Question and answer method and system for a work machine, work machine

Abstract

The application relates to the technical field of working machines, and provides a question and answer method and system for a working machine and the working machine, wherein the method comprises the following steps: performing entity recognition on a fault problem of the working machine inputted by using a pre-trained entity recognition model; by using a pre-trained relation extraction model, the correlation of each two entities in the fault problem under a preset entity relation is determined for each preset entity relation, and if the correlation of the two entities meets a correlation threshold value, it is determined that the preset entity relation exists between the two entities; and based on the recognized entities, the determined preset entity relation and a pre-constructed knowledge graph, the answer of the fault problem is determined. The method can solve the defects of low fault diagnosis and maintenance efficiency of the working machine in the prior art, and can realize efficient diagnosis and maintenance of the fault problem of the working machine.

Description

Technical Field

[0001] This invention relates to the field of work machinery technology, and in particular to a question-and-answer method and system for work machinery, and work machinery. Background Technology

[0002] Currently, when excavators and other construction machinery malfunction, users typically report the problem through a platform, which then assigns a maintenance engineer to the site for diagnosis and repair. However, the diagnosis and repair process still relies on the engineer's personal experience. This dependence on the engineer's experience level, which is often insufficient, makes diagnosing and repairing unfamiliar faults more difficult, resulting in low efficiency in diagnosing and repairing construction machinery. Summary of the Invention

[0003] This invention provides a question-and-answer method and system for operating machinery, and the operating machinery itself, to solve the shortcomings of low efficiency in fault diagnosis and repair of operating machinery in the prior art, and to achieve high-efficiency diagnosis and repair of fault problems of operating machinery.

[0004] This invention provides a question-and-answer method for operating machinery, comprising:

[0005] Using a pre-trained entity recognition model, entity recognition is performed on the input fault problem of the operating machinery, and the entities are related to the fault of the operating machinery.

[0006] Using a pre-trained relation extraction model, for each preset entity relation, the correlation between every two entities in the fault problem is determined. If the correlation between two entities meets the correlation threshold, it is determined that there is a preset entity relation between the two entities.

[0007] Based on the identified entities, the established pre-defined entity relationships, and the pre-constructed knowledge graph, the answer to the fault problem is determined. The knowledge graph is formed by multiple entities and various pre-defined entity relationships.

[0008] According to the question-and-answer method for operating machinery provided by the present invention, for each preset entity relationship, the correlation between every two entities in the fault problem under the preset entity relationship is determined. If the correlation between two entities meets a correlation threshold, a preset entity relationship is determined to exist between the two entities, including:

[0009] Obtain the vector representation of the fault problem;

[0010] Based on the vector representation of the fault problem, a first vector representation and a second vector representation are obtained, where the entity in the first vector representation is the head entity and the entity in the second vector representation is the tail entity.

[0011] Based on the first vector representation and the second vector representation, a correlation matrix between the first vector representation and the second vector representation is obtained under each preset entity relationship. The correlation matrix contains the correlation between each head entity and each tail entity. If the correlation between the head entity and the tail entity meets the correlation threshold, it is determined that there is a preset entity relationship between the head entity and the tail entity.

[0012] According to a question-and-answer method for operating machinery provided by the present invention, the vector representation of the fault question includes sentence vector representation and word vector representation; based on the vector representation of the fault question, a first vector representation and a second vector representation are obtained, including:

[0013] Obtain the vector representation of the layer preceding the vector representation of the fault problem;

[0014] At least each word vector representation in the vector representation of the previous layer is added to the sentence vector representation of the fault problem to obtain the first vector representation;

[0015] The vector representation of the fault problem is used as the second vector representation.

[0016] According to the present invention, a question-answering method for operating machinery determines the answer to a fault question based on identified entities, predetermined entity relationships, and a pre-constructed knowledge graph, including:

[0017] Obtain subgraphs from the knowledge graph that correspond to the identified entities and the predetermined entity relationships;

[0018] Based on the subgraph, determine at least one candidate solution to the fault problem;

[0019] The probability of each candidate answer appearing is determined based on the pre-statistical count of each candidate answer's occurrences.

[0020] The answer to the faulty problem is the at least one candidate answer with the highest probability of occurrence.

[0021] According to the question-answering method for operating machinery provided by the present invention, the knowledge graph is constructed in the following manner:

[0022] Multiple initial knowledge graphs are obtained, each constructed based on a different data source;

[0023] Generate a vector representation of each node in each initial knowledge graph;

[0024] Calculate the similarity between the vector representations of nodes in different initial knowledge graphs;

[0025] Nodes whose similarity meets the similarity threshold are merged to obtain a knowledge graph.

[0026] According to the question-answering method for operating machinery provided by the present invention, the initial knowledge graph is constructed in the following manner:

[0027] Obtain data sources, which should include at least the maintenance records, maintenance manuals, failure cases, and failure experience of the operating machinery;

[0028] Entity recognition models are used to identify entities from data sources.

[0029] Using a relation extraction model, various preset entity relations are extracted from entities identified in the data source;

[0030] Based on the results of entity recognition and the extracted pre-defined entity relationships, an initial knowledge graph is constructed.

[0031] According to the question-answering method for operating machinery provided by the present invention, the relation extraction model is obtained based on a pre-trained model, and / or the entity recognition model is obtained based on a pre-trained model;

[0032] The pre-trained model is obtained through natural language pre-training based on the field of operational machinery.

[0033] According to a question-and-answer method for operating machinery provided by the present invention, before performing entity recognition on the input malfunction question of the operating machinery, the method further includes:

[0034] The system identifies the intent of input questions about the operating machinery. The intent includes malfunction questions, common sense questions, equipment operation instructions, and casual conversations.

[0035] If the input problem is determined to be a fault problem based on the result of intent recognition, the pre-trained entity recognition model is used to perform entity recognition on the input fault problem of the operating machinery.

[0036] If the result of intent recognition determines that the input question is not a fault question, the answer corresponding to the input question is selected from the preset question-answer pairs.

[0037] The present invention also provides a question-and-answer system for operating machinery, comprising:

[0038] The entity recognition unit is used to identify entities related to the faults of the operating machinery by using a pre-trained entity recognition model.

[0039] The relation extraction unit is used to use a pre-trained relation extraction model to determine the correlation between every two entities in the fault problem under each preset entity relation. If the correlation between two entities meets the correlation threshold, it is determined that there is a preset entity relation between the two entities.

[0040] The answer determination unit is used to determine the answer to the fault question based on the identified entities, the determined preset entity relationships, and the pre-constructed knowledge graph. The knowledge graph is formed by multiple entities and various preset entity relationships.

[0041] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement any of the above-described question-and-answer methods for working machinery.

[0042] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the above-described question-and-answer methods for working machinery.

[0043] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements any of the above-described question-and-answer methods for working machinery.

[0044] The present invention also provides a work machine for performing any of the question-and-answer methods for work machines described above, or including the question-and-answer system for work machines described above, or including the electronic device described above, or including the non-transitory computer-readable storage medium described above, or including the computer program product described above.

[0045] The question-answering method for operating machinery provided by this invention uses a pre-trained entity recognition model to identify entities in the input fault problem of the operating machinery. After identifying the entities present in the fault problem, a pre-trained relation extraction model is used to determine the correlation between every two entities in the fault problem under preset entity relations. If the correlation between two entities meets a correlation threshold, it is determined that there is a preset entity relationship between these two entities. In this way, determining the existence of preset entity relations by the correlation between entities can more accurately extract the preset entity relations contained in the fault problem. Based on the identified entities and the determined preset entity relations, combined with a pre-constructed knowledge graph, the answer to the fault problem can be determined. Since the knowledge graph can intuitively represent complex semantic knowledge, it can accurately find the answer to the fault problem, assisting in the diagnosis and repair of fault problems, thereby effectively improving the efficiency of fault diagnosis and repair of operating machinery. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0047] Figure 1 This is one of the flowcharts illustrating the question-and-answer method for operating machinery provided by the present invention;

[0048] Figure 2 This is the second flowchart of the question-and-answer method for operating machinery provided by the present invention;

[0049] Figure 3 This is a schematic diagram of the network structure of the relation extraction model provided by the present invention;

[0050] Figure 4 This is a schematic diagram of the network structure of the entity recognition model provided by the present invention;

[0051] Figure 5 This is the third flowchart of the question-and-answer method for operating machinery provided by the present invention;

[0052] Figure 6 This is one of the flowcharts for entity alignment provided by the present invention;

[0053] Figure 7 This is the fourth flowchart of the question-and-answer method for operating machinery provided by the present invention;

[0054] Figure 8 This is the fifth flowchart of the question-and-answer method for operating machinery provided by the present invention;

[0055] Figure 9 This is the sixth flowchart of the question-and-answer method for operating machinery provided by the present invention;

[0056] Figure 10 This is one of the structural schematic diagrams of the question-and-answer system for operating machinery provided by the present invention;

[0057] Figure 11 This is the second schematic diagram of the question-and-answer system for operating machinery provided by the present invention;

[0058] Figure 12 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0060] When a user's machinery malfunctions, the user typically reports the problem through the platform, which then assigns a maintenance engineer to the site for repair. However, during the repair process, the engineer usually relies on their personal experience to diagnose the problem, making the process heavily dependent on their individual skill level.

[0061] Currently, some fault diagnosis systems are based on the analysis of fault cases compiled by experts. They assist users in diagnosing faults by retrieving corresponding faults from a knowledge base through an interactive question-and-answer format. However, they do not make rational use of the large amount of fault record data generated by the operating machinery in the past. Furthermore, since the fault cases compiled by experts rely entirely on the experts' personal experience and the listed faults are limited, they are one-sided. In addition, the fault cases compiled by experts list all possible occurrences of the fault, and users need to check them in sequence, which is difficult and highly professional. They are generally only suitable for maintenance engineers to learn.

[0062] Furthermore, most existing auxiliary diagnostic systems can only meet the question-and-answer needs of fault-related questions, lacking service content. Moreover, when performing a full-match search in the knowledge base based on the user's question, they cannot find corresponding answers for different descriptions of the same fault.

[0063] Based on this, the present invention provides a question-answering method for operating machinery. This method can intuitively represent complex semantic knowledge using a knowledge graph, identify entities and extract pre-defined entity relationships for fault problems, and then determine the answer to the fault problem by combining the knowledge graph. This can effectively improve the efficiency of fault diagnosis and repair of operating machinery. The following is a combination of... Figures 1-9 The present invention describes a question-and-answer method for operating machinery.

[0064] Figure 1 This is one of the flowcharts illustrating the question-and-answer method for operating machinery provided by the present invention.

[0065] like Figure 1 As shown, the question-and-answer method for operating machinery provided in this embodiment can be executed by a question-and-answer device for operating machinery, such as a mobile phone, tablet computer, or vehicle-mounted terminal of the operating machinery, etc. The method includes at least the following steps:

[0066] Step 101: Using a pre-trained entity recognition model, perform entity recognition on the input fault problem of the operating machinery. The entities are related to the fault of the operating machinery.

[0067] Step 102: Using a pre-trained relation extraction model, for each preset entity relation, determine the correlation between every two entities in the fault problem under the preset entity relation. If the correlation between two entities meets the correlation threshold, it is determined that there is a preset entity relation between the two entities.

[0068] Step 103: Based on the identified entities, the determined preset entity relationships, and the pre-constructed knowledge graph, determine the answer to the fault problem. The knowledge graph is formed by multiple entities and various preset entity relationships.

[0069] The operating machinery in this embodiment can be excavators, pile drivers, mixers, etc.

[0070] Knowledge graphs are a technology used to represent various entity concepts and their relationships in the real world. In this embodiment, entities may include faulty parts, phenomena, causes of faults, and handling methods in operating machinery. Correspondingly, preset entity relationships refer to a certain relationship between two entities, such as "phenomenon is," "cause is," and "method is," etc. Specifically, a pre-trained relation extraction model can be used to extract the preset entity relationships contained in the fault problem, determine the correlation between each pair of entities identified in the fault problem, and judge the relationship between the correlation between two entities and a correlation threshold. If the correlation between two entities is greater than the correlation threshold, then the correlation between the two entities meets the correlation threshold. If the correlation between two entities meets the correlation threshold, then it is determined that there is a pre-existing relationship between the two entities. Let's define entity relationships. For example, in a defined fault problem, entities A and B are included. First, a pre-defined entity relationship of "phenomenon" is determined between entities A and B. If the correlation between the two entities meets a correlation threshold, then a pre-defined entity relationship of "phenomenon" is determined between entities A and B, meaning the phenomenon of entity A is entity B. Otherwise, a pre-defined entity relationship of "cause" is determined between entities A and B. If the correlation between the two entities meets a correlation threshold, then a pre-defined entity relationship of "cause" is determined between entities A and B, meaning the cause of entity A is entity B. Otherwise, a pre-defined entity relationship of "method" is determined between entities A and B. The order of entities A and B in the above steps cannot be reversed; that is, it is possible that there is no pre-defined entity relationship of "cause" between entities A and B, but there is a pre-defined entity relationship of "cause" between entities B and A.

[0071] In practical applications, after receiving user input about the malfunction of the operating machinery, it is necessary to extract information and perform semantic understanding on the malfunction. Specifically, semantic understanding can be performed through Natural Language Understanding (NLU) in Natural Language Processing (NLP).

[0072] After establishing the pre-defined entity relationships between entities, triples of (entity, relationship, entity) can be formed. Determining these triples defines the semantics of the fault problem. As key information, the answer to the fault problem can then be determined based on a pre-built knowledge graph. This answer assists in the diagnosis and repair of the fault problem. The knowledge graph, composed of multiple entities and various pre-defined entity relationships, is used for unified, standardized, and normalized management of all fault problems.

[0073] In the question-answering method for operating machinery provided in this embodiment, a pre-trained entity recognition model is used to identify entities in the input fault problem of the operating machinery. After identifying the entities present in the fault problem, a pre-trained relation extraction model is used to determine the correlation between every two entities in the fault problem under preset entity relations. If the correlation between two entities meets the correlation threshold, it is determined that there is a preset entity relationship between the two entities. In this way, determining the existence of preset entity relations by the correlation between entities can more accurately extract the preset entity relations contained in the fault problem. Based on the identified entities and the determined preset entity relations, combined with a pre-constructed knowledge graph, the answer to the fault problem can be determined. Since the knowledge graph can intuitively represent complex semantic knowledge, the answer to the fault problem can be found accurately, assisting in the diagnosis and repair of fault problems, thereby effectively improving the efficiency of fault diagnosis and repair of operating machinery.

[0074] In addition, for simple and easy-to-repair faults that do not require replacement parts, users of the operating machinery can also perform self-diagnosis and repair by answering the fault questions, without having to wait for maintenance engineers, thereby improving the efficiency of fault diagnosis and repair.

[0075] In an exemplary embodiment, before performing entity recognition on the input fault problem of the operating machinery, it is also necessary to perform word segmentation and part-of-speech tagging on the fault problem. In practice, the tagged part of speech can be verbs or nouns, which can help with entity recognition of the fault problem and facilitate accurate entity recognition.

[0076] In an exemplary embodiment, after identifying the entities in the fault problem and determining the preset entity relationships, the entire question-and-answer process can be processed through slot filling, session strategies, and response mechanisms. At the same time, the process of asking and answering questions with the user can be stored. When asking and answering questions with the same user, the user's historical question-and-answer records can be called. By combining the fault problem input by the user, the answer to the fault problem can be obtained, thereby achieving a comprehensive judgment. For example, Redis can be used to store and call the user's historical question-and-answer records.

[0077] In an exemplary embodiment, the answer to the fault problem is determined based on the identified entities, the determined preset entity relationships, and the pre-constructed knowledge graph, including:

[0078] Obtain subgraphs from the knowledge graph that correspond to the identified entities and the predetermined entity relationships;

[0079] Based on the subgraph, determine at least one candidate solution to the fault problem;

[0080] The probability of each candidate answer appearing is determined based on the pre-statistical count of each candidate answer's occurrences.

[0081] The answer to the faulty problem is the at least one candidate answer with the highest probability of occurrence.

[0082] The probability of each candidate answer appearing is the ratio of the number of times that candidate answer appears to the total number of times all candidate answers appear. The number of times a candidate answer appears reflects its frequency of occurrence.

[0083] In implementation, the answer to the fault question can be determined in the knowledge graph using Bayesian probability. Specifically, after determining the triples between entities and relations, the system identifies identical entities in the knowledge graph based on the entities in the fault question and the predetermined entity relations. These identical entities and related entities are then separated to form subgraphs. At least one candidate answer to the fault question can be determined within these subgraphs. In practice, there can be multiple candidate answers. Next, the probability of each candidate answer (i.e., the prior probability) can be determined based on the pre-statistical occurrence frequency of each candidate answer. This allows the system to determine the answer to the fault question and present it to the user. In the example of a fault problem, the entities identified are the engine and the abnormal noise. The relationship between the two is further determined as the phenomenon, i.e., the determined triple is (engine, phenomenon, abnormal noise). From the subgraph, it can be determined that there are two causes for the abnormal engine noise. Cause A occurs 8 times and cause B occurs 2 times. Based on this, it can be determined that the probability of cause A being the cause of the abnormal engine noise is 80%, and the probability of cause B being the cause is 20%. The answer to the fault problem can be determined as cause A. In practice, multiple answers to the fault problem can also be determined, such as determining cause A and cause B together as the answer to the fault problem.

[0084] In this embodiment, the answer to the fault problem is determined by using Bayesian probability. Since Bayesian probability is calculated from the frequency observed in a given distribution, the answer to the fault problem can be determined more accurately.

[0085] In an exemplary embodiment, for each preset entity relationship, the correlation between every two entities in the fault problem under the preset entity relationship is determined. If the correlation between two entities meets a correlation threshold, it is determined that a preset entity relationship exists between the two entities, including:

[0086] Obtain the vector representation of the fault problem;

[0087] Based on the vector representation of the fault problem, a first vector representation and a second vector representation are obtained, where the entity in the first vector representation is the head entity and the entity in the second vector representation is the tail entity.

[0088] Based on the first vector representation and the second vector representation, a correlation matrix between the first vector representation and the second vector representation is obtained under each preset entity relationship. The correlation matrix contains the correlation between each head entity and each tail entity. If the correlation between the head entity and the tail entity meets the correlation threshold, it is determined that there is a preset entity relationship between the head entity and the tail entity.

[0089] Specifically, the vector representation of the fault problem can be generated through embedding encoding.

[0090] Figure 2 This is the second flowchart illustrating the question-and-answer method for operating machinery provided by the present invention.

[0091] like Figure 2 As shown, in an exemplary embodiment, the vector representation of the fault problem includes a sentence vector representation and a word vector representation; based on the vector representation of the fault problem, obtaining a first vector representation and a second vector representation includes:

[0092] Step 201: Obtain the vector representation of the layer preceding the vector representation of the fault problem;

[0093] Step 202: Add the sentence vector representation of the fault problem to each word vector representation in the vector representation of the previous layer to obtain the first vector representation;

[0094] Step 203: Use the vector representation of the fault problem as the second vector representation.

[0095] In the vector representation of the fault problem, the first node is the CLS node, which is the sentence vector representation, used to store the complete sentence information of the fault problem. The remaining nodes are word vector representations. Since the sentence vector representation in the final vector representation of the fault problem can more accurately express the semantics of the entire sentence, adding the sentence vector representation of the fault problem vector representation to the vector representation of the previous layer results in a first vector representation that can also more accurately express the semantics of each word, thus accurately representing the entity. Each word vector representation in the previous layer can be added to the sentence vector representation of the fault problem; furthermore, the sentence vector representation in the previous layer can also be added to the sentence vector representation of the fault problem.

[0096] The following section will provide a detailed introduction using the network structure of the relation extraction model.

[0097] like Figure 3 As shown, the relation extraction model may include a vector representation layer, an interaction layer (indicated by "+" in the figure), a correlation calculation layer (indicated by "F" in the figure), and an activation layer (Sigmoid).

[0098] The vector representation layer is used to obtain a vector representation of the fault problem, which includes sentence vector representation and word vector representation. The vector representation layer can be an embedding layer or a pre-trained model.

[0099] The interaction layer is used to add the sentence vector representation of the fault problem to each word vector representation in the vector representation of the previous layer to obtain the first vector representation, in which the entity is the head entity. In practice, each word vector representation of the head entity can be added to the sentence vector representation of the tail entity. In implementation, the sentence vector representation of the head entity can also be added to the sentence vector representation of the tail entity, that is, the relationship between entities is judged by combining the information of the context.

[0100] The vector representation of the fault problem is used as the second vector representation, and the entities in it are called tail entities.

[0101] The correlation calculation layer is used to calculate the correlation matrix based on the first vector representation and the second vector representation. For example, the first vector representation and the second vector representation can be cross-multiplied to obtain the correlation matrix under different preset entity relationships. It should be noted that different preset network parameters are used under different preset entity relationships, thus resulting in different correlation matrices.

[0102] The activation layer is used to normalize the correlation matrix, which can control the result of the cross product to be within the range of 0-1.

[0103] For example, such as Figure 3As shown, the header entity includes entity 1-B, entity 2-B, entity 2-I, 0, and entity 3-B. Entity 1-B represents entity 1, entity 2-B and entity 2-I represent entity 2, and entity 3-B represents entity 3. The difference between entity 2-B and entity 2-I is that entity 2-B is the first character in the entity, and entity 2-I is the second character in the entity. For example, if entity 2 can be "engine", then entity 2-B represents "engine" and entity 2-I represents "engine".

[0104] In each correlation matrix, the correlation between two entities can be determined based on the scores (scores representing the magnitude of the correlation) corresponding to the positions of the head and tail entities in the matrix. In the correlation matrix, the positions of entities are represented by the coordinates of the matrix's rows and columns. For example, if coordinates (1,2) and (1,3) are determined to be the intersection of entity 1 and entity 2, the correlation between entity 1 and entity 2 can be determined based on the scores corresponding to coordinates (1,2) and (1,3). If the score corresponding to coordinate (1,2) is 0.5 and the score corresponding to coordinate (1,3) is 0.3, then the correlation between entity 1 and entity 2 is 0.8 (0.5 + 0.3 = 0.8). The question then becomes whether 0.8 is a positive correlation. The relevance threshold can also be reached by determining the relevance between entity 1 and entity 2 based on the average score corresponding to coordinates (1,2) and (1,3). For example, if the score corresponding to coordinate (1,2) is 0.5 and the score corresponding to coordinate (1,3) is 0.3, then the relevance between entity 1 and entity 2 is 0.4 ((0.5+0.3) / 2=0.4). It is then determined whether 0.4 meets the relevance threshold. If it does, then a pre-defined entity relationship corresponding to this relevance matrix exists between entity 1 and entity 2. The existence of the aforementioned pre-defined entity relationships between entity 1 and entity 3, entity 2 and entity 1, entity 2 and entity 3, entity 3 and entity 1, and entity 3 and entity 2 can also be determined using the same example.

[0105] In this embodiment, a correlation matrix between entities is constructed to determine the preset entity relationship between entities in the fault problem. The correlation matrix can determine the magnitude of the correlation between two entities. If a certain correlation threshold is met, it is determined that there is a preset entity relationship between the two entities. In this way, by systematically and quantitatively analyzing the relationship between entities, the preset entity relationship between entities can be determined more accurately, thereby completing the relationship extraction.

[0106] The following section will provide a detailed explanation of the entity recognition process, using the network structure of the entity recognition model as an example.

[0107] In an exemplary embodiment, entity identification is performed on the input malfunction problem of the operating machinery, including:

[0108] Obtain the vector representation of the fault problem;

[0109] Entity recognition is based on the vector representation of fault problems and the context of fault problems.

[0110] like Figure 4 As shown, the entity recognition model provided in this embodiment may include an encoding layer, a recognition layer, and a decoding layer.

[0111] The encoding layer is used to encode the fault problem, generating a vector representation of the fault problem. Specifically, after embedding the fault problem using the encoding layer of the entity recognition model, a vector representation of the fault problem is generated. In practice, embedding encoding can be performed on each character in the fault problem to obtain a vector representation of each character. Therefore, the encoding layer can be an embedding layer or the pre-trained model mentioned above.

[0112] The recognition layer is used to perform entity recognition based on the vector representation of the fault problem and the context of the fault problem, and to obtain the label output representation of each word of the entity.

[0113] The decoding layer is used to decode the tag output representation of each character of the entity to obtain the output tag of each character in order to obtain the recognized entity. The output tags include a first output tag and a second output tag. The first output tag corresponds to the first character of the entity, and the second output tag corresponds to the characters after the first character of the entity.

[0114] Since the same entity can have different meanings in different contexts, entity recognition for fault problems requires consideration of the context. For example, the entity "solenoid valve" alone cannot determine which solenoid valve is located, but considering the context can aid in semantic analysis, thus helping to identify the solenoid valve's location. Based on this, the aforementioned recognition layer can be a bidirectional LSTM model, thereby leveraging contextual information to improve the accuracy of entity recognition.

[0115] In practical applications, after obtaining the character vector representation of each character in the fault problem, the label output representation of each character can be obtained through the recognition layer in the entity recognition model. Then, the decoding layer in the entity recognition model, such as conditional random fields (CRF), can be used to decode the label output representation of the character. In this way, the output label of each character can be obtained. After further decoding, the entity in the fault problem can be determined, which means that the entity recognition of the fault problem is completed.

[0116] See Figure 4For example, for the entity recognition result of a fault problem, it is B-fault part, I-fault part, I-fault part, B-phenomenon, and I-phenomenon. The fault part here refers to the engine, and the phenomenon refers to abnormal noise. There will be a prefix B or prefix I before the fault part and the phenomenon. The prefix B represents the beginning of the entity, and the prefix I represents the characters after the beginning of the entity.

[0117] In implementation, after obtaining the label output representation of each word through the recognition layer, the entities in the fault problem can also be determined by the enumeration method. Exemplarily, for the fault problem of "engine abnormal noise", "fa" can be enumerated first. By comparing it with the preset entities, it can be determined whether "fa" is an entity, then whether "fadong" is an entity, and then whether "engine" is an entity. In this way, after enumerating all the words in the fault problem, it can be determined that the entities existing in the fault problem are "engine" and "abnormal noise".

[0118] In this embodiment, by vectorizing the fault problem and combining the context, the entities in the fault problem can be determined more accurately, and the ambiguity can be effectively eliminated, improving the efficiency of diagnosing and repairing the fault problems of construction machinery.

[0119] Figure 5 It is the third flow schematic diagram of the question-answering method for construction machinery provided by the present invention.

[0120] Figure 6 It is one of the flow schematic diagrams for entity alignment provided by the present invention.

[0121] Such as Figure 5 and Figure 6 As shown, in an exemplary embodiment, after entity recognition of the input fault problem of construction machinery, it further includes:

[0122] Step 501, mapping the vector space of each recognized entity to a standard Gaussian space;

[0123] Step 502, calculating the similarity between the entity and the standard entity in the standard Gaussian space;

[0124] Step 503, if the similarity reaches the first set threshold, standardize and align the entity with the standard entity.

[0125] For the same entity, different people may have different ways of describing it, that is, there may be colloquial expressions. For example, the entity identified by the fault problem input by user A is "motor", the entity identified by the fault problem input by user B is "engine", and the entity identified by the fault problem input by user C is "engine". It can be seen that although "motor", "engine" and "engine" are different expressions, they all refer to the entity "engine". Therefore, information extraction from fault problems also includes entity alignment. In this embodiment, entity alignment is used to standardize and align the colloquial entities in the fault problems input by different users with standard entities. The standard entities can be predefined, such as defining "engine", "motor" and "engine" as "engine", or they can be defined as "motor".

[0126] In this embodiment, entity alignment can determine whether two or more entities from different information sources point to the same object in the real world. If multiple entities are found to represent the same entity, such as "engine", "motor" and "engine", these entities all represent the same entity. Alignment relationships need to be established between these entities, and the information contained in the entities is fused and aggregated.

[0127] In this embodiment, a vector space mapping method can be used to map the vector space of each entity to a standard Gaussian space. The vector space of the entity and the standard Gaussian space are invertible. In practice, standard entities in the standard Gaussian space can be pre-defined.

[0128] In this embodiment, by standardizing and aligning entities, the true semantics of the fault problem can be better determined, avoiding the problem of entities being unrecognizable due to different description methods.

[0129] Figure 7 This is the fourth flowchart illustrating the question-and-answer method for operating machinery provided by this invention.

[0130] like Figure 7 As shown in the exemplary embodiment, the knowledge graph is constructed in the following manner:

[0131] Step 701: Obtain multiple initial knowledge graphs, each of which is constructed based on a different data source.

[0132] Step 702: Generate a vector representation of each node in each initial knowledge graph.

[0133] Step 703: Calculate the similarity between the vector representations of nodes in different initial knowledge graphs.

[0134] Step 704: Merge nodes whose similarity meets the similarity threshold to obtain the knowledge graph.

[0135] Since knowledge graphs are a technical method that uses graph models to describe the relationships between different entities, their essence is a semantic network describing the relationships between entities. Through knowledge graphs, complex semantic knowledge can be represented intuitively. Therefore, each entity in a knowledge graph can be regarded as a node, and the relationships between entities can be regarded as the lines connecting the nodes. Furthermore, by analyzing the nodes and the relationships between them, potential patterns of events can be inferred. During the construction of a knowledge graph, if different data sources are used, the resulting knowledge graphs may not be entirely identical. Therefore, deep learning can be used to construct multiple initial knowledge graphs using multiple different data sources, and then these initial knowledge graphs can be merged to form the final knowledge graph.

[0136] Multiple initial knowledge graphs can be merged by merging nodes within the initial knowledge graph. Specifically, the nodes in the initial knowledge graph can be embedded to obtain a vector representation of each node. This vector representation is used to calculate the similarity between nodes in different initial knowledge graphs. If the similarity between two nodes meets a similarity threshold, the two nodes are merged. For example, nodes with a similarity greater than or equal to the similarity threshold are merged, thus merging nodes with high similarity, which are considered to be the same node and can be merged.

[0137] In implementation, the constructed knowledge graph can be stored using a graph database, such as the Neo4j database. Specifically, the knowledge graph can employ graph network models such as GCN.

[0138] In this embodiment, initial knowledge graphs are constructed using different data sources, and then the different initial knowledge graphs are merged by merging nodes. This results in a more complete knowledge graph, which can provide more accurate answers when determining the solutions to fault problems based on the knowledge graph. This enables more efficient diagnosis and repair of fault problems in the operating machinery.

[0139] In the exemplary embodiment, the initial knowledge graph is constructed as follows:

[0140] Obtain data sources, which should include at least the maintenance records, maintenance manuals, failure cases, and failure experience of the operating machinery;

[0141] Entity recognition models are used to identify entities from data sources.

[0142] Using a relation extraction model, various preset entity relations are extracted from entities identified in the data source;

[0143] Based on the results of entity recognition and the extracted pre-defined entity relationships, an initial knowledge graph is constructed.

[0144] Customer relationship management systems typically contain a large number of unstructured equipment failure records. These unstructured equipment failure records can be structured and used to build a knowledge graph. In addition, maintenance manuals of the operating machinery, classic failure cases, and failure experience summarized by maintenance engineers can also be used as data sources to build a knowledge graph.

[0145] After obtaining the data source, entity recognition and relation extraction can be performed on the data from that data source to construct an initial knowledge graph based on that data source. The specific steps for entity recognition and relation extraction can be implemented as described in the above embodiments, and will not be repeated here.

[0146] In this embodiment, by performing entity recognition and relation extraction on the data source to construct an initial knowledge graph, the semantic information of the data source can be effectively determined, thereby making the constructed initial knowledge graph more accurate.

[0147] As mentioned earlier, relation extraction models and entity recognition models can both be derived from pre-trained models. Furthermore, the pre-trained models are obtained through natural language pre-training in the field of operational machinery.

[0148] In practical applications, pre-trained models can be trained using a large amount of data from the field of operational machinery. The trained pre-trained model is a natural language pre-trained model; for example, a BERT model (Bidirectional Encoder Representation from Transformers) can be trained using this model. The aforementioned entity recognition model can be constructed using the pre-trained model and trained using data sources. Similarly, the aforementioned relation extraction model can be constructed using the pre-trained model and trained using data sources.

[0149] In the prior art, pre-trained models are generally trained using large amounts of open-domain datasets from various fields. However, in this embodiment, the pre-trained model is trained using data from the field of construction machinery. In this way, the pre-trained model can learn more knowledge from the field of construction machinery, and thus perform better in the field of construction machinery.

[0150] Because the pre-trained models used in existing technologies have a very large number of parameters, the models run very slowly, causing great inconvenience to users in practical applications. Therefore, in the exemplary embodiment, model distillation can be used to train the pre-trained model. Specifically, a pre-trained model with a large number of parameters can be trained first, and then a pre-trained model with a smaller number of parameters can be redesigned. The pre-trained model with a smaller number of parameters can then be used to train the pre-trained model with a smaller number of parameters, so that the pre-trained model with a smaller number of parameters learns the prediction probability and knowledge distribution of the pre-trained model with a larger number of parameters. In this way, the function of using a pre-trained model with a smaller number of parameters to achieve the function of a pre-trained model with a larger number of parameters is achieved.

[0151] For example, a BERT pre-trained model can be trained using Transformer as the network structure as a pre-trained model with a large number of parameters, while a Long Short-Term Memory (LSTM) network pre-trained model can be trained using Long Short-Term Memory (LSTM) as the network structure as a pre-trained model with a small number of parameters.

[0152] In this embodiment, model distillation can reduce the number of parameters in the pre-trained model and greatly improve the speed of model operation, for example, it can improve the model inference speed by more than 10 times.

[0153] Figure 8 This is the fifth flowchart illustrating the question-and-answer method for operating machinery provided by the present invention.

[0154] like Figure 8 As shown in the exemplary embodiment, after training and obtaining the pre-trained model, and before training the entity recognition model and the relation extraction model, it is also necessary to annotate the acquired data source.

[0155] Specifically, the data annotation process includes:

[0156] Organize vocabulary, that is, organize the terminology related to some operating machinery.

[0157] Discussing rules involves compiling relevant terminology and then jointly discussing and formulating rules for data labeling.

[0158] The trial annotation process involves labeling a portion of the data based on the data annotation rules discussed above. The rules are then continuously improved based on the issues raised during the annotation process until no more issues are reported.

[0159] Labeling training refers to training labeling personnel on the rules.

[0160] Batch annotation involves distributing data to annotation personnel in batches, who then annotate the data.

[0161] Quality control and approval involves verifying the labeling results to ensure the quality of the data labeling. Cross-validation can be used for this purpose, which involves distributing the same data to different people for labeling. If the labeling results from different people are consistent, the data is considered to be correctly labeled. If the labeling results from different people are inconsistent, further quality control can be carried out.

[0162] Organizing and archiving involves integrating the verified and labeled data for use in training entity recognition models and relation extraction models.

[0163] Data annotation refers to labeling entities in the fault problem and the predefined relationships between them. Specifically, it can be labeling data from the data source with predefined labels.

[0164] In practical applications, entity recognition models and relation extraction models can be constructed using a combination of small-batch annotation and large-batch machine-automated extraction techniques. Small-batch annotation refers to manually annotating a portion of the data source. Specifically, technical terms can be manually compiled and data annotation rules can be discussed and determined. These rules are then continuously improved based on feedback from the small-batch annotation process. Once the data annotation rules are perfected, the entity recognition model and relation extraction model can be initially constructed. Further, machine-assisted annotation can be used, with manual annotation as the primary method, and the initially constructed entity recognition model and relation extraction model serving as a reference for manual annotation. In this way, the entity recognition model and relation extraction model can be further improved during the data annotation process.

[0165] In practice, large-scale automated machine extraction refers to the process of machine-assisted annotation in which all data in the data source can be annotated in batches. This allows for adjustments to the entity recognition model and relation extraction model based on problems encountered during the data annotation process.

[0166] During implementation, the data annotation results can be quality checked and cross-validated to ensure the quality of data annotation. Cross-validation refers to distributing the same data to different people for annotation. If the data annotation results of different people are consistent, the data annotation is considered correct. If the data annotation structures of different people are inconsistent, further quality checks can be carried out to finally determine the data annotation results.

[0167] In this embodiment, the entity recognition model and relation extraction model are finally trained by using a combination of small-batch annotation and large-batch automatic machine extraction. This breaks the limitation that all data in the traditional knowledge graph construction process is manually organized. It establishes and standardizes the knowledge graph and data annotation process. By combining manual annotation with machine-assisted annotation, the training speed of the entity recognition model and relation extraction model can be accelerated, and the accuracy of the entity recognition model and relation extraction model can be improved.

[0168] In an exemplary embodiment, before performing entity recognition on the input malfunction problem of the operating machinery, the method further includes:

[0169] The system identifies the intent of input questions about the operating machinery. The intent includes malfunction questions, common sense questions, equipment operation instructions, and casual conversations.

[0170] If the input problem is determined to be a fault problem based on the result of intent recognition, the pre-trained entity recognition model is used to perform entity recognition on the input fault problem of the operating machinery.

[0171] If the result of intent recognition determines that the input question is not a fault question, the answer corresponding to the input question is selected from the preset question-answer pairs.

[0172] In other words, the question-and-answer method for operating machinery provided in this embodiment can not only answer user-inputted fault questions, but also respond to general knowledge questions, equipment operation instructions, and casual conversations.

[0173] In practical applications, question-and-answer pairs for common sense questions, equipment operation instructions, and casual conversations can be pre-set and stored. After a user inputs a question, the intent of the user's input is first identified by an intent recognition model. Based on the different intents, the system automatically jumps to the corresponding specific mechanism for question-and-answer processing. If the user's input question is determined to be a fault problem, it is processed according to the method provided in the above embodiments, specifically including: using a pre-trained entity recognition model to identify entities related to the fault of the operating machinery; using a pre-trained relation extraction model to determine the correlation between every two entities in the fault problem under each preset entity relation; if the correlation between two entities meets the correlation threshold, a preset entity relation is determined to exist between the two entities; and based on the identified entities, the determined preset entity relations, and the pre-constructed knowledge graph, the answer to the fault problem is determined.

[0174] If it is determined that the input question is not a fault problem, then the answer corresponding to the input question is selected from pre-set question-and-answer pairs related to common sense questions, equipment operation instructions, and casual conversations. Each question-and-answer pair includes a standard question and its corresponding answer. The degree of matching (e.g., semantic similarity) between the input question and the standard question in the question-and-answer pair can be calculated. If the matching degree threshold is met, then the answer to the standard question is the answer to the input question.

[0175] In practice, after selecting the answers corresponding to the input question, Natural Language Generating (NLG) technology can be used to respond to the user. Specifically, the answers corresponding to the input question can be automatically modified in terms of sentence order, grammatical logic, etc., to generate natural language that humans can understand.

[0176] In the process of filtering answers to an input question based on question-answer pairs, recall ranking can be used to determine the answer. Specifically, recall ranking refers to quickly recalling answers corresponding to the user's input question from pre-generated question-answer pairs, and then ranking them by matching degree before pushing several answers to the user in sequence. In implementation, a multi-path recall plus ranking approach can be used to determine the answer to the input question. Multi-path recall refers to using different strategies to recall from multiple angles, which can make the filtered answers corresponding to the input question more accurate.

[0177] In practical applications, intent recognition models can also be trained based on pre-trained models.

[0178] In this embodiment, not only can it answer questions about malfunctions of the operating machinery, but it can also provide functions such as casual conversation, general knowledge Q&A, and equipment operation guidance. This can effectively improve service quality and solve the problem that general users do not systematically learn the operating manual of the operating machinery and have difficulty finding solutions when they encounter problems they do not know how to operate.

[0179] Figure 9 This is the sixth flowchart illustrating the question-and-answer method for operating machinery provided by this invention.

[0180] like Figure 9As shown, in a specific embodiment, after a user inputs a question, the answer to the question can be obtained after question preprocessing, natural language understanding, session management, and graph computation steps. Question preprocessing includes Chinese word segmentation, semantic disambiguation, part-of-speech tagging, text correction, syntactic analysis, and sensitive word filtering. Natural language understanding includes intent recognition, entity recognition, relation extraction, and entity alignment. Session management includes slot filling, state tracking, session strategy, state storage, response mechanism, and emergency mechanism. Graph computation includes information retrieval, subgraph generation, probability calculation, natural language generation, reordering, and result return.

[0181] The question-and-answer system for operating machinery provided by the present invention will be described below. The question-and-answer system for operating machinery described below can be referred to in correspondence with the question-and-answer method for operating machinery described above.

[0182] Figure 10 This is one of the structural schematic diagrams of the question-and-answer system for operating machinery provided by the present invention.

[0183] Figure 11 This is the second schematic diagram of the question-and-answer system for operating machinery provided by the present invention.

[0184] like Figure 10 As shown, the question-and-answer system for operating machinery provided in this embodiment includes:

[0185] The entity recognition unit 1001 is used to perform entity recognition on the input fault problem of the operating machinery using a pre-trained entity recognition model, wherein the entity is related to the fault of the operating machinery.

[0186] The relation extraction unit 1002 is used to use a pre-trained relation extraction model to determine the correlation between every two entities in the fault problem under the preset entity relationship for each preset entity relationship. If the correlation between two entities meets the correlation threshold, it is determined that there is a preset entity relationship between the two entities.

[0187] The answer determination unit 1003 is used to determine the answer to the fault problem based on the identified entities, the determined preset entity relationships and the pre-constructed knowledge graph. The knowledge graph is formed by multiple entities and multiple preset entity relationships.

[0188] like Figure 11 As shown, the question-and-answer system for operating machinery provided in this embodiment includes a data layer, a technology layer, an access layer, and an application layer. The application layer is set on a terminal device, such as a mobile phone or a computer. The access layer, technology layer, and data layer are set on the server side. The access layer enables users to interact between the terminal and the server.

[0189] Sensitive word filtering in the technical layer refers to filtering out predefined sensitive words through rule matching, such as sensitive words related to pornography or violence.

[0190] Subgraph generation refers to identifying identical and related entities in a knowledge graph based on the entities in the user's input question, and then separating them from the knowledge graph to form a subgraph.

[0191] Clause splitting refers to breaking down a sentence into multiple clauses.

[0192] Status tracking refers to tracking a user's historical question and answer records.

[0193] Slot filling refers to filling the extracted two entities into a triplet. For example, both the faulty component and the phenomenon are critical slots, with the phenomenon being more critical. If the extracted entities only contain the faulty component, then the user needs to be asked about the phenomenon. If the extracted entities only contain the phenomenon, then the user does not need to be asked about the faulty component.

[0194] In the data layer, the artificial middleware platform refers to the central server.

[0195] Artificial inspiration refers to rules being formulated by humans. For example, in the process of training a relation extraction model, the labelers annotate the data, which is artificial inspiration.

[0196] In an exemplary embodiment, the relation extraction unit 1002 is specifically used for:

[0197] Obtain the vector representation of the fault problem;

[0198] Based on the vector representation of the fault problem, a first vector representation and a second vector representation are obtained, where the entity in the first vector representation is the head entity and the entity in the second vector representation is the tail entity.

[0199] Based on the first vector representation and the second vector representation, a correlation matrix between the first vector representation and the second vector representation is obtained under each preset entity relationship. The correlation matrix contains the correlation between each head entity and each tail entity. If the correlation between the head entity and the tail entity meets the correlation threshold, it is determined that there is a preset entity relationship between the head entity and the tail entity.

[0200] In an exemplary embodiment, the relation extraction unit 1002 is specifically used for:

[0201] The vector representation of the previous layer of the fault problem is obtained;

[0202] At least each word vector representation in the vector representation of the previous layer is added to the sentence vector representation of the fault problem to obtain the first vector representation;

[0203] The vector representation of the fault problem is used as the second vector representation.

[0204] In an exemplary embodiment, the answer determination unit 1003 is specifically used for:

[0205] Obtain subgraphs from the knowledge graph that correspond to the identified entities and the predetermined entity relationships;

[0206] Based on the subgraph, determine at least one candidate solution to the fault problem;

[0207] The probability of each candidate answer appearing is determined based on the pre-statistical count of each candidate answer's occurrences.

[0208] The answer to the faulty problem is the at least one candidate answer with the highest probability of occurrence.

[0209] In an exemplary embodiment, a solid alignment unit is also included, which is specifically used for:

[0210] Map the vector space of each identified entity to a standard Gaussian space;

[0211] Calculate the similarity between an entity and a standard entity in standard Gaussian space;

[0212] If the similarity reaches the first set threshold, the entity will be standardized and aligned with the standard entity.

[0213] In an exemplary embodiment, the knowledge graph is constructed as follows:

[0214] Multiple initial knowledge graphs are obtained, each constructed based on a different data source;

[0215] Generate a vector representation of each node in each initial knowledge graph;

[0216] Calculate the similarity between the vector representations of nodes in different initial knowledge graphs;

[0217] Nodes whose similarity meets the similarity threshold are merged to obtain a knowledge graph.

[0218] In an exemplary embodiment, the initial knowledge graph is constructed as follows:

[0219] Obtain data sources, which should include at least the maintenance records, maintenance manuals, failure cases, and failure experience of the operating machinery;

[0220] Entity recognition models are used to identify entities from data sources.

[0221] Using a relation extraction model, various preset entity relations are extracted from entities identified in the data source;

[0222] Based on the results of entity recognition and the extracted pre-defined entity relationships, an initial knowledge graph is constructed.

[0223] In an exemplary embodiment, the relation extraction model is obtained based on a pre-trained model, and / or the entity recognition model is obtained based on a pre-trained model;

[0224] The pre-trained model is obtained through natural language pre-training based on the field of operational machinery.

[0225] In an exemplary embodiment, an intent recognition unit is further included, which is specifically used for:

[0226] The system identifies the intent of input questions about the operating machinery. The intent includes malfunction questions, common sense questions, equipment operation instructions, and casual conversations.

[0227] If the input problem is determined to be a fault problem based on the result of intent recognition, the pre-trained entity recognition model is used to perform entity recognition on the input fault problem of the operating machinery.

[0228] If the result of intent recognition determines that the input question is not a fault question, the answer corresponding to the input question is selected from the preset question-answer pairs.

[0229] Figure 12 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 12 As shown, the electronic device may include a processor 1210, a communications interface 1220, a memory 1230, and a communication bus 1240. The processor 1210, communications interface 1220, and memory 1230 communicate with each other via the communication bus 1240. The processor 1210 can call logical instructions from the memory 1230 to execute a question-and-answer method for the operating machinery. This method includes: using a pre-trained entity recognition model to perform entity recognition on the input fault question of the operating machinery, where the entities are related to the fault of the operating machinery.

[0230] Using a pre-trained relation extraction model, for each preset entity relation, the correlation between every two entities in the fault problem is determined. If the correlation between two entities meets the correlation threshold, it is determined that there is a preset entity relation between the two entities.

[0231] Based on the identified entities, the established pre-defined entity relationships, and the pre-constructed knowledge graph, the answer to the fault problem is determined. The knowledge graph is formed by multiple entities and various pre-defined entity relationships.

[0232] Furthermore, the logical instructions in the aforementioned memory 1230 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0233] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the question-and-answer method for operating machinery provided by the above methods. The method includes: using a pre-trained entity recognition model to perform entity recognition on the input fault question of the operating machinery, wherein the entity is related to the fault of the operating machinery.

[0234] Using a pre-trained relation extraction model, for each preset entity relation, the correlation between every two entities in the fault problem is determined. If the correlation between two entities meets the correlation threshold, it is determined that there is a preset entity relation between the two entities.

[0235] Based on the identified entities, the established pre-defined entity relationships, and the pre-constructed knowledge graph, the answer to the fault problem is determined. The knowledge graph is formed by multiple entities and various pre-defined entity relationships.

[0236] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the question-answering method for operating machinery provided by the above methods, the method comprising: using a pre-trained entity recognition model to perform entity recognition on an input question about a fault of the operating machinery, wherein the entity is related to the fault of the operating machinery;

[0237] Using a pre-trained relation extraction model, for each preset entity relation, the correlation between every two entities in the fault problem is determined. If the correlation between two entities meets the correlation threshold, it is determined that there is a preset entity relation between the two entities.

[0238] Based on the identified entities, the established pre-defined entity relationships, and the pre-constructed knowledge graph, the answer to the fault problem is determined. The knowledge graph is formed by multiple entities and various pre-defined entity relationships.

[0239] The present invention also provides a work machine for performing the question-and-answer method for work machines as provided in the above embodiments, or including the question-and-answer system for work machines provided in the embodiments, or including the electronic device provided in the embodiments, or including the non-transitory computer-readable storage medium provided in the embodiments, or including the computer program product provided in the embodiments.

[0240] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0241] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments or some parts of embodiments.

[0242] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A question-and-answer method for operating machinery, characterized in that, include: Using a pre-trained entity recognition model, entity recognition is performed on the input fault problem of the operating machinery, where the entity is related to the fault of the operating machinery; The entity recognition model includes an encoding layer, a recognition layer, and a decoding layer. The encoding layer encodes the fault problem to generate a vector representation of the fault problem. The recognition layer performs entity recognition based on the vector representation of the fault problem and its context to obtain a label output representation for each character of the entity. The decoding layer decodes the label output representation for each character of the entity to obtain an output label for each character, thus obtaining the recognized entity. The output labels include a first output label and a second output label. The first output label corresponds to the first character of the entity, and the second output label corresponds to the characters after the first character of the entity. Using a pre-trained relation extraction model, for each preset entity relation, the correlation between every two entities in the fault problem is determined. If the correlation between two entities meets a correlation threshold, it is determined that the preset entity relation exists between the two entities. Based on the identified entities, the determined preset entity relationships, and the pre-constructed knowledge graph, the answer to the fault problem is determined. The knowledge graph is formed by multiple entities and multiple preset entity relationships.

2. The question-and-answer method for operating machinery according to claim 1, characterized in that, For each preset entity relationship, determining the correlation between every two entities in the fault problem under the preset entity relationship, and determining that the preset entity relationship exists between the two entities if the correlation between the two entities meets a correlation threshold, includes: Obtain the vector representation of the fault problem; Based on the vector representation of the fault problem, a first vector representation and a second vector representation are obtained, wherein the entity in the first vector representation is the head entity and the entity in the second vector representation is the tail entity; Based on the first vector representation and the second vector representation, a correlation matrix between the first vector representation and the second vector representation is obtained under each preset entity relationship. The correlation matrix includes the correlation between each head entity and each tail entity. If the correlation between the head entity and the tail entity satisfies the correlation threshold, it is determined that the preset entity relationship exists between the head entity and the tail entity.

3. The question-and-answer method for operating machinery according to claim 2, characterized in that, The vector representation of the fault problem includes sentence vector representation and character vector representation; obtaining the first vector representation and the second vector representation based on the vector representation of the fault problem includes: Obtain the vector representation of the layer preceding the vector representation of the fault problem; The first vector representation is obtained by adding the sentence vector representation of the fault problem to each word vector representation in the vector representation of the previous layer; The vector representation of the fault problem is used as the second vector representation.

4. The question-and-answer method for operating machinery according to any one of claims 1 to 3, characterized in that, The step of determining the answer to the fault problem based on the identified entities, the determined preset entity relationships, and the pre-constructed knowledge graph includes: Obtain subgraphs from the knowledge graph that correspond to the identified entities and the determined preset entity relationships; Based on the subgraph, at least one candidate answer to the fault problem is determined; The probability of each candidate answer appearing is determined based on the pre-statistical occurrence count of each candidate answer. The candidate answer with the highest probability of occurrence is selected as the answer to the faulty problem.

5. The question-and-answer method for operating machinery according to any one of claims 1 to 3, characterized in that, The knowledge graph is constructed in the following way: Multiple initial knowledge graphs are obtained, each of which is constructed based on a different data source; Generate a vector representation for each node of each of the initial knowledge graphs; Calculate the similarity between the vector representations of the nodes in different initial knowledge graphs; The nodes whose similarity meets the similarity threshold are merged to obtain the knowledge graph.

6. The question-and-answer method for operating machinery according to claim 5, characterized in that, The initial knowledge graph was constructed in the following manner: The data source is obtained, and the data source includes at least the maintenance records, maintenance manuals, fault cases and fault experience of the operating machinery; The entity recognition model is used to perform entity recognition on the data source; Using the relationship extraction model, various preset entity relationships are extracted from the entities identified in the data source; Based on the results of entity recognition and the extracted preset entity relationships, the initial knowledge graph is constructed.

7. The question-and-answer method for operating machinery according to claim 6, characterized in that, The relation extraction model is obtained based on a pre-trained model, and / or the entity recognition model is obtained based on a pre-trained model; The pre-trained model is obtained through natural language pre-training based on the field of the operating machinery.

8. The question-and-answer method for operating machinery according to any one of claims 1 to 3, characterized in that, Before performing entity identification on the input malfunctions of the operating machinery, the method further includes: The intent of the input questions about the operating machinery is identified, and the intent includes fault questions, common sense questions and answers, equipment operation instructions and casual conversations; If the input problem is determined to be the fault problem based on the result of the intent recognition, the input fault problem of the operating machinery is identified by using a pre-trained entity recognition model. If, based on the result of the intent recognition, it is determined that the input question is not the fault question, the answer corresponding to the input question is selected from the preset question-answer pairs.

9. A question-and-answer system for operating machinery, characterized in that, include: An entity recognition unit is used to perform entity recognition on an input fault problem of a working machine using a pre-trained entity recognition model. The entity is related to the fault of the working machine. The entity recognition model includes an encoding layer, a recognition layer, and a decoding layer. The encoding layer encodes the fault problem to generate a vector representation of the fault problem. The recognition layer performs entity recognition based on the vector representation of the fault problem and its context to obtain a label output representation for each character of the entity. The decoding layer decodes the label output representation for each character of the entity to obtain an output label for each character, thus obtaining the recognized entity. The output labels include a first output label and a second output label. The first output label corresponds to the first character of the entity, and the second output label corresponds to the characters after the first character of the entity. The relation extraction unit is used to use a pre-trained relation extraction model to determine the correlation between every two entities in the fault problem under each preset entity relation. If the correlation between two entities meets the correlation threshold, it is determined that the preset entity relation exists between the two entities. The answer determination unit is used to determine the answer to the fault question based on the identified entities, the determined preset entity relationships, and the pre-constructed knowledge graph, wherein the knowledge graph is formed by multiple entities and multiple preset entity relationships.

10. A type of operating machinery, characterized in that, The operating machinery is used to perform the question-and-answer method for operating machinery as described in any one of claims 1 to 8, or includes the question-and-answer system for operating machinery as described in claim 9.

Images