Maintenance spare part prediction system based on big language model plug-in knowledge base

Abstract

The invention provides a maintenance spare part prediction system based on a big language model plug-in knowledge base, and the system comprises the steps: carrying out the information data extraction and preprocessing of a historical work order, and obtaining a preprocessing text; performing semantic vector conversion and association binding processing on the preprocessed text to obtain a vector database; extracting attribute information from the new work order, and performing semantic vector conversion on the attribute information to obtain a query vector; retrieving a work order related to the query vector from a vector database; semantic reasoning is carried out on a new work order and a related work order, a maintenance spare part prediction result is generated, natural language fault description and equipment model information in a maintenance record are fully utilized, natural language information is fully understood in the face of complexity and diversity, the spare part prediction accuracy and practicability are improved, reasonable configuration of maintenance resources is achieved, and the maintenance efficiency is improved. Spare part inventory cost and inventory waste are reduced, and maintenance efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and in particular to a maintenance spare parts prediction system based on a large language model plug-in knowledge base. Background Art

[0002] In after-sales service support systems, accurate forecasting of spare parts is crucial for improving operational efficiency and reducing inventory costs. Traditional spare parts forecasting methods rely primarily on statistical analysis of historical consumption data or expert experience. However, these methods suffer from low forecast accuracy when faced with incomplete data, diverse equipment models, or complex failure modes, making them unable to meet the high-precision requirements of after-sales spare parts forecasting.

[0003] Currently, regression analysis, decision trees, or neural networks are used to model historical maintenance records to predict spare parts demand. However, these methods still have significant limitations in processing fault information described in natural language. They are unable to fully utilize the large amount of unstructured text information in maintenance work orders, resulting in weak generalization capabilities of the prediction models and prediction results that are susceptible to sample bias. In addition, most existing prediction methods are difficult to adapt to new equipment models or rare fault scenarios and lack dynamic learning and knowledge transfer capabilities, resulting in insufficient adaptability and scalability in actual operation and maintenance scenarios, restricting the optimal configuration of maintenance resources and affecting the efficiency and reliability of equipment operation and maintenance. It can be seen that combining structured and unstructured information to develop a spare parts prediction method with contextual understanding and semantic association is of great significance to improving the accuracy and practicality of spare parts prediction. Summary of the Invention

[0004] The purpose of the present invention is to provide a maintenance spare parts prediction system based on a large language model plug-in knowledge base, extracting and preprocessing information data from historical work orders to obtain preprocessed text; performing semantic vector conversion and association binding processing on the preprocessed text to obtain a vector database; extracting attribute information from new work orders, performing semantic vector conversion on the attribute information to obtain a query vector; retrieving work orders related to the query vector from the vector database; performing semantic reasoning on new work orders and related work orders to generate maintenance spare parts prediction results, making full use of natural language fault descriptions and equipment model information in maintenance records, fully understanding natural language information in the face of complexity and diversity, improving spare parts prediction accuracy and practicality, achieving reasonable allocation of maintenance resources, reducing spare parts inventory costs and inventory waste, and improving maintenance efficiency.

[0005] The present invention is achieved through the following technical solutions: A maintenance spare parts prediction system based on a large language model plug-in knowledge base, comprising: The historical work order preprocessing module is used to extract and preprocess information data of historical work orders to obtain preprocessed text; A vector database construction module is used to perform semantic vector conversion and association binding processing on the preprocessed text to obtain a vector database; A new work order conversion processing module is used to extract attribute information from the new work order, perform semantic vector conversion on the attribute information, and obtain a query vector; a hybrid retrieval module, configured to retrieve work orders related to the query vector from the vector database; The large language model prediction module is used to perform semantic reasoning on the new work order and the related work orders to generate a maintenance spare parts prediction result.

[0006] Optionally, the historical work order preprocessing module is used to extract and preprocess information data of historical work orders to obtain preprocessed text, including: Identify the information type of the data contained in historical work orders and extract the corresponding type of information data; The information data is subjected to abnormality elimination, text standardization, and clustering processing in sequence to obtain a preprocessed text.

[0007] Optionally, the historical work order preprocessing module sequentially performs exception elimination, text standardization, and clustering on the information data to obtain a preprocessed text, including: Eliminating abnormal data from the information data; wherein the abnormal data includes at least one of a field missing portion, a semantically invalid portion, a fault description blank portion, and a logical conflict portion; Performing text field format unification processing and invalid stop word removal processing on the information data; A large language model is used to perform fault-related field clustering analysis on the information data, and semantically similar field descriptions are aggregated and normalized to obtain preprocessed text.

[0008] Optionally, the vector library construction module is used to perform semantic vector conversion and association binding processing on the preprocessed text to obtain a vector database, including: Extracting key content fields of the preprocessed text, and performing embedding vector conversion on the key content fields to obtain semantic vectors; The semantic vector is associated and bound with the historical work order corresponding to the pre-processed text, and then stored in a vector database.

[0009] Optionally, the vector library construction module extracts key content fields of the preprocessed text, including: The fault description field and the device model field of the preprocessed text are extracted as key content fields.

[0010] Optionally, the new work order conversion processing module is configured to extract attribute information from the new work order, perform semantic vector conversion on the attribute information, and obtain a query vector, including: Extract the fault description field and device model field from the new work order as attribute information; The attribute information is converted into an embedding vector using a large language model to obtain a query vector.

[0011] Optionally, the hybrid retrieval module is configured to retrieve work orders related to the query vector from the vector database, including: Vector similarity calculation and keyword similarity calculation are performed on the vector database and the query vector, and work orders related to the query vector are retrieved from the vector database.

[0012] Optionally, the hybrid retrieval module performs vector similarity calculation and keyword similarity calculation on the vector database and the query vector, and retrieves work orders related to the query vector from the vector database, including: Based on the query vector, use cosine distance to search for Top-K similar work orders in the vector space of the vector database; Perform keyword similarity matching on the query vector and the vector database to obtain several similar work orders; Perform inverted fusion on the Top-K similar work orders and the several similar work orders to ultimately determine the work orders related to the query vector.

[0013] Optionally, the large language model prediction module is used to perform semantic reasoning on the new work order and the related work orders to generate a repair spare parts prediction result, including: Assembling summary information of the new work order and the related work orders into a structured prompt word; The structured prompt words are input into a large language model for semantic reasoning to generate maintenance spare parts prediction results; wherein the maintenance context prediction results include the model and quantity of spare parts expected to be used.

[0014] Optionally, the large language model prediction module further sends the maintenance spare parts prediction result to the maintenance material preparation end; The maintenance material preparation terminal is used to obtain a spare parts list and a spare parts preparation waiting time period according to the maintenance spare parts prediction result.

[0015] Compared with the prior art, the present invention has the following beneficial effects: The present application provides a maintenance spare parts prediction system based on a large language model plug-in knowledge base, which uses a large language model to semantically encode historical maintenance records and new work orders, and combines it with a vector library to achieve efficient similarity retrieval. It can quickly find historical records that are most similar to the current fault scenario in millions of work orders, and significantly reduce maintenance waiting time compared to traditional spare parts prediction methods based on rules or historical averages. With the help of a large language model, new work orders and historical experience are integrated and reasoned, and detailed reasons for spare parts recommendations and past case basis are given in maintenance prompts, which significantly improves the efficiency of judging spare parts required for maintenance. By supporting the construction of an abnormal spare parts early warning mechanism based on historical data and prediction results, it can assist in discovering hidden design defects or material problems in equipment. Both the Embedding model and the large language model used support cross-product line expansion and have strong versatility and scalability. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work. Among them: Figure 1 This is a structural diagram of a maintenance spare parts prediction system based on a large language model plug-in knowledge base provided by the present invention.

[0017] Figure 2 It is the preprocessing process of historical work orders.

[0018] Figure 3 It is the process of building a vector database.

[0019] Figure 4 It is the process of retrieving the work orders related to the query vector. DETAILED DESCRIPTION

[0020] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are described in detail below in conjunction with the accompanying drawings. It will be understood that the specific embodiments described herein are only used to explain the present application, rather than to limit the present application. It should also be noted that, for ease of description, only some, rather than all, structures related to the present application are shown in the accompanying drawings. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0021] As used herein, the terms "comprise," "comprising," and "having," and any variations thereof, are intended to cover non-exclusive inclusions. For example, a process, system, product, or apparatus comprising a series of steps or elements is not limited to the listed steps or elements but may optionally include steps or elements not listed, or may optionally include other steps or elements that are inherent to the process, system, product, or apparatus.

[0022] References herein to "embodiments" mean that a particular feature, structure, or characteristic described in connection with the embodiments may be included in at least one embodiment of the present application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it constitute an independent or alternative embodiment that is mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0023] See also Figure 1 As shown, an embodiment of the present application provides a maintenance spare parts prediction system based on a large language model plug-in knowledge base. The maintenance spare parts prediction system based on a large language model plug-in knowledge base includes: The historical work order preprocessing module is used to extract and preprocess information data of historical work orders to obtain preprocessed text; The vector library construction module is used to convert the pre-processed text into semantic vectors and perform association and binding processing to obtain a vector database; A new work order conversion processing module is used to extract attribute information from the new work order, convert the attribute information into a semantic vector, and obtain a query vector; A hybrid retrieval module, used to retrieve work orders related to the query vector from the vector database; The large language model prediction module is used to perform semantic reasoning on new work orders and related work orders to generate maintenance spare parts prediction results.

[0024] The beneficial effects of the above embodiments are as follows: the maintenance spare parts prediction system based on a large language model plug-in knowledge base extracts and preprocesses information data from historical work orders to obtain preprocessed text; performs semantic vector conversion and association binding processing on the preprocessed text to obtain a vector database; extracts attribute information from new work orders, performs semantic vector conversion on the attribute information to obtain a query vector; retrieves work orders related to the query vector from the vector database; performs semantic reasoning on new work orders and related work orders to generate maintenance spare parts prediction results, makes full use of natural language fault descriptions and equipment model information in maintenance records, fully understands natural language information in the face of complexity and diversity, improves the accuracy and practicality of spare parts prediction, realizes reasonable allocation of maintenance resources, reduces spare parts inventory costs and inventory waste, and improves maintenance efficiency.

[0025] In another embodiment, the historical work order preprocessing module is used to extract and preprocess information data from historical work orders to obtain preprocessed text, including: Identify the information type of the data contained in historical work orders and extract the corresponding type of information data; The information data is subjected to abnormal elimination, text standardization and clustering processing in sequence to obtain the preprocessed text.

[0026] In practice, the maintenance system reads data and collects historical work orders for various equipment models, potentially numbering in the millions. Data is indexed and identified for each historical work order, obtaining index information for each data element within the historical work order. This allows the identification of the information type of all data within the historical work order, thereby extracting the corresponding information data. This information data may include, but is not limited to, fields such as a description of the fault phenomenon (using natural language), equipment model, maintenance history, and the type and data of the actual replacement parts. Given that historical work orders are filled out by different staff members, errors, duplications, and irregularities are inevitable in each historical work order's information data. Directly using this extracted information data for vector conversion and large language models can lead to prediction errors and confusion. To address this issue, the information data undergoes sequential outlier removal, text normalization, and clustering to generate preprocessed text. This ensures data standardization and uniformity, thereby improving the quality of vector library construction and prediction consistency.

[0027] In another embodiment, the historical work order preprocessing module sequentially performs abnormal elimination, text standardization, and clustering on the information data to obtain preprocessed text, including: Eliminate abnormal data from the information data; wherein the abnormal data includes at least one of a field missing portion, a semantically invalid portion, a fault description blank portion, and a logical conflict portion; Unify the text field format of information data and remove invalid stop words; A large language model is used to perform cluster analysis on fault-related fields in information data, and semantically similar field descriptions are aggregated and normalized to obtain preprocessed text.

[0028] See also Figure 2In order to provide a high-quality and unified data source for the construction of the vector library, the data parts or data segments with missing fields, invalid semantics, blank fault descriptions, and logical conflicts in the information data are first deleted, and then the text field format, unit, and invalid stop words of the information data are unified to ensure the standard consistency and simplicity of the expression of the information data; the Qwen3-235B large language model is used to perform cluster analysis on fault-related fields (such as fields describing fault phenomena), and semantically similar field descriptions are aggregated and normalized (for example, semantically similar field descriptions such as "unable to power on", "no power response", and "unable to power on" are aggregated into the same set), thereby improving the subsequent vector library construction quality and prediction consistency.

[0029] In another embodiment, the vector library construction module is used to perform semantic vector conversion and association binding processing on the preprocessed text to obtain a vector database, including: Extract key content fields from the preprocessed text, convert the key content fields into embedded vectors, and obtain semantic vectors; The semantic vector is associated with the historical work order corresponding to the preprocessed text and stored in the vector database.

[0030] See also Figure 3 , the Qwen-Embedding-0.8B large language model can be used to extract the key content fields of the preprocessed text, and the above large language model can be used to calculate the Embedding embedding vector for the key content fields corresponding to each historical work order, converting the natural language model into a semantic vector in the form of a dense vector to facilitate semantic retrieval. The semantic vector is also associated and bound with the historical work order corresponding to the preprocessed text and uniformly stored in the Milvus vector database. Considering that Milvus is an open source vector retrieval system, it supports approximate nearest neighbor retrieval of high-dimensional data and can support efficient calling of millions of samples, facilitating subsequent fast and accurate similarity retrieval in the Milvus vector database.

[0031] In another embodiment, the vector library construction module extracts key content fields of the preprocessed text, including: Extract the fault description field and device model field from the preprocessed text and use them as key content fields.

[0032] Using the Qwen-Embedding-0.8B large language model to extract the fault description field and device model field from the preprocessed text as key content fields can ensure that the semantic vector accurately covers the key content of historical work orders and improve the data comprehensiveness and effectiveness of the vector database.

[0033] In another embodiment, the new work order conversion processing module is used to extract attribute information from the new work order, perform semantic vector conversion on the attribute information, and obtain a query vector, including: Extract the fault description field and device model field from the new work order as attribute information; The attribute information is converted into an embedding vector using a large language model to obtain a query vector.

[0034] When the maintenance system receives a new work order, it first performs text recognition on the new work order, extracting the fault description and device model fields within the new work order, accurately and comprehensively refining the new work order's key information. It also uses the Qwen-Embedding-0.8B large language model to calculate embedding vectors for the fault description and device model fields, converting the natural language model into a dense semantic vector. This generates a query vector, thus transforming the new work order content from natural language to a dense semantic vector.

[0035] In another embodiment, the hybrid retrieval module is used to retrieve work orders related to the query vector from the vector database, including: Perform vector similarity calculation and keyword similarity calculation on the vector database and the query vector, and retrieve work orders related to the query vector from the vector database.

[0036] A hybrid similarity search is performed on the vector database and query vector at both the vector similarity and keyword similarity levels. Combining vector similarity search with full-text similarity search, multi-dimensional matching is performed to ensure that historical work orders related to the query vector are quickly and accurately retrieved.

[0037] In another embodiment, the hybrid retrieval module performs vector similarity calculation and keyword similarity calculation on the vector database and the query vector, and retrieves work orders related to the query vector from the vector database, including: Based on the query vector, use the cosine distance to find the top-K similar work orders in the vector space of the vector database; Perform keyword similarity matching on the query vector and the vector database to obtain several similar work orders; Perform inverted fusion on the Top-K similar work orders and several similar work orders to ultimately determine the work orders related to the query vector.

[0038] See also Figure 4In practice, we use cosine distance calculation to find the top-K similar tickets (i.e., the first K most similar tickets) corresponding to the query vector in the vector space of the vector database. Traditional text retrieval algorithms, such as BM25 or TF-IDF, are then used to perform keyword similarity matching between the query vector and the vector database to obtain a number of similar tickets. Using a weighted strategy, we perform an inverted fusion of these two search results to ultimately determine the tickets relevant to the query vector. This process takes into account both semantic and keyword relevance.

[0039] In another embodiment, the large language model prediction module is used to perform semantic reasoning on the new work order and related work orders to generate a repair parts prediction result, including: Assemble the summary information of the new work order and related work orders into a structured prompt word; The structured prompt words are input into the large language model for semantic reasoning to generate maintenance spare parts prediction results; among them, the maintenance context prediction results include the model and quantity of spare parts expected to be used.

[0040] The summary information of the new work order and the retrieved related work orders is assembled into a structured prompt (i.e., a structured prompt). The structured prompt is input into the Qwen3-235B large language model for semantic reasoning to generate a maintenance spare parts prediction result that includes the expected spare parts model and quantity. This accurately identifies the type and quantity of spare parts required for equipment maintenance, providing a reliable reference for maintenance centers to prepare spare parts.

[0041] In another embodiment, the large language model prediction module further sends the maintenance spare parts prediction result to the maintenance and material preparation end; The maintenance and material preparation end is used to obtain the spare parts list and the spare parts preparation waiting time period based on the maintenance spare parts prediction results.

[0042] The large language model prediction module sends the maintenance spare parts prediction results to the maintenance and material preparation end (such as the maintenance and material preparation service system). In this way, the maintenance and material preparation end is used to obtain the spare parts list and the spare parts preparation waiting time period based on the maintenance spare parts prediction results, thereby automatically recommending the spare parts list, preparing materials in advance to reduce the waiting period, and comparing with the actual spare parts consumption to perform error learning and model feedback iteration, supporting subsequent abnormal consumables analysis and prediction model training and other functions.

[0043] In general, the maintenance spare parts prediction system based on the large language model plug-in knowledge base extracts and preprocesses information data from historical work orders to obtain preprocessed text; performs semantic vector conversion and association binding processing on the preprocessed text to obtain a vector database; extracts attribute information from new work orders, performs semantic vector conversion on the attribute information to obtain a query vector; retrieves work orders related to the query vector from the vector database; performs semantic reasoning on new work orders and related work orders to generate maintenance spare parts prediction results, making full use of natural language fault descriptions and equipment model information in maintenance records, fully understanding natural language information when faced with complexity and diversity, improving the accuracy and practicality of spare parts prediction, realizing reasonable allocation of maintenance resources, reducing spare parts inventory costs and inventory waste, and improving maintenance efficiency.

[0044] The above is only a specific embodiment of the present invention, and any other improvements made based on the concept of the present invention are considered to be within the protection scope of the present invention.

Claims

1. A maintenance spare parts prediction system based on a large language model plug-in knowledge base, characterized in that: include: The historical work order preprocessing module is used to extract and preprocess information data of historical work orders to obtain preprocessed text; A vector database construction module is used to perform semantic vector conversion and association binding processing on the preprocessed text to obtain a vector database; A new work order conversion processing module is used to extract attribute information from the new work order, perform semantic vector conversion on the attribute information, and obtain a query vector; a hybrid retrieval module, configured to retrieve work orders related to the query vector from the vector database; A large language model prediction module, configured to perform semantic reasoning on the new work order and the related work orders to generate a maintenance spare parts prediction result; The historical work order preprocessing module is used to extract and preprocess information data of historical work orders to obtain preprocessed text, including: Identify the information type of the data contained in historical work orders and extract the corresponding type of information data; The information data is sequentially subjected to abnormality elimination, text standardization, and clustering processing to obtain a preprocessed text; The historical work order preprocessing module sequentially performs abnormal elimination, text standardization, and clustering on the information data to obtain preprocessed text, including: Eliminating abnormal data from the information data; wherein the abnormal data includes at least one of a field missing portion, a semantically invalid portion, a fault description blank portion, and a logical conflict portion; Performing text field format unification processing and invalid stop word removal processing on the information data; A large language model is used to perform fault-related field clustering analysis on the information data, and semantically similar field descriptions are aggregated and normalized to obtain preprocessed text.

2. The maintenance spare parts prediction system based on a large language model plug-in knowledge base according to claim 1, characterized in that: The vector library construction module is used to perform semantic vector conversion and association binding processing on the preprocessed text to obtain a vector database, including: Extracting key content fields of the preprocessed text, and performing embedding vector conversion on the key content fields to obtain semantic vectors; The semantic vector is associated and bound with the historical work order corresponding to the pre-processed text, and then stored in a vector database.

3. The maintenance spare parts prediction system based on a large language model plug-in knowledge base according to claim 2, characterized in that: The vector library construction module extracts key content fields of the preprocessed text, including: The fault description field and the device model field of the preprocessed text are extracted as key content fields.

4. The maintenance spare parts prediction system based on a large language model plug-in knowledge base according to claim 1, characterized in that: The new work order conversion processing module is used to extract attribute information from the new work order, perform semantic vector conversion on the attribute information, and obtain a query vector, including: Extract the fault description field and device model field from the new work order as attribute information; The attribute information is converted into an embedding vector using a large language model to obtain a query vector.

5. The maintenance spare parts prediction system based on a large language model plug-in knowledge base according to claim 1, characterized in that: The hybrid retrieval module is used to retrieve work orders related to the query vector from the vector database, including: Vector similarity calculation and keyword similarity calculation are performed on the vector database and the query vector, and work orders related to the query vector are retrieved from the vector database.

6. The maintenance spare parts prediction system based on a large language model plug-in knowledge base according to claim 5, characterized in that: The hybrid retrieval module performs vector similarity calculation and keyword similarity calculation on the vector database and the query vector, and retrieves work orders related to the query vector from the vector database, including: Based on the query vector, use cosine distance to search for Top-K similar work orders in the vector space of the vector database; Perform keyword similarity matching on the query vector and the vector database to obtain several similar work orders; Perform inverted fusion on the Top-K similar work orders and the several similar work orders to ultimately determine the work orders related to the query vector.

7. The maintenance spare parts prediction system based on a large language model plug-in knowledge base according to claim 1, characterized in that: The large language model prediction module is used to perform semantic reasoning on the new work order and the related work orders to generate a maintenance spare parts prediction result, including: Assembling summary information of the new work order and the related work orders into a structured prompt word; The structured prompt words are input into a large language model for semantic reasoning to generate maintenance spare parts prediction results; wherein the maintenance context prediction results include the model and quantity of spare parts expected to be used.

8. The maintenance spare parts prediction system based on a large language model plug-in knowledge base according to claim 7, characterized in that: The large language model prediction module also sends the maintenance spare parts prediction result to the maintenance material preparation end; The maintenance material preparation terminal is used to obtain a spare parts list and a spare parts preparation waiting time period according to the maintenance spare parts prediction result.

Images