relationship types and storing and defining the attributes associated with the entities, we can store triples of 3D printer knowledge in the Neo4j graph database. For example, we first define the label "Product_Component"; then add a node named "Sprinkler" and assign node attributes to it (including detailed information about the sprinkler); finally, we define relationships to establish relationships between the entity nodes. In this way, by storing triples of 3D printer knowledge in the Neo4j graph database (e.g., <Printer Head Module, Product_Module, Include, Sprinkler, Product_Component>), we construct a knowledge base and form a basic knowledge graph for 3D printers.

[0174] Step 2: Improve the basic knowledge graph of 3D printers built in step 1 based on crowd intelligence perception;

[0175] Step 2.1: Obtain the industrial product triples represented by the structure and attributes of the 3D printer;

[0176] The FDM 3D printer module set is defined as follows:

[0177] P M ={M1: print head module, M2: print platform module, M3: motion system module,

[0178] M4: Extrusion system module}

[0179] The components and component attributes of each product module are defined as follows:

[0180] The components and component properties of the print head module are as follows:

[0181] M1={C1: nozzle, C2: heating block, C3: thermistor}

[0182] C1={F1: maximum nozzle temperature, F2: nozzle diameter, F3: nozzle flow rate}

[0183] C2={F1: heating power, F2: heating area, F3: heating block material}

[0184] C3={F1: working range, F2: measurement accuracy, F3: installation dimensions}

[0185] The components and component properties of the print platform module are as follows:

[0186] M2 = {C1: Print bed, C2: Print model, C3: Heating device}

[0187] C1={F1: surface material, F2: size and shape, F3: surface flatness}

[0188] C2={F1: model layer thickness, F2: model material, F3: support structure}

[0189] C3={F1: heating power, F2: overall size, F3: working range}

[0190] The components and component properties of the motion system module are as follows:

[0191] M3={C1: X-axis, Y-axis, Z-axis motors, C2: X-axis, Y-axis belts, C3: Z-axis screw}

[0192] C1={F1: torque, F2: step angle, F3: maximum speed}

[0193] C2={F1: belt type, F2: belt width, F3: belt material}

[0194] C3={F1: screw type, F2: screw diameter, F3: screw pitch}

[0195] A collection of components and component properties for the Extrusion System module:

[0196] M4={C1: extrusion motor, C2: feed gear, C3: cooling fan}

[0197] C1={F1: torque, F2: step angle, F3: maximum speed}

[0198] C2={F1: gear material, F2: tooth shape, F3: gear diameter}

[0199] C3={F1: fan speed, F2: fan air volume, F3: working area}

[0200] The above modules, components, and component attributes can be obtained from the basic knowledge graph of FDM 3D printers constructed in step 1. Combining modules, components, and attributes can obtain the triple P representing the structure and attributes of the FDM 3D printer:

[0201] P = {M, C, F};

[0202] Step 2.2: Construct the impact matrix;

[0203] Considering that there may be influence relationships between components of different modules in the FDM 3D printer, an influence matrix is ​​constructed based on the influence relationships between components of different modules in step 2.1.

[0204] The nozzle in the print head module affects each other with the print bed and print model in the print platform module, the Z-axis motor and Z-axis lead screw in the motion system module, and the extrusion motor, feed gear, and cooling fan in the extrusion system module; the heating block in the print head module affects each other with the print model in the print platform module and the extrusion motor in the extrusion system module; the print bed in the print platform module also affects each other with the X-axis Y-axis motor and X-axis Y-axis belt in the motion system module; the print model in the print platform module also affects each other with the X-axis Y-axis Z-axis motor, X-axis Y-axis belt, Z-axis lead screw in the motion system module, and the extrusion motor, feed gear, and cooling fan in the extrusion system module. A two-dimensional influence matrix E is constructed to represent the influence relationship. Each element in the influence matrix E represents the influence relationship between different modules, and the element value is 0 or 1. Only the influence between different module components is considered, and the influence relationship between different components in the same module is not considered. If the different module components represented by the rows and columns affect each other, the corresponding influence matrix element value is 1, otherwise it is 0. The influence matrix E constructed in this embodiment is:

[0205]

[0206] Among them, the rows and columns of the influence matrix E represent the nozzle (PT), heating block (JK), thermistor (RZ) in the print head module, the print bed (DC), print model (DM), heating device (JZ) in the print platform module, the X-axis Y-axis Z-axis motor (XJ), X-axis Y-axis belt (XD), Z-axis screw (ZG) in the motion system module, and the extrusion motor (JJ), feed gear (JL), and cooling fan (LS) in the extrusion system module. The elements in the matrix are determined according to whether the components represented by the rows and columns affect each other. If there is a mutual influence relationship, the corresponding element is 1, otherwise it is 0. At the same time, only the influence between components of different modules is considered. When establishing the influence matrix E, it is recommended that the column names and row names correspond to the components of each module in the 3D printer, and are arranged in sequence to facilitate subsequent search indexes;

[0207] Step 2.3: Collect user experiences and express them with keywords;

[0208] The collected user experiences include questions and answers. Keywords are extracted from the collected user experiences. The keywords extracted from the questions include the module components and component attributes of the 3D printer problem. The keywords extracted from the answers include the module components and component attributes in the questions and the components of different modules that affect each other. Finally, based on the extracted keywords, each collected user experience can be represented as U_e:

[0209] U_e={Q、A}

[0210] Among them, Q represents the keyword set of the question, A represents the keyword set of the answer, and the user experience can be assigned to the corresponding module in the 3D printer through the keyword of the question and the triplet P obtained in step 2.1.

[0211] Step 2.4: Evaluate the reliability of user experience based on fuzzy evaluation and select reliable user experience;

[0212] The existence of answer keywords (first indicator), the matching of answer keywords with the 3D printer modules mentioned in the corresponding questions (second indicator), and the matching of answer keywords with 3D printer product modules other than the corresponding questions (third indicator) are used as evaluation indicators to score each user experience, and then the comprehensive evaluation result is calculated based on the fuzzy evaluation matrix and the weight of each indicator. Considering that user experience can be used as an alternative for operation and maintenance, the selected user experience must meet the first indicator and at least one of the second and third indicators. The evaluation weights and thresholds are set according to the above rules. When calculating the comprehensive evaluation results, the weights of the first, second, and third evaluation indicators can be defined as 0.4, 0.4, and 0.2, respectively, and the threshold is set to 0.5.

[0213] Compare the calculated comprehensive evaluation results of each user experience with the threshold. If the comprehensive evaluation result of a user experience is greater than or equal to the threshold, the user experience is considered reliable.

[0214] Step 2.5: Add the reliable user experience screened out in step 2.4 to the 3D printer basic knowledge graph established in step 1 to obtain the 3D printer operation and maintenance knowledge graph.

[0215] Step 3: Use physical sensors to build an FDM-based 3D printer with sensing capabilities;

[0216] Step 3.1: The FDM 3D printer realizes the process of layer-by-layer accumulation by controlling the movement of the extrusion head and the printing platform, and controlling the heating and cooling of the molten material. Therefore, the key components of the FDM 3D printer are determined as the nozzle, the printing platform and the motor, and the operating parameters of these key components are: the temperature, moving speed, pressure and position of the nozzle, the temperature, moving speed and position of the printing platform, the rotating speed, vibration parameter and voltage of the motor.

[0217] Step 3.2: The corresponding physical sensors are respectively configured on the key components determined in step 3.1, so that the FDM 3D printer can monitor its own operating state during the printing process and has the sensing ability of its operating state.

[0218] Step 4: Real-time operating state monitoring and fault warning of the FDM 3D printer;

[0219] Step 4.1: The corresponding fault warning threshold range is set for each key component of the FDM 3D printer;

[0220] First, the historical operating data of the key components (nozzle, printing platform and motor) of the 3D printer are obtained, and based on the statistical law, the 3σ principle is used to set the corresponding fault warning threshold range of each key component. The 3σ principle is: the probability of operating data distributed in (μ-σ, μ+σ) is 0.6827; the probability of operating data distributed in (μ-2σ, μ+2σ) is 0.9545; the probability of operating data distributed in (μ-3σ, μ+3σ) is 0.9973. Based on the 3σ principle, when the mean μ and standard deviation σ of the historical operating data of each key component of the 3D printer are calculated, the fault warning threshold range (μ-3σ, μ+3σ) is set for each key component.

[0221] Further, the historical data can also be reviewed regularly, and the fault warning threshold range of each key component can be flexibly adjusted according to the performance change and actual situation of the 3D printer.

[0222] Step 4.2: The FDM 3D printer with sensing ability monitors the operating state data of each key component in real time;

[0223] The data collected by the physical sensors carried on the 3D printer during the process of machining parts include: the temperature, moving speed, pressure and position of the nozzle, the temperature, moving speed and position of the printing platform, and the rotating speed, vibration parameter and voltage of the motor.

[0224] Each installed sensor is uniquely numbered and stored with its type, location, fault warning threshold, current status, and a Bool value indicating whether it is abnormal. A MySQL database is used to store the information for each deployed sensor in the format {ID, Type, Location, Threshold, Value, Bool}.

[0225] After the physical sensor collects the data, first, the data collected by the physical sensor is denoised, filtered, and calibrated, and then packaged in JSON format. A back-end server program is written, and the packaged data is sent to the back-end server using the MQTT protocol. After the back-end server receives the sensor data, it parses the data to extract the sensor monitoring parameters and verify the integrity of the sensor data. The verified data is stored in the above-mentioned database connected to the back-end server. The back-end server updates the current status value of each sensor in real time according to the sensor number. Then, WebSocket communication is established between the front-end and the back-end of the server, and the front-end application is built using the Web framework. A visual interface is designed. The server uses WebSocket to push new sensor data to the front-end. The front-end receives real-time data updates through WebSocket and uses the chart library in the visual interface to display the real-time operating parameter data of the key components of the 3D printer in real time. At the same time, basic information of the 3D printer is displayed, including the 3D printer model, processing principle, processing accuracy, processing materials, motion system parameters, printing platform parameters, and the properties of each component. Step 4.3: Implement real-time monitoring and early warning logic on the back-end of the server;

[0226] The back-end server updates the current status value of the corresponding sensor in real time according to the sensor number ID, and compares it with the fault warning threshold range Threshold set in step 4.1 in real time. If the current status value of a sensor is not within the fault warning threshold range of the key component it monitors, a parameter abnormality is detected. At this time, the Bool value in the sensor information corresponding to the sensor is updated to True. Otherwise, the Bool value in the sensor information corresponding to the sensor always remains False.

[0227] When the Bool value in a certain sensor information is updated to True, the sensor information with data abnormality is read according to the sensor number ID, and fault description data is generated according to the fault description template. WebSocket real-time communication is used, and the front-end uses the browser, pop-up window or other methods to send an abnormal alarm notification, that is, the fault description data, to the user.

[0228] The above fault description template is:

[0229] The [Type] of the current [Location] is abnormal, the current value is [Value], and the threshold range is [Threshold].

[0230] Step 5: Generate FDM 3D printer maintenance decision;

[0231] Step 5.1: Train the TextCNN model.

[0232] Step 5.1.1: Based on the common fault types of FDM 3D printers, generate corresponding fault description data using the same fault description template as in Step 4.3, thereby constructing a fault description dataset. This fault description dataset includes different FDM 3D printer fault descriptions and their corresponding fault types, such as "The current printer nozzle temperature is abnormal. The current value is 270°C, and the threshold range is 250-260°C." The corresponding fault type is annotated, such as "Abnormal nozzle temperature." Furthermore, a Neo4j query statement is created for each fault type in the fault description dataset.

[0233] Step 5.1.2: Use the constructed fault description dataset and use existing conventional methods to train the TextCNN model.

[0234] Step 5.2: Obtain maintenance decision;

[0235] Step 5.2.1: After inputting the fault description data obtained in step 4.3 into the TextCNN model trained in step 5.1, the TextCNN model outputs the fault type to which the fault description data belongs;

[0236] Step 5.2.2: Determine the corresponding Neo4j query language based on the fault type obtained in step 5.2.1;

[0237] Step 5.2.3: Input the Neo4j query language obtained in step 5.2.2 into the 3D printer's maintenance knowledge graph constructed in step 2.5. The 3D printer's maintenance knowledge graph outputs a maintenance decision, including maintenance steps, recommended replacement parts, and actions that maintenance personnel can take. For example, if the fault description data obtained in step 4.3 is: "The current temperature of the printer's nozzle is abnormal. The current value is 270°C, and the threshold is 250-260°C," the TextCNN model outputs that this fault description belongs to the "Abnormal nozzle temperature" fault type. Using the Neo4j query language corresponding to this fault type, combined with the 3D printer's maintenance knowledge graph to query relevant information, the maintenance decision is generated: "Check the nozzle temperature sensor. The temperature sensor may need to be replaced or cleaned."

[0238] From the above process, it can be seen that the maintenance decision obtained in this embodiment is obtained by combining the inference of the machine learning model and the information in the operation and maintenance knowledge graph of the 3D printer. It can provide users with practical and feasible fault solutions and ultimately realize the trusted operation and maintenance process of the FDM 3D printer.

Claims

1. A method for constructing an operation and maintenance knowledge graph for industrial products based on crowd intelligence perception, characterized in that: The following steps are involved: Step 1: Based on official data of industrial products, build a basic knowledge graph of industrial products based on deep learning; Step 2: Based on crowd intelligence perception, improve the basic knowledge graph of industrial products constructed in step 1 to obtain the operation and maintenance knowledge graph of industrial products; Step 2.1: Obtain the industrial product triple P: P = {M, C, F}; M is the module that constitutes the industrial product, C is the component of the module, and F is the attribute of the component; Step 2.2: Based on the mutual influence relationships between components belonging to different modules in the industrial product, construct an influence matrix E. The element values ​​in the influence matrix E are 0 or 1, determined according to the following rules: Step 2.3: Collect user experiences and represent each user experience with keywords as U_e = {Q, A}, where Q is the keyword set of the question and A is the keyword set of the answer. Based on the keyword set Q of the question and the industrial product triple P, the user experience can be assigned to the corresponding module of the industrial product. Step 2.4: Filter out reliable user experiences from the collected user experiences; Step 2.4.1: For each user experience, score the presence of the key words in its answer; Step 2.4.2: For each user experience, rate the match between the keywords in the answer and the product module mentioned in the question; Step 2.4.3: For each user experience, score the matching between the answer keywords and the product modules other than the question based on the influence matrix E; Step 2.4.4: Based on the scores from steps 2.4.1-2.4.3, use the fuzzy evaluation method to calculate the column vector B consisting of the comprehensive evaluation results of all user experiences: Step 2.4.5: Set a threshold to filter out reliable user experiences. If the value of an element in column vector B is greater than or equal to the threshold, the user experience corresponding to the element is considered reliable. Otherwise, the user experience corresponding to the element is unreliable. Step 2.5: Add the screened reliable user experience to the basic knowledge graph of industrial products constructed in step 1 to obtain the operation and maintenance knowledge graph of industrial products.

2. The method for constructing an operation and maintenance knowledge graph of industrial products based on crowd intelligence perception according to claim 1 is characterized in that: Step 1 is as follows: Step 1.1: Collect official data on industrial products and remove duplicates, missing values, and standardize the data format; Step 1.2: Ontology modeling, including defining entities and relationships between them; Step 1.3: Entity annotation; For each category of official data processed in step 1.1, randomly select some data and annotate them with BIOE based on the entities defined in step 1.2, so that they correspond one-to-one with the entities defined in step 1.2; Step 1.4: Knowledge extraction; Step 1.4.1: Use the official data labeled with entities from step 1.3 to train the BERT+Bi-LSTM+CRF neural network model. Use the trained BERT+Bi-LSTM+CRF model to extract entities from the remaining unlabeled official data to obtain industrial product entities. Step 1.4.2: For the types of industrial product entities extracted in step 1.4.1, form triples of product knowledge based on the relationships between the defined entities; Step 1.4.3: Modify or delete the triplet information obtained in step 1.4.2, and retain the valid triplet information; Step 1.5: Knowledge storage; The Neo4j graph database is used to store the valid triple information obtained in step 1.4 to obtain the basic knowledge graph of industrial products.

3. The method for constructing an operation and maintenance knowledge graph of industrial products based on crowd intelligence perception according to claim 2 is characterized in that: Step 2.4 is as follows: Step 2.4.1: For each user experience, score the presence of the key words in its answer; For each user experience U_e, determine whether the answer keyword set A is empty. If it is empty, the score is 0, and steps 2.4.2 and 2.4.3 are skipped. At the same time, the score of the two evaluation indicators of steps 2.4.2 and 2.4.3 for this user experience is 0; if it is not empty, the score is 1, and step 2.4.2 is entered. Step 2.4.2: For each user experience, score the match between the answer keywords and the product module in the question; For each user experience U_e, first assign the user experience to the corresponding product module based on its question keyword Q. Then, match the answer keyword A in the user experience with the components and component attributes in the product module assigned based on the question keyword. If the match is successful, the score is 0.5 and the process proceeds to step 2.4.

3. If the match is unsuccessful, the score is 0 and the process proceeds to step 2.4.

3. Step 2.4.3: For each user experience, score the matching between the answer keywords and other product modules besides the question based on the influence matrix E; For each user experience U_e, if the answer keyword A in a user experience mentions a component of a module other than the product module assigned to the question keyword in step 2.4.2, and the influence matrix E shows that the component of the other module and the component mentioned in the question keyword of the user experience mutually influence each other, then the score is 0.5 and the process proceeds to step 2.4.

4. Otherwise, the score is 0 and the process proceeds to step 2.4.

4. Step 2.4.4: Use the following formula to calculate the column vector B consisting of the comprehensive evaluation results of all user experiences: B=R·W T Among them, W is the weight vector of the evaluation index, W = [ω1 ω2 ω3], ω1 ω2 ω3 are the weights of the existence score of the answer keyword, the matching score between the answer keyword and the product module in the question, and the matching score between the answer keyword and the product module outside the question, respectively. R is the fuzzy evaluation matrix. The elements in matrix R are the scores of the three evaluation indicators obtained in steps 2.4.1-2.4.

3. The rows in matrix R represent different user experiences, and the columns represent the three evaluation indicators of user experience; Step 2.4.5: Set a threshold to filter out reliable user experiences. If the value of an element in column vector B is greater than or equal to the threshold, the user experience corresponding to the element is considered reliable. Otherwise, the user experience corresponding to the element is unreliable.

4. The method for constructing an operation and maintenance knowledge graph of industrial products based on crowd intelligence perception according to claim 3 is characterized in that: In step 2.4.4, the values ​​of ω1ω2ω3 are 0.4, 0.4, and 0.2, respectively, and the threshold in step 2.4.5 is set to 0.

5.

5. The method for constructing an operation and maintenance knowledge graph of industrial products based on crowd intelligence perception according to claim 3 or 4, characterized in that: Step 2.5 is as follows: Step 2.5.1: De-duplicate reliable user experiences; Step 2.5.2: Based on the ontology model constructed in step 1.2, divide the reliable user experience after deduplication into entities and add relationships to form triple information of user experience; Step 2.5.3: Use the Neo4j graph database to add and save the triple information of user experience obtained in step 2.5.2 to the basic knowledge graph of industrial products obtained in step 1 to obtain the improved operation and maintenance knowledge graph of industrial products.

6. The trusted operation and maintenance method for industrial products based on crowd intelligence perception is characterized by: The following steps are involved: Step 1: Use physical sensors to build industrial products with sensing capabilities; Step 2: Industrial product status monitoring and fault warning; Step 2.1: Set corresponding fault warning threshold ranges for each key component of the industrial product; Step 2.2: Industrial products with sensing capabilities monitor the operating status data of their key components in real time; Step 2.3: Display the operating status data of each key component monitored in step 2.2 in real time and regularly compare it with its corresponding fault warning threshold range. If all are within the fault warning threshold range, it indicates that the industrial product is operating normally, and return to step 2.

2. If the operating status data of a key component is not within its corresponding fault warning threshold range, it indicates that the industrial product is operating abnormally, and a fault description is generated and sent to the front end to notify the user, and then proceed to step 3. Step 3: Generate industrial product maintenance decisions; Step 3.1: Input the fault description obtained in step 2.2 into a pre-trained TextCNN model to obtain the fault type to which the fault description data belongs. The TextCNN model is trained using a fault description dataset constructed based on common fault types of industrial products. Each fault type in the fault description dataset is corresponding to a Neo4j query statement. Step 3.2: Determine the corresponding Neo4j query language based on the fault type obtained in step 3.1; Step 3.3: Input the Neo4j query language obtained in step 3.2 into the operation and maintenance knowledge graph constructed by the operation and maintenance knowledge graph construction method of industrial products based on crowd intelligence perception described in any one of claims 1-5 to obtain a maintenance decision.

7. The method for trusted operation and maintenance of industrial products based on crowd intelligence perception according to claim 6 is characterized by: In step 2.1, based on the historical operating data and statistical laws of each key component, the 3σ principle is used to set the fault warning threshold range of each key component to (μ-3σ, μ+3σ), where μ and σ are the mean and standard deviation of the historical operating data of each key component, respectively.

8. The method for trusted operation and maintenance of industrial products based on crowd intelligence perception according to claim 7 is characterized in that: In step 2.3, basic information of industrial products is also displayed in real time, including product model, product size, product processing principle, product processing accuracy, product processing size, product components and their component parameters.

Images