Commercial dishwashing equipment fault knowledge base construction method and system

Abstract

The invention belongs to the technical field of data processing, and particularly relates to a commercial dishwashing equipment fault knowledge base construction method and system. The method comprises the following steps: identifying an equipment component entity and a fault code entity by acquiring fault text data, extracting noun phrases as candidate anaphora items, and calculating a co-reference score for an anaphora pair formed by two candidate anaphora items; an incremental method is adopted to construct a co-reference chain, a to-be-clustered reference item is linked to the co-reference chain where the antecedent with the highest score is located, the co-reference score and the internal compactness of the co-reference chain are considered during decision making, the co-reference chain containing a plurality of different fault code entities is segmented after construction, and based on the segmented co-reference chain aggregation information, the fault code entities are clustered according to the segmented co-reference chain aggregation information. And constructing a knowledge triple containing fault phenomena, reasons and solutions, and forming a structured fault knowledge base. According to the method, domain knowledge features and a co-reference chain segmentation mechanism are introduced, the problem of entity reference confusion in a multi-fault scene is solved, and the accuracy of fault knowledge base construction is improved.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, specifically relating to a method and system for constructing a fault knowledge base for commercial dishwasher equipment. Background Technology

[0002] Commercial dishwashers are commonly used in restaurants, hotels, and canteens, serving as essential core equipment for ensuring daily operations. Through long-term equipment maintenance and customer service, the industry has accumulated massive amounts of unstructured text data, primarily encompassing maintenance manuals, technical documents, after-sales service records, and user feedback reports. This text data contains a wealth of valuable knowledge about fault phenomena, causes, and solutions. Automatically extracting structured fault knowledge from this heterogeneous and fragmented text to build a computer-usable fault knowledge base would have significant application value in improving the intelligence level of equipment operation and maintenance and developing intelligent diagnostic systems.

[0003] However, these technical documents are typically written by different engineers, resulting in significant differences in writing style and a prevalence of pronouns, ellipses, and alternative names. Within the same document, different terms such as "main wash pump," "this pump," and the personal pronoun "it" may refer to the same equipment component. This ambiguity constitutes a major obstacle to information extraction. Failure to accurately identify all expressions referring to the same entity in the text will directly lead to broken or incorrectly aggregated fault information, severely impacting the accuracy and completeness of the knowledge base construction.

[0004] Existing coreference resolution methods mostly rely on general linguistic features such as lexical characteristics, syntactic structure, and semantic similarity for judgment. These general models lack an understanding of domain-specific knowledge and cannot effectively utilize the hierarchical relationships between equipment components or the meaning of specific fault codes to assist in judgment. For example, the models struggle to understand the logical relationships inherent in areas such as nozzles being sub-components of the spray arm, or specific error codes specifically indicating heating faults, resulting in low accuracy when dealing with technical terms and implicit associations.

[0005] Furthermore, a single fault report in a real-world scenario may record multiple related or independent fault issues, each involving different components and fault codes. Traditional coreference chain construction methods struggle to reliably distinguish entity references belonging to different fault contexts, easily leading to their erroneous clustering into the same coreference chain, causing confusion and misleading information. Therefore, devising a coreference resolution method that deeply integrates equipment domain expertise and effectively handles complex fault scenarios, thereby improving the level and quality of commercial dishwasher equipment fault knowledge base construction, is a pressing technical challenge in this field. Summary of the Invention

[0006] The purpose of this invention is to propose a method and system for constructing a fault knowledge base for commercial dishwashers, in order to solve the problem that existing technologies, when constructing a fault knowledge base for commercial dishwashers from unstructured maintenance texts, suffer from low accuracy in fault information extraction and incorrect associations due to the lack of domain knowledge understanding in general models and the difficulty in distinguishing complex referential relationships under multiple fault scenarios.

[0007] To address the above problems, the technical solution for constructing a fault knowledge base for commercial dishwashers proposed in this invention is as follows: The method for building a fault knowledge base for commercial dishwasher equipment includes the following steps: Acquire text data of commercial dishwasher equipment malfunctions, preprocess and identify equipment component entities and fault code entities, and extract noun phrases as candidate references; calculate the coreference score of any two reference pairs consisting of the candidate references. An incremental method is used to construct coreference chains. The candidate references are traversed in text order. The currently processed candidate reference is taken as the reference to be clustered. The candidate references that appear before the reference to be clustered are taken as potential antecedents. The link decision score between the reference to be clustered and each potential antecedent is calculated. The reference to be clustered is linked to the coreference chain of the potential antecedent with the highest link decision score. The constructed core chain is post-processed. When a single core chain contains more than one different fault code entity, the core chain is split according to the different fault code entities to form multiple sub-chains. Based on the fault-related information aggregated from the segmented sub-chains, a knowledge triple containing fault phenomena, causes, and solutions is constructed to obtain a fault knowledge base for commercial dishwasher equipment.

[0008] Furthermore, the acquisition of commercial dishwasher equipment fault text data, after preprocessing to identify equipment component entities and fault code entities, and extracting noun phrases as candidate referents, includes: Denoising, sentence segmentation, and word segmentation preprocessing are performed on faulty text data; A bidirectional long short-term memory network-conditional random field model is used to identify equipment component entities and fault code entities in the fault text data. Noun phrases in fault text data are extracted as candidate pronouns based on part-of-speech tagging and chunking analysis rules.

[0009] Further, calculating the coreference score of a reference pair consisting of any two candidate references includes: Let the two candidate referents in the referent pair be denoted as and Mapping to the device component hierarchy map, calculate the shortest path distance between the two in the map. The value after exponential transformation is used as the domain knowledge feature value; Will and Input the pre-trained language model to obtain word vector representations and The cosine similarity of word vectors is calculated as a semantic feature value; statistics and Number of sentence intervals in the faulty text data The text distance decay feature value is calculated by applying the exponential decay function; The co-exponential score is obtained by linearly weighting and summing the three eigenvalues ​​mentioned above. .

[0010] Furthermore, the calculation process for the link decision score is as follows: Two candidate references and In As a descriptor for the cluster to be clustered As a potential antecedent, calculation and Coreference score between ; calculate The core chain Internal tightness The internal tightness is the arithmetic mean of the co-reference scores of all reference pairs within the co-reference chain; Commonly pointed out With internal tightness The link decision is obtained by performing a linear weighted combination. .

[0011] Furthermore, the incremental method for constructing corelink chains includes: Each candidate pronoun to be clustered is processed sequentially according to the order of appearance of the candidate pronouns in the faulty text data from left to right. For the current reference to be clustered, all candidate references that have appeared before the reference to be clustered are taken as the set of potential antecedents; Calculate the link decision score between the current clustering reference and each potential antecedent in the set of potential antecedents; If the highest value of the link decision score is greater than the preset threshold, the current reference to be clustered is added to the core association chain of the potential antecedent corresponding to the highest score; otherwise, a new core association chain is created for the current reference to be clustered.

[0012] Furthermore, the step of segmenting the corelink chain according to the different fault code entities to form multiple sub-chains includes: Traverse all candidate references in the core reference chain in text order; Starting from the beginning of the coreference chain, candidate references in the coreference chain are sequentially assigned to a current subchain until the first fault code entity is encountered. The fault code entity is then assigned to the current subchain, thus completing the construction of the current subchain. Starting with the candidate reference immediately following the previous fault code entity, a new subchain is constructed, and subsequent candidate references are sequentially assigned to this new subchain until the next different fault code entity is encountered, and the different fault code entity is assigned to the new subchain. Repeat the above process until all candidate references in the coreference chain are assigned to the corresponding subchains.

[0013] Furthermore, repeating the above process until all candidate references in the core reference chain are divided into corresponding sub-chains also includes: if no new different fault code entities are encountered after traversing to the end of the core reference chain, then all remaining candidate references are divided into the last sub-chain separately, ensuring that each sub-chain contains at most one fault code entity.

[0014] Furthermore, based on the fault-related information aggregated from the segmented sub-chains, a knowledge triple containing fault phenomena, causes, and solutions is constructed, including: For each segmented subchain, determine the unique fault code entity contained in the subchain; In the original sentence set corresponding to the candidate referents in the sub-chain, the expression related to "phenomenon is" and "displayed as" is extracted as the fault phenomenon, the expression related to "reason" and "due to" is extracted as the fault cause, and the expression related to "solution is" and "needs to be replaced" is extracted as the solution. The extracted fault phenomena, causes, and solutions are combined into knowledge triples and associated with the unique fault code entity corresponding to the sub-chain.

[0015] Furthermore, the structure of the bidirectional long short-term memory network-conditional random field model, from bottom to top, includes an embedding layer, a bidirectional long short-term memory network layer, and a conditional random field layer.

[0016] The technical solution for the commercial dishwasher equipment fault knowledge base construction system proposed in this invention is as follows: A system for building a fault knowledge base for commercial dishwashers includes a processor and a memory; the memory stores computer program instructions, which, when executed by the processor, implement the method for building a fault knowledge base for commercial dishwashers in any of the above technical solutions.

[0017] The beneficial effects of this invention are as follows: By integrating a preset equipment component hierarchy map into the coreference score calculation stage, this invention successfully inputs prior knowledge such as the structural relationship between equipment components into the coreference resolution model, thereby improving the accuracy of identifying the referential relationship between professional terms and component aliases. This enables the model to use subordinate logical relationships between components for auxiliary judgment, overcoming the limitations of relying solely on literal similarity or general semantic features.

[0018] When deciding whether to add a reference to a potential antecedent to the coreference chain, this invention not only considers the local coreference score between the two, but also introduces the global index of the internal tightness of the coreference chain. This ensures the semantic and logical consistency between the newly added member and the existing members in the chain, effectively suppressing the chain drift phenomenon that is common in incremental clustering and significantly reducing the risk of incorrect links.

[0019] This invention divides a long common index chain into multiple independent sub-chains based on the order in which different fault code entities appear, reliably distinguishing multiple different fault events involved in the same document. This effectively prevents erroneous association and confusion of information on the phenomena, causes, and solutions of different faults, ensuring the logical correctness of knowledge extraction.

[0020] This invention, based on the aggregated information of the clean sub-chains after segmentation, can accurately extract and construct knowledge triples containing fault phenomena, causes, and solutions. This not only solves the problem of the difficulty in utilizing massive amounts of heterogeneous text, but also provides a complete and clear data foundation and knowledge support for the subsequent development of intelligent diagnostic systems and the improvement of the intelligence level of equipment operation and maintenance. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the steps of the method for constructing a fault knowledge base for commercial dishwasher equipment according to the present invention. Detailed Implementation

[0022] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0023] Specific embodiments of the method for constructing a fault knowledge base for commercial dishwashers proposed in this invention: like Figure 1 As shown, the method for constructing a fault knowledge base for commercial dishwasher equipment includes steps S1-S4, which are as follows: S1. Obtain text data of commercial dishwasher equipment malfunctions, preprocess and identify equipment component entities and fault code entities, and extract noun phrases as candidate references; calculate the coreference score of any two reference pairs consisting of the candidate references.

[0024] Specifically, raw data such as maintenance manuals and after-sales records are collected through web scraping or exporting from internal enterprise databases. The preprocessing stage involves text cleaning, removing irrelevant characters and formatting, and performing sentence segmentation and part-of-speech tagging. Using pre-labeled fault domain corpora, a BiLSTM-CRF model combining a bidirectional long short-term memory network and a conditional random field is trained. This model automatically identifies equipment component entities in the text, such as the main wash pump and spray arm, as well as fault code entities, such as E01 and E04. Based on the part-of-speech tagging results, a noun phrase segmentation technique is used to identify nouns, pronouns, and phrases composed of adjectives and nouns in the text, serving as a candidate set for subsequent coreference resolution.

[0025] A hierarchical graph representing the relationships between components of a commercial dishwasher is pre-constructed. For example, the spray system is the parent node of the main wash pump, and the nozzles are the child nodes of the spray arms. For any pair of candidate referents, the domain knowledge feature is represented by calculating the shortest path distance between the corresponding entities of the two candidate referents in the hierarchical graph; the closer the distance, the higher the score. Semantic features are obtained using the cosine similarity of word vectors obtained through the BERT model. The text distance decay feature calculates the number of sentences separating the two candidate referents in the text; the fewer the sentences, the higher the score. The coreference score is a comprehensive score obtained by normalizing the above three types of feature values ​​and then weighting and summing them according to the set weights.

[0026] In an optional embodiment, the step of acquiring commercial dishwasher equipment fault text data, preprocessing it to identify equipment component entities and fault code entities, and extracting noun phrases as candidate referents includes: Denoising, sentence segmentation, and word segmentation preprocessing are performed on faulty text data; A bidirectional long short-term memory network-conditional random field model is used to identify equipment component entities and fault code entities in the fault text data. Noun phrases in fault text data are extracted as candidate pronouns based on part-of-speech tagging and chunking analysis rules.

[0027] The original maintenance record text is cleaned. For example, for the input text E-101 pump # vibration fault code F01, irrelevant symbols such as well numbers are removed, and the text is segmented into sentences and words. After processing, the sequence E-101 pump vibration fault code F01 is obtained. This word sequence is then input into the bidirectional long short-term memory network-conditional random field model.

[0028] Optionally, the structure of the bidirectional long short-term memory network-conditional random field model, from bottom to top, includes an embedding layer, a bidirectional long short-term memory network layer, and a conditional random field layer. The embedding layer is responsible for converting each word in the input text sequence into a low-dimensional, dense word vector. The bidirectional long short-term memory network layer consists of a forward long short-term memory network and a backward long short-term memory network, processing the word vector sequence from both directions respectively, and concatenating the hidden state vectors from both directions to generate a feature representation for each word that incorporates contextual information. The conditional random field layer, located at the top layer, receives the output features from the bidirectional long short-term memory network layer, learns the transition probabilities between labels, and decodes to obtain the globally optimal entity label sequence. The training set of this model consists of a large amount of manually annotated equipment maintenance record text. During the annotation process, each word in the text is assigned a predefined label, such as using the BIO annotation system, where B-DEV represents the beginning of an equipment component entity, I-DEV represents the middle part of an equipment component entity, and O represents that it does not belong to any entity type. The training process is conducted end-to-end. The model's objective is to minimize the negative log-likelihood loss function. Backpropagation and optimizers such as Adam are used to update the model parameters until convergence. The model input is a sequence of maintenance record text after word segmentation preprocessing, such as "E-101 pump," "occur," and "vibration." The model output is a sequence of entity labels of the same length as the input sequence, such as "B-DEV," "O," and "O."

[0029] The Conditional Random Field (CRF) layer decodes the globally optimal entity label sequence based on feature vectors and label transition probabilities. For example, the model might identify pump E-101 as an equipment component entity and F01 as a fault code entity. Part-of-speech tagging and chunking techniques are used to identify and extract all noun phrases from the text. In the example above, pump E-101 and the fault code are identified as noun phrases, serving as candidate references for subsequent coreference resolution.

[0030] In an optional embodiment, calculating the coreference score of a pair of references consisting of any two candidate references includes: Let the two candidate referents in the referent pair be denoted as and Mapping to the device component hierarchy map, calculate the shortest path distance between the two in the map. The value after exponential transformation is used as the domain knowledge feature value; Will and Input the pre-trained language model to obtain word vector representations and The cosine similarity of word vectors is calculated as a semantic feature value; statistics and Number of sentence intervals in the faulty text data The text distance decay feature value is calculated by applying the exponential decay function; The co-exponential score is obtained by linearly weighting and summing the three eigenvalues ​​mentioned above. The specific calculation formula is as follows: ; in, These are the weight coefficients for each feature, and the sum of the three is 1; This is the distance attenuation factor.

[0031] To calculate two candidate pronouns, for example For pumps and To calculate the coreference score between the devices, we leverage domain knowledge, assuming that in a pre-built device component hierarchy graph, the pump is a subclass of the device, and the shortest path distance between them is calculated. If the value is 1, then the domain knowledge feature value is... The value is approximately 0.368. Then, semantic similarity is calculated by inputting the pump and the device into pre-trained language models such as BERT to obtain their word vector representations. and Assuming the calculated cosine similarity A value of 0.95 indicates a high semantic correlation between the two. Next, the text distance is calculated. Assuming the pump appears in the second sentence and the device in the third sentence, the sentence interval is... The distance attenuation factor is 1. If we set it to 0.1, then the text distance decay feature value is... It is approximately equal to 0.905. Set the weighting coefficients, for example... The co-index score is calculated using a linear weighted summation method, and the result is: A high score indicates that the pump and the device likely refer to the same entity.

[0032] In this step, the pre-trained language model is built on the Transformer encoder architecture. Its core structure consists of multiple stacked encoder units, each containing a multi-head self-attention mechanism and a feedforward neural network. The model is pre-trained on massive amounts of general text corpora and fine-tuned on specialized data in fields such as equipment repair. Pre-training employs self-supervised learning, with the main task being a masked language model, i.e., predicting randomly masked words in a sentence. The training objective is to minimize the prediction loss. The model's input is a text sequence, such as a candidate pronoun pump, which is segmented and converted into a sequence of numeric indices. The model's output is a context-dependent high-dimensional word vector corresponding to each word in the input sequence, such as a 768-dimensional floating-point vector, encoding the semantic information of the word in a specific context.

[0033] S2. Construct coreference chains using an incremental method. Traverse the candidate references in text order, take the currently processed candidate reference as the reference to be clustered, take the candidate references that appeared before the reference to be clustered as potential antecedents, calculate the link decision score between the reference to be clustered and each potential antecedent, and link the reference to be clustered to the coreference chain where the potential antecedent with the highest link decision score is located.

[0034] This step processes each candidate pronoun sequentially from left to right in the faulty text data. For the current candidate pronoun, all candidate pronouns preceding it are treated as a set of potential antecedents; the link decision score between the current candidate pronoun and each potential antecedent in the set is calculated. If the highest link decision score is greater than a preset threshold, the current candidate pronoun is added to the coreference chain of the potential antecedent corresponding to the highest score; otherwise, a new coreference chain is created for the current candidate pronoun.

[0035] In an optional embodiment, the calculation process of the link decision score is as follows: Two candidate references and In As a descriptor for the cluster to be clustered As a potential antecedent, calculation and Coreference score between ; calculate The core chain Internal tightness The internal tightness is the arithmetic mean of the co-reference scores of all reference pairs within the co-reference chain; Commonly pointed out With internal tightness The link decision is obtained by performing a linear weighted combination. .

[0036] Specifically, internal density The calculation formula is as follows: ;in, , For corelinks Members of, For corelinks The number of members, and Not less than 2.

[0037] Commonly pointed out With internal tightness The link decision is obtained by performing a linear weighted combination. The calculation formula is as follows: ; in, and For preset weighting coefficients, and These are the clustering references and potential antecedents, respectively.

[0038] Example 1: Assume For the clustering reference, and For it, there exists a constructed coreference chain. Contains two referents and , It is an E-101 compressor. For this compressor. Now a decision needs to be made as to whether to... Link to In the middle, select the last member in the chain. As a potential antecedent. Calculation. and The coreference score, i.e. Assume the calculated score is 0.85. Calculate the corefinite chains. Internal tightness There is only one pair of referential terms in this chain, namely and Assuming The calculated value is 0.90, therefore the internal tightness of this chain is... That is, 0.90. Set the weighting coefficient. Calculate the link decision score The score takes into account the direct relationship between the new referent and the members in the chain, as well as the stability of the chain itself.

[0039] Example 2: Suppose that three candidate references appear in the text in sequence. , and ,in, It is an E-101 pump. For this device, For this purpose, the preset link threshold T is 0.6. The processing flow begins by processing... .because It is the first candidate referent, and there is no potential antecedent preceding it, therefore it is... Create a new cofinite chain Then process. The set of potential antecedents is .calculate and Link decision segment Assume the calculated result is 0.82. Because 0.82 is greater than the threshold of 0.6, therefore... Join The core chain In the middle, after the update Next, proceed with the processing. The set of potential antecedents is Calculate the decision scores of the two links. and Assuming the calculated results are 0.75 and 0.88 respectively, the highest link decision score of 0.88 is selected. The antecedent corresponding to this score is... Since 0.88 is also greater than the threshold of 0.6, join in The core chain In the middle, we obtain the co-point chain. .

[0040] S3. Post-process the constructed corelink chain. When a single corelink chain contains more than one different fault code entity, the corelink chain is split according to the different fault code entities to form multiple sub-chains.

[0041] Specifically, after processing all candidate references in the coreference chain, each generated coreference chain is traversed. It checks if there are two or more different fault code entities within the chain. If so, the original coreference chain is divided into multiple sub-chains according to the order in which the candidate references appear in the text, using the first occurrence of each different fault code entity as the splitting point. Specifically: starting from the beginning of the chain, references are added sequentially to the current sub-chain until a fault code entity is encountered, denoted as . And add the entity to the current subchain; continue traversing, if the fault code entity encountered later is related to... If they are the same, continue adding them to the current subchain; if a different fault code entity is encountered, it is denoted as... Then Starting with this, we begin building a new subchain and will Add a new subchain. Repeat this process until the entire core pointer chain has been traversed. If no new distinct fault code entity is encountered by the end of the chain, then divide all remaining candidate references into a separate subchain, ensuring that each subchain contains at most one fault code entity.

[0042] Example 3: Suppose that coreference resolution yields a mixed coreference chain containing device components and fault codes, with members arranged in the order of their text appearance. .in, For compressors, For this device, The fault code entity is F005. For it, It is the fault code entity F012, for this machine. The segmentation process starts from the first member of the chain and traverses. Add and to the first sub-chain. When traversing to , it is recognized as a fault code entity F005. Add to the first sub-chain as well, completing the construction of the first sub-chain to obtain , which is associated with the fault code F005. Starting from the next member of , a new sub-chain is constructed. Add to the second sub-chain. When traversing to , it is recognized as another different fault code entity F012. Then add to the second sub-chain, completing the construction to obtain , which is associated with the fault code F012. Process the remaining and put it into the third sub-chain

[0043] . By this method, a long co-reference chain is segmented into multiple short sub-chains related to specific fault events.

[0044] S4. Aggregate fault-related information based on the segmented sub-chains, construct knowledge triples including fault phenomena, causes, and solutions, and obtain a commercial dishwashing equipment fault knowledge base.

[0045] Specifically, each co-reference chain is regarded as an independent fault event. For a certain co-reference chain, extract the sentences where all candidate referents in the chain are located to form a sentence group related to this fault event. Plan rule templates based on keyword and dependency syntactic analysis, and extract knowledge triples from the sentence group. For example, match sentences containing fault codes and phenomenon representation words such as display and manifested as, and extract fault phenomena; match sentences containing words such as cause and due to, and extract fault causes; match sentences containing operational verbs such as replace, adjust, and clean, and extract solutions. Associate the extracted fault phenomena, fault causes, and solutions with the core entities in the chain, such as fault codes or key components, to form triples such as <E01, fault cause, heating tube blockage>. The set of all triples is stored in a graph database or a relational database to constitute a complete commercial dishwashing equipment fault knowledge base.

[0045] In an optional embodiment, the aggregating fault-related information based on the segmented co-reference chains to construct knowledge triples including fault phenomena, causes, and solutions includes: For each segmented sub-chain, determine the unique fault code entity included in the sub-chain; In the original sentence set corresponding to the candidate referents in the sub-chain, the expression related to "phenomenon is" and "displayed as" is extracted as the fault phenomenon, the expression related to "reason" and "due to" is extracted as the fault cause, and the expression related to "solution is" and "needs to be replaced" is extracted as the solution. The extracted fault phenomena, causes, and solutions are combined into knowledge triples and associated with the unique fault code entity corresponding to the sub-chain.

[0046] The subchains obtained from the splitting in step S3 For example, we determine that the unique fault code entity associated with this sub-chain is F005, and collect the sentences in the original text corresponding to all candidate referents in this sub-chain. Assume the sentences are as follows: Sentence A1 indicates that the compressor displays low outlet pressure, log code F005; Sentence A2 states the cause is worn seals; Sentence A3 states the solution is to replace the sealing components of the equipment.

[0047] The above sentences are analyzed using keyword template matching rules. In sentence A1, the template is matched, and the extracted fault symptom is low outlet pressure. In sentence A2, the template is matched for the cause, and the extracted cause is worn seal. In sentence A3, the template indicates replacement is needed, and the extracted solution is to replace the sealing assembly. The extracted structured information is then associated with fault code F005 to form knowledge triples, such as F005, fault symptom: low outlet pressure; F005, fault cause: worn seal; and F005, solution: replace the sealing assembly. These triples can be stored in a knowledge graph for subsequent fault diagnosis and intelligent question answering.

[0048] Specific embodiments of the commercial dishwasher equipment fault knowledge base construction system proposed in this invention: The commercial dishwasher equipment fault knowledge base construction system includes a processor and a memory; the memory stores computer program instructions, and when the computer program instructions are run by the processor, the commercial dishwasher equipment fault knowledge base construction method in the above embodiments is implemented.

[0049] The commercial dishwasher equipment fault knowledge base construction system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces. Their settings and functions are known in the art and will not be described in detail here.

[0050] While various embodiments of the invention have been shown and described in this specification, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will occur to those skilled in the art without departing from the spirit and essence of the invention.

Claims

1. A method for constructing a fault knowledge base for commercial dishwasher equipment, characterized in that, Includes the following steps: Acquire text data of commercial dishwasher equipment malfunctions, preprocess and identify equipment component entities and fault code entities, and extract noun phrases as candidate references; calculate the coreference score of any two reference pairs consisting of the candidate references. An incremental method is used to construct coreference chains. The candidate references are traversed in text order. The currently processed candidate reference is taken as the reference to be clustered. The candidate references that appear before the reference to be clustered are taken as potential antecedents. The link decision score between the reference to be clustered and each potential antecedent is calculated. The reference to be clustered is linked to the coreference chain of the potential antecedent with the highest link decision score. The constructed core chain is post-processed. When a single core chain contains more than one different fault code entity, the core chain is split according to the different fault code entities to form multiple sub-chains. Based on the fault-related information aggregated from the segmented sub-chains, a knowledge triple containing fault phenomena, causes, and solutions is constructed to obtain a fault knowledge base for commercial dishwasher equipment.

2. The method for constructing a fault knowledge base for commercial dishwashers according to claim 1, characterized in that, The process involves acquiring textual data about faults in commercial dishwashers, preprocessing it to identify equipment component entities and fault code entities, and extracting noun phrases as candidate referents, including: Denoising, sentence segmentation, and word segmentation preprocessing are performed on faulty text data; A bidirectional long short-term memory network-conditional random field model is used to identify equipment component entities and fault code entities in the fault text data. Noun phrases in fault text data are extracted as candidate pronouns based on part-of-speech tagging and chunking analysis rules.

3. The method for constructing a fault knowledge base for commercial dishwashers according to claim 1, characterized in that, For any two candidate references, the coreference score of the reference pair is calculated, including: Let the two candidate references in the reference pair be denoted as and Mapping to the device component hierarchy map, calculate the shortest path distance between the two in the map. The value after exponential transformation is used as the domain knowledge feature value; Will and Input the pre-trained language model to obtain word vector representations and The cosine similarity of word vectors is calculated as a semantic feature value; statistics and Number of sentence intervals in the faulty text data The text distance decay feature value is calculated by applying the exponential decay function; The co-exponential score is obtained by linearly weighting and summing the three eigenvalues ​​mentioned above. .

4. The method for constructing a fault knowledge base for commercial dishwashers according to claim 3, characterized in that, The calculation process for the link decision score is as follows: Two candidate references and In As a descriptor for the cluster to be clustered As a potential antecedent, calculation and Coreference score between ; calculate The core chain Internal tightness The internal tightness is the arithmetic mean of the co-reference scores of all reference pairs within the co-reference chain; Commonly pointed out With internal tightness The link decision is obtained by performing a linear weighted combination. .

5. The method for constructing a fault knowledge base for commercial dishwashers according to claim 1, characterized in that, The incremental method for constructing corelink chains includes: Each candidate pronoun to be clustered is processed sequentially according to the order of appearance of the candidate pronouns in the faulty text data from left to right. For the current reference to be clustered, all candidate references that have appeared before the reference to be clustered are taken as the set of potential antecedents; Calculate the link decision score between the current clustering reference and each potential antecedent in the set of potential antecedents; If the highest value of the link decision score is greater than the preset threshold, the current reference to be clustered is added to the core association chain of the potential antecedent corresponding to the highest score; otherwise, a new core association chain is created for the current reference to be clustered.

6. The method for constructing a fault knowledge base for commercial dishwashers according to claim 1, characterized in that, The step of segmenting the corelink chain according to the different fault code entities to form multiple sub-chains includes: Traverse all candidate references in the core reference chain in text order; Starting from the beginning of the coreference chain, candidate references in the coreference chain are sequentially assigned to a current subchain until the first fault code entity is encountered. The fault code entity is then assigned to the current subchain, thus completing the construction of the current subchain. Starting with the candidate reference immediately following the previous fault code entity, a new subchain is constructed, and subsequent candidate references are sequentially assigned to this new subchain until the next different fault code entity is encountered, and the different fault code entity is assigned to the new subchain. Repeat the above process until all candidate references in the coreference chain are assigned to the corresponding subchains.

7. The method for constructing a fault knowledge base for commercial dishwashers according to claim 6, characterized in that, The process of repeating the above steps until all candidate references in the core reference chain are assigned to the corresponding sub-chains also includes: if no new different fault code entities are encountered after traversing to the end of the core reference chain, then all remaining candidate references are assigned to the last sub-chain separately, ensuring that each sub-chain contains at most one fault code entity.

8. The method for constructing a fault knowledge base for commercial dishwashers according to claim 1, characterized in that, The method of aggregating fault-related information based on the segmented sub-chains constructs a knowledge triple containing fault phenomena, causes, and solutions, including: For each segmented subchain, determine the unique fault code entity contained in the subchain; In the original sentence set corresponding to the candidate referents in the sub-chain, the expression related to "phenomenon is" and "displayed as" is extracted as the fault phenomenon, the expression related to "reason" and "due to" is extracted as the fault cause, and the expression related to "solution is" and "needs to be replaced" is extracted as the solution. The extracted fault phenomena, causes, and solutions are combined into knowledge triples and associated with the unique fault code entity corresponding to the sub-chain.

9. The method for constructing a fault knowledge base for commercial dishwashers according to claim 2, characterized in that, The structure of the bidirectional long short-term memory network-conditional random field model, from bottom to top, includes an embedding layer, a bidirectional long short-term memory network layer, and a conditional random field layer.

10. A system for building a fault knowledge base for commercial dishwasher equipment, characterized in that, It includes a processor and a memory; the memory stores computer program instructions, which, when executed by the processor, implement the method for constructing a fault knowledge base for commercial dishwashers as described in any one of claims 1-9.

Images