Fault operation and maintenance large model training and using method, device and medium

Abstract

The invention relates to a fault operation and maintenance large model training and using method, equipment and a medium. The method comprises the following steps: based on a pre-trained fault operation and maintenance large model, generating a path capture loss of a reference reasoning path related to an operation and maintenance problem sample from a fault operation and maintenance knowledge graph; wherein the entity of the fault operation and maintenance knowledge graph comprises a problem description, a problem reason and an operation and maintenance strategy of an operation and maintenance problem; the edge relationship between the entities comprises a causal relationship; according to the operation and maintenance problem sample, the reference reasoning path and the fault operation and maintenance knowledge graph, determining a pre-trained fault operation and maintenance large model to generate retrieval reasoning loss of an operation and maintenance strategy label corresponding to the operation and maintenance problem sample; and adjusting the pre-trained fault operation and maintenance large model according to the path capture loss and / or the retrieval reasoning loss. By adopting the method, the accuracy and efficiency of determining the operation and maintenance strategy by the fault operation and maintenance large model can be improved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a method, device and medium for training and using a large fault operation and maintenance model. Background Art

[0002] As a revolutionary force in the automotive industry, new energy vehicles have a core technology that is more complex and sophisticated than traditional internal combustion engine vehicles, namely the "three-electric system". The troubleshooting and maintenance of the three-electric system is highly professional, and the technical threshold is significantly increased.

[0003] However, in reality, the speed of technical updates of maintenance personnel is difficult to keep up with the rapid development of new energy vehicle technology, which often leads to decision-making difficulties when facing failures. In addition, maintenance personnel rely too much on personal experience when performing fault repairs, which may lead to misjudgment and reduce maintenance effectiveness.

[0004] Based on this, a method of determining fault operation and maintenance strategies based on a fault operation and maintenance big model has emerged. However, the current fault operation and maintenance big model has low accuracy and efficiency in determining operation and maintenance strategies, which urgently needs to be solved. Summary of the invention

[0005] Based on this, it is necessary to provide a training and use method, equipment and medium for a fault operation and maintenance big model to address the above technical problems, which can improve the accuracy and efficiency of the fault operation and maintenance big model in determining operation and maintenance strategies.

[0006] In a first aspect, the present application provides a method for training a large fault operation and maintenance model, including:

[0007] Based on the pre-trained fault operation and maintenance big model, the path capture loss of the reference reasoning path related to the operation and maintenance problem sample is generated from the fault operation and maintenance knowledge graph; the entities of the fault operation and maintenance knowledge graph include the problem description, cause of the problem and operation and maintenance strategy of the operation and maintenance problem; the edge relationship between the entities includes the causal relationship;

[0008] Based on the operation and maintenance problem samples, reference reasoning paths, and fault operation and maintenance knowledge graph, determine the retrieval reasoning loss of the operation and maintenance strategy labels corresponding to the operation and maintenance problem samples generated by the pre-trained fault operation and maintenance large model;

[0009] The pre-trained fault operation and maintenance large model is adjusted according to the path capture loss and / or retrieval reasoning loss.

[0010] In one of the embodiments, based on a pre-trained fault operation and maintenance big model, a path capture loss of a reference reasoning path related to an operation and maintenance problem sample is generated from a fault operation and maintenance knowledge graph, including: based on the pre-trained fault operation and maintenance big model, generating an initial reasoning path related to the operation and maintenance problem sample; determining the initial reasoning path to be a first probability distribution of a reference reasoning path in the fault operation and maintenance knowledge graph; and determining the path capture loss based on the difference between the first probability distribution and a preset path distribution.

[0011] In one of the embodiments, the method further includes: taking at least one shortest reasoning path between the operation and maintenance problem sample and the corresponding operation and maintenance strategy label in the fault operation and maintenance knowledge graph as a reference reasoning path.

[0012] In one of the embodiments, the path capture loss is determined based on the difference between the first probability distribution and the preset path distribution, including: determining the probability distribution of the first element in the reference reasoning path in the fault operation and maintenance knowledge graph for each generated element in the initial reasoning path; and determining the path capture loss based on the first element probability distribution and the preset path distribution.

[0013] In one of the embodiments, based on the operation and maintenance problem samples, the reference reasoning path and the fault operation and maintenance knowledge graph, the retrieval reasoning loss of the operation and maintenance strategy label corresponding to the operation and maintenance problem sample generated by the pre-trained fault operation and maintenance big model is determined, including: generating at least one key reasoning path based on the reference reasoning path; based on the pre-trained fault operation and maintenance big model, generating a second probability distribution of the operation and maintenance strategy label given the operation and maintenance problem samples, the key reasoning path and the fault operation and maintenance knowledge graph; generating the retrieval reasoning loss based on the second probability distribution.

[0014] In one of the embodiments, at least one key reasoning path is generated based on a reference reasoning path, including: based on a pre-trained fault operation and maintenance large model, determining the confidence of a reference reasoning path related to an operation and maintenance problem sample, and selecting at least one key reasoning path from the reference reasoning path based on the confidence of the reference reasoning path; and / or, based on a breadth-first algorithm, searching for at least one search reasoning path between an operation and maintenance problem sample and a corresponding operation and maintenance strategy label in a fault operation and maintenance knowledge graph, and generating a key reasoning path including a reference reasoning path and a search reasoning path.

[0015] In one of the embodiments, based on a pre-trained fault operation and maintenance big model, a second probability distribution of operation and maintenance strategy labels is generated given operation and maintenance problem samples, key reasoning paths and fault operation and maintenance knowledge graphs, including: inputting the key reasoning paths and operation and maintenance problem samples into the pre-trained fault operation and maintenance big model to generate a predicted reasoning path; determining the probability of each generated element in the predicted reasoning path being located at the second element of the operation and maintenance strategy label; and determining the second probability distribution corresponding to the predicted reasoning path based on the second element probability.

[0016] In a second aspect, the present application also provides a method for using a large fault operation and maintenance model, including:

[0017] Obtain the operation and maintenance problem data to be tested;

[0018] Based on the trained fault operation and maintenance big model, determine the answer strategy corresponding to the operation and maintenance problem data to be tested; wherein the fault operation and maintenance big model is trained using the above-mentioned fault operation and maintenance big model training method.

[0019] In a third aspect, the present application also provides a training device for a large fault operation and maintenance model, including:

[0020] A generation module is used to generate a path capture loss of a reference reasoning path related to an operation and maintenance problem sample from a fault operation and maintenance knowledge graph based on a pre-trained fault operation and maintenance large model; wherein the entities of the fault operation and maintenance knowledge graph include a problem description of the operation and maintenance problem, a cause of the problem, and an operation and maintenance strategy; and the edge relationship between the entities includes a causal relationship;

[0021] The first determination module is used to determine the retrieval reasoning loss of the operation and maintenance problem sample corresponding to the operation and maintenance strategy label generated by the pre-trained fault operation and maintenance large model based on the operation and maintenance problem sample, the reference reasoning path and the fault operation and maintenance knowledge graph;

[0022] The adjustment module is used to adjust the pre-trained fault operation and maintenance model according to the path capture loss and / or the retrieval reasoning loss.

[0023] In a fourth aspect, the present application also provides a device for using a large fault operation and maintenance model, including:

[0024] The acquisition module is used to obtain the operation and maintenance problem data to be tested;

[0025] The second determination module is used to determine the answer strategy corresponding to the operation and maintenance problem data to be tested based on the trained fault operation and maintenance big model; wherein the fault operation and maintenance big model is trained using the above-mentioned fault operation and maintenance big model training method.

[0026] In a fifth aspect, the present application further provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0027] Based on the pre-trained fault operation and maintenance big model, the path capture loss of the reference reasoning path related to the operation and maintenance problem sample is generated from the fault operation and maintenance knowledge graph; the entities of the fault operation and maintenance knowledge graph include the problem description, cause of the problem and operation and maintenance strategy of the operation and maintenance problem; the edge relationship between the entities includes the causal relationship;

[0028] Based on the operation and maintenance problem samples, reference reasoning paths, and fault operation and maintenance knowledge graph, determine the retrieval reasoning loss of the operation and maintenance strategy labels corresponding to the operation and maintenance problem samples generated by the pre-trained fault operation and maintenance large model;

[0029] The pre-trained fault operation and maintenance large model is adjusted according to the path capture loss and / or retrieval reasoning loss.

[0030] In a sixth aspect, the present application further provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0031] Obtain the operation and maintenance problem data to be tested;

[0032] Based on the trained fault operation and maintenance big model, determine the answer strategy corresponding to the operation and maintenance problem data to be tested; wherein the fault operation and maintenance big model is trained using the above-mentioned fault operation and maintenance big model training method.

[0033] In a seventh aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the following steps are implemented:

[0034] Based on the pre-trained fault operation and maintenance big model, the path capture loss of the reference reasoning path related to the operation and maintenance problem sample is generated from the fault operation and maintenance knowledge graph; the entities of the fault operation and maintenance knowledge graph include the problem description, cause of the problem and operation and maintenance strategy of the operation and maintenance problem; the edge relationship between the entities includes the causal relationship;

[0035] Based on the operation and maintenance problem samples, reference reasoning paths, and fault operation and maintenance knowledge graph, determine the retrieval reasoning loss of the operation and maintenance strategy labels corresponding to the operation and maintenance problem samples generated by the pre-trained fault operation and maintenance large model;

[0036] The pre-trained fault operation and maintenance large model is adjusted according to the path capture loss and / or retrieval reasoning loss.

[0037] In an eighth aspect, the present application further provides a computer program product, including a computer program, which implements the following steps when executed by a processor:

[0038] Obtain the operation and maintenance problem data to be tested;

[0039] Based on the trained fault operation and maintenance big model, determine the answer strategy corresponding to the operation and maintenance problem data to be tested; wherein the fault operation and maintenance big model is trained using the above-mentioned fault operation and maintenance big model training method.

[0040] The training and use method, device and medium of the above-mentioned fault operation and maintenance big model, in the model training stage, based on the pre-trained fault operation and maintenance big model, generates the path capture loss of the reference reasoning path related to the operation and maintenance problem sample from the fault operation and maintenance knowledge graph. In this process, since the fault operation and maintenance knowledge graph records the historical fault records, the accuracy of the reference reasoning path is relatively high. In other words, adjusting the pre-trained fault operation and maintenance big model according to the path capture loss can minimize the probability that the pre-trained fault operation and maintenance big model generates problematic or inaccurate operation and maintenance strategies when answering, thereby improving the model accuracy of the fault operation and maintenance big model. In addition, since the retrieval reasoning loss is used to characterize the probability that the pre-trained fault operation and maintenance big model generates the operation and maintenance strategy label corresponding to the operation and maintenance problem sample, therefore, adjusting the pre-trained fault operation and maintenance big model according to the retrieval reasoning loss can increase the probability that the fault operation and maintenance big model generates the operation and maintenance strategy label when answering, thereby improving the model efficiency of the fault operation and maintenance strategy and reducing the loss of computing resources. Furthermore, in the model use phase, the aforementioned adjusted fault operation and maintenance big model is directly reused to process the operation and maintenance problem data to be tested, thereby improving the convenience and versatility of operation and maintenance. In addition, since the calculation efficiency and accuracy of the adjusted fault operation and maintenance big model are improved to a certain extent, the processing efficiency of the operation and maintenance problem data to be tested and the accuracy of the answer strategy determination results are improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the drawings required for use in the embodiments of the present application or related technical descriptions will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0042] Figure 1A It is a flowchart of a method for training a large fault operation and maintenance model in one embodiment;

[0043] Figure 1B A schematic diagram of a fault operation and maintenance knowledge graph in one embodiment;

[0044] Figure 2 A schematic diagram of a flow chart of a path capture loss determination step in one embodiment;

[0045] Figure 3 A schematic flow chart of a path capture loss determination step in another embodiment;

[0046] Figure 4 A flowchart of a step of determining a retrieval reasoning loss in one embodiment;

[0047] Figure 5 A training structure diagram of a fault operation and maintenance large model in an embodiment;

[0048] Figure 6 It is a flowchart of a method for training a large fault operation and maintenance model in another embodiment;

[0049] Figure 7 It is a structural block diagram of a training device for a large fault operation and maintenance model in an embodiment;

[0050] Figure 8 It is a structural block diagram of a device for using a large fault operation and maintenance model in an embodiment;

[0051] Fig. 9 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0053] In some embodiments, Figure 1A As shown, a method for training a large fault operation and maintenance model is provided, and the method is applied to a terminal for example. It can be understood that the method can also be applied to a server, and can also be applied to a system including a terminal and a server, and is implemented through the interaction between the terminal and the server. In some embodiments, the method includes the following steps:

[0054] S110, based on the pre-trained fault operation and maintenance large model, generates the path capture loss of the reference reasoning path related to the operation and maintenance problem sample from the fault operation and maintenance knowledge graph.

[0055] Among them, the pre-trained fault operation and maintenance large model can be a generative model, which is used to process the problem description data of new energy vehicle failures and generate corresponding fault operation and maintenance strategies.

[0056] Among them, the fault operation and maintenance knowledge graph can be constructed based on historical fault records in the operation and maintenance scenario. For easy understanding, the fault operation and maintenance knowledge graph is first briefly introduced.

[0057] like Figure 1B As shown in the figure, the entities of the fault operation and maintenance knowledge graph can include the problem description of the operation and maintenance problem, the cause of the problem (such as Figure 1Bhigh temperature environment, excessive charging current, high voltage and battery overload) and operation and maintenance strategies (such as Figure 1B Liquid cooling, avoiding overcharging and avoiding overloading in the problem description includes the problem result description (such as Figure 1B thermal failure in the Figure 1B thermal failure, sparks, temperature anomalies, and smoke in the Figure 1B The phenomena, causes and solutions.

[0058] Correspondingly, the operation and maintenance problem samples may be problem description data under the fault operation and maintenance scenario in the model training phase, and illustratively, may be problem descriptions under historical operation and maintenance scenarios.

[0059] The reference reasoning path is used to characterize the reasoning path related to the operation and maintenance problem sample and in the fault operation and maintenance knowledge graph. In some embodiments, the operation and maintenance problem sample can be processed based on the pre-trained fault operation and maintenance large model to generate an operation and maintenance strategy, and the operation and maintenance strategy in the fault operation and maintenance knowledge graph in the operation and maintenance strategy generation result is used as the reference reasoning path.

[0060] Among them, the path capture loss can be used to characterize the probability of the existence of a reference reasoning path in the reasoning path of the operation and maintenance strategy generated by the pre-trained fault operation and maintenance model.

[0061] Exemplarily, the operation and maintenance problem samples can be processed based on the pre-trained fault operation and maintenance large model. For any operation and maintenance problem sample, generate at least one operation and maintenance strategy, and determine the inclusion of reference reasoning paths in each operation and maintenance strategy; generate reference losses based on the inclusion of reference reasoning paths in the generated operation and maintenance strategies, such as the proportion of reference reasoning paths in the generated operation and maintenance strategies; determine the path capture loss based on the reference losses corresponding to each operation and maintenance problem sample. For example, the weighted sum of the reference losses corresponding to different operation and maintenance problem samples can be used as the path capture loss. Among them, the weight of each reference loss can be determined based on manual experience, and this application does not impose any restrictions on this.

[0062] S120, based on the operation and maintenance problem samples, the reference reasoning path and the fault operation and maintenance knowledge graph, determine the retrieval reasoning loss of the operation and maintenance strategy labels corresponding to the operation and maintenance problem samples generated by the pre-trained fault operation and maintenance large model.

[0063] Among them, the operation and maintenance strategy label corresponding to the operation and maintenance problem sample is the pre-labeled standard operation and maintenance strategy of the operation and maintenance problem sample, which can be manually labeled or other traditional labeling methods, and this application does not impose any restrictions on this.

[0064] Among them, the retrieval reasoning loss can be used to characterize the probability that the pre-trained fault operation and maintenance model can predict the operation and maintenance strategy label.

[0065] Exemplarily, the operation and maintenance problem samples, reference reasoning paths and fault operation and maintenance knowledge graph can be input into a pre-trained probability determination model to obtain the probability that the pre-trained fault operation and maintenance large model generates the operation and maintenance problem samples corresponding to the operation and maintenance strategy labels; the retrieval reasoning loss is determined based on the probability.

[0066] Among them, the probability determination model can be constructed based on a common neural network. For example, when training the probability determination model, operation and maintenance problem samples, sample reasoning paths and sample knowledge graphs can be input into the probability determination model to obtain the predicted probability.

[0067] S130: Adjust the pre-trained fault operation and maintenance large model according to the path capture loss and / or the retrieval reasoning loss.

[0068] Optionally, the pre-trained fault operation and maintenance large model can be adjusted (also known as fine-tuning) according to the path capture loss; the pre-trained fault operation and maintenance large model can also be adjusted according to the retrieval reasoning loss; the pre-trained fault operation and maintenance large model can also be adjusted according to the path capture loss and the retrieval reasoning loss.

[0069] Exemplarily, the pre-trained fault operation and maintenance large model can be adjusted according to the path capture loss and / or the retrieval reasoning loss based on at least one of gradient descent, momentum optimized coordinate descent, and adaptive learning rate optimization algorithm.

[0070] In the above-mentioned training method of the fault operation and maintenance big model, based on the pre-trained fault operation and maintenance big model, the path capture loss of the reference reasoning path related to the operation and maintenance problem sample is generated from the fault operation and maintenance knowledge graph. In this process, since the fault operation and maintenance knowledge graph records the historical fault records, the accuracy of the reference reasoning path is relatively high. In other words, adjusting the pre-trained fault operation and maintenance big model according to the path capture loss can minimize the probability that the pre-trained fault operation and maintenance big model generates problematic or inaccurate operation and maintenance strategies when answering, thereby improving the model accuracy of the fault operation and maintenance big model. In addition, since the retrieval reasoning loss is used to characterize the probability that the pre-trained fault operation and maintenance big model generates the operation and maintenance strategy label corresponding to the operation and maintenance problem sample, therefore, adjusting the pre-trained fault operation and maintenance big model according to the retrieval reasoning loss can increase the probability that the fault operation and maintenance big model generates the operation and maintenance strategy label when answering, thereby improving the model efficiency of the fault operation and maintenance strategy and reducing the loss of computing resources.

[0071] Based on the technical solutions of the above embodiments, the present application also provides an optional embodiment. In this optional embodiment, the process of generating the path capture loss of the reference reasoning path related to the operation and maintenance problem sample from the fault operation and maintenance knowledge graph based on the pre-trained fault operation and maintenance large model is refined.

[0072] See also Figure 2 The steps for determining the path capture loss shown include:

[0073] S210, based on the pre-trained fault operation and maintenance large model, generate an initial reasoning path related to the operation and maintenance problem sample.

[0074] The initial reasoning path may be a reasoning path in the operation and maintenance strategy generated after the pre-trained fault operation and maintenance large model processes the operation and maintenance problem sample. It is understandable that the initial reasoning path is not necessarily in the fault operation and maintenance knowledge graph.

[0075] Exemplarily, the operation and maintenance problem sample can be input into a pre-trained fault operation and maintenance large model to obtain at least one operation and maintenance strategy related to the operation and maintenance problem sample, and the reasoning path in the at least one operation and maintenance strategy is used as the initial reasoning path related to the operation and maintenance problem sample.

[0076] S220, determining that the initial reasoning path is a first probability distribution of a reference reasoning path in the fault operation and maintenance knowledge graph.

[0077] The first probability distribution may be used to characterize the probability that the initial reasoning path belongs to the reference reasoning path.

[0078] In an optional implementation, when the pre-trained fault operation and maintenance large model outputs the initial reasoning path, it will simultaneously output the probability that the corresponding initial reasoning path is a reference reasoning path in the fault operation and maintenance knowledge graph, thereby obtaining a first probability distribution of different initial reasoning paths.

[0079] S230: Determine the path capture loss according to the difference between the first probability distribution and the preset path distribution.

[0080] The preset path distribution may be a posterior distribution, for example, may be at least one of a normal distribution, a t distribution, a binomial distribution, a Poisson distribution, and a mean distribution, etc. In order to facilitate calculation and accelerate convergence, in some embodiments, the preset path distribution is described by taking an average distribution as an example.

[0081] In an optional implementation, the path capture loss may be determined based on the distance between the first probability distribution and the preset path distribution; or, the log-likelihood function loss between the first probability distribution and the preset path distribution may be used as the path capture loss.

[0082] In another optional implementation, the path capture loss may be generated according to the KL divergence (Kullback-Leibler Divergence, KLD) between the first probability distribution and the preset path distribution. Exemplarily, the path capture loss may be determined according to the following formula:

[0083] ;

[0084] In the formula, represents the path capture loss; Represents the first probability distribution With preset path distribution KL divergence between; first probability distribution Characterize the probability of the fault operation and maintenance large model generating a reference reasoning path z when given an operation and maintenance problem sample q; Indicates the parameters to be adjusted in the fault operation and maintenance model.

[0085] Exemplarily, the path capture module extracts knowledge from the fault operation and maintenance knowledge graph to generate a reference reasoning path in the fault operation and maintenance large model, and improves the accuracy of path reasoning by minimizing the KL divergence between the reference reasoning path and the posterior distribution. Take uniform distribution as an example to illustrate. For example, given an operation and maintenance problem sample q and an operation and maintenance strategy label a, the problem entity corresponding to the connection operation and maintenance problem sample q can be obtained through the fault operation and maintenance knowledge graph. The policy entity corresponding to the operation and maintenance policy label a Reference reasoning path between , the corresponding relationship set is , where r1, r2...ri represent the relationship between entities in the fault operation and maintenance knowledge graph. Accordingly, the preset path distribution can be determined based on the following formula:

[0086] ;

[0087] In the formula, represents uniform distribution; It represents the posterior probability of generating a reference reasoning path z given the operation problem sample q, the operation strategy label a and the fault operation knowledge graph G; Z represents the number of reference reasoning paths; Represents the problem entity in the fault operation and maintenance knowledge graph and policy entities There are reference reasoning paths between them; Indicates the parameters to be adjusted in the fault operation and maintenance model.

[0088] In some embodiments, by minimizing the path capture loss, the probability that the fault operation and maintenance model extracts knowledge from the fault operation and maintenance knowledge graph and obtains a reliable reasoning path can be maximized.

[0089] In the above embodiment, an implementation method for determining the path capture loss is provided. The path capture loss is determined based on the difference between the first probability distribution of the reference reasoning path in the fault operation and maintenance knowledge graph and the preset path distribution, and then the fault operation and maintenance big model is trained based on the path capture loss, which can promote the fault operation and maintenance big model to generate reliable reasoning paths.

[0090] Optionally, since the reference reasoning path is the operation and maintenance strategy generated after processing the operation and maintenance problem samples based on the pre-trained fault operation and maintenance large model, and is in the fault operation and maintenance knowledge graph, and the number of reasoning paths corresponding to the operation and maintenance problem samples in the fault knowledge graph is relatively large, therefore, determining the path capture loss based on the full amount of reasoning paths corresponding to the operation and maintenance problem samples in the fault knowledge graph will bring about an increase in data calculations and increase computing power costs.

[0091] In order to reduce computing power cost investment and improve data computing efficiency, in some embodiments, at least one shortest reasoning path between the operation and maintenance problem sample and the corresponding operation and maintenance strategy label in the fault operation and maintenance knowledge graph can be used as a reference reasoning path, thereby reducing the amount of data computing corresponding to the non-shortest reasoning path.

[0092] Accordingly, the path capture loss can be determined by the following formula:

[0093] ;

[0094] In the formula, represents the path capture loss; It means that given the operation and maintenance problem sample q, the operation and maintenance strategy label a and the fault operation and maintenance knowledge graph G, the posterior probability of the reference reasoning path z is generated, corresponding to the preset path distribution; represents the first probability distribution; Z represents the reference reasoning path set; It represents the set of shortest reasoning paths between the entity corresponding to the operation and maintenance problem sample q and the entity corresponding to the operation and maintenance strategy label a; Indicates the parameters to be adjusted in the fault operation and maintenance model.

[0095] Optionally, in order to further narrow the scope of reference reasoning paths, a preset number of shortest reasoning paths with higher confidence among at least one shortest reasoning path between the operation and maintenance problem sample and the corresponding operation and maintenance strategy label in the fault operation and maintenance knowledge graph can be used as reference reasoning paths; or, the shortest reasoning path whose confidence exceeds a preset confidence threshold can be used as a reference reasoning path.

[0096] Exemplarily, the above-mentioned preset number and preset confidence threshold can be determined based on manual experience or through a large number of experiments, and this application does not impose any limitation on this.

[0097] Based on the technical solutions of the above embodiments, the present application also provides an optional embodiment. In this optional embodiment, an implementation method for determining the path capture loss is also provided.

[0098] See also Figure 3 The steps for determining the path capture loss shown include:

[0099] S310, determining the probability distribution of the first element in the reference reasoning path in the fault operation and maintenance knowledge graph for each generated element in the initial reasoning path.

[0100] Each generated element (token) in the initial reasoning path may be the smallest processing unit involved in the initial reasoning path, such as a word segment. The first element probability distribution is used to characterize the probability that each generated element in the initial reasoning path is located in the reference reasoning path.

[0101] Exemplarily, each generated element in the initial reasoning path and each generated element in the reference reasoning path can be determined, and further, the probability that each generated element in the initial reasoning path belongs to each generated element in the reference reasoning path can be determined as the first element probability distribution.

[0102] S320, determining a path capture loss according to the first element probability distribution and a preset path distribution.

[0103] Exemplarily, the path capture loss may be determined according to the following formula:

[0104] ;

[0105] In the formula, Represents the probability of each generated element (token) in the initial reasoning path generated by the fault operation and maintenance large model in the reference reasoning path z, corresponding to the first element probability distribution; represents the path capture loss; represents the first probability distribution; Indicates the parameters to be adjusted in the fault operation and maintenance model.

[0106] In the above embodiment, another specific method for determining the path capture loss is given. The path capture loss is determined based on the generated elements in the initial reasoning path, the probability distribution of the first element in the reference reasoning path in the fault operation and maintenance knowledge graph, and the preset path distribution, thereby improving the accuracy of the path capture loss and providing better data support for model adjustment of the fault operation and maintenance large model.

[0107] It can be understood that the process of adjusting the network parameters of the fault operation and maintenance model based on the path capture loss can be understood as finding the maximum Therefore, the process of optimizing the objective function of the path capture module can be expressed as:

[0108] ;

[0109] In the formula, Characterizes the generated elements r in the initial reasoning path generated by the fault operation and maintenance large model i The probability of (token) in the reference reasoning path z corresponds to the first element probability distribution; represents the path capture loss; represents the first probability distribution; Indicates the parameters to be adjusted in the fault operation and maintenance model.

[0110] Based on the technical solutions of the above embodiments, the present application also provides an optional embodiment. In this optional embodiment, the process of determining the retrieval reasoning loss of the operation and maintenance problem sample corresponding to the operation and maintenance strategy label generated by the pre-trained fault operation and maintenance large model based on the operation and maintenance problem sample, the reference reasoning path and the fault operation and maintenance knowledge graph is described in detail.

[0111] See also Figure 4 The steps for determining the retrieval inference loss shown include:

[0112] S410: Generate at least one key reasoning path according to the reference reasoning path.

[0113] Among them, the key reasoning path can be understood as a path used to assist retrieval reasoning.

[0114] In an optional implementation, at least one key reasoning path may be selected from the reference reasoning paths based on the confidence corresponding to each reference reasoning path output by the pre-trained fault operation and maintenance large model.

[0115] For example, a preset number of reference reasoning paths with a higher confidence level (e.g., the highest) can be used as the key reasoning path; or a reference reasoning path with a confidence level exceeding a preset confidence threshold can be used as the key reasoning path. The preset number and the preset confidence threshold can be determined based on manual experience or through a large number of experiments, and this application does not impose any limitation on this.

[0116] Furthermore, in the process of training the fault operation and maintenance large model, in order to avoid overfitting of the fault operation and maintenance large model and thus improve the robustness of the fault operation and maintenance large model. In another optional implementation, based on a breadth-first algorithm, at least one search reasoning path between the operation and maintenance problem sample and the corresponding operation and maintenance strategy label can be searched in the fault operation and maintenance knowledge graph, and a key reasoning path including a reference reasoning path and a search reasoning path can be generated.

[0117] The search reasoning path is the other reasoning path between the operation and maintenance problem sample and the corresponding operation and maintenance strategy label in the fault operation and maintenance knowledge graph, except for the reference reasoning path. Usually, the number of entities carried by the search reasoning path is greater than the number of entities carried by the reference reasoning path; or, the number of relationships carried by the search reasoning path is greater than the number of relationships carried by the reference reasoning path.

[0118] Exemplarily, at least one search reasoning path can be expressed as:

[0119] ;

[0120] In the formula, is a set of search reasoning paths; Representing the problem entity To answer entity The path between them; G represents the fault operation and maintenance knowledge graph.

[0121] Optionally, the search reasoning path and the reference reasoning path may be taken together as at least one key reasoning path.

[0122] S420, based on the pre-trained fault operation and maintenance large model, a second probability distribution of the operation and maintenance strategy label is generated given the operation and maintenance problem sample, the key reasoning path and the fault operation and maintenance knowledge graph.

[0123] Exemplarily, when the pre-trained fault operation and maintenance large model processes the operation and maintenance problem samples given the key reasoning path and the fault operation and maintenance knowledge graph, it will synchronously output the second probability distribution of the generated operation and maintenance strategy label.

[0124] Exemplarily, the second probability distribution may be as follows:

[0125] ;

[0126] Where P is the second probability distribution, which represents the key reasoning path for a given operation and maintenance problem sample q. and the fault operation and maintenance knowledge graph G, to generate the probability of the operation and maintenance strategy label a; Indicates the parameters to be adjusted in the fault operation and maintenance model.

[0127] Optionally, the key reasoning path and operation and maintenance problem samples can be input into a pre-trained fault operation and maintenance large model to generate a predictive reasoning path; determine the probability of the second element of each generated element in the predictive reasoning path being located in the operation and maintenance strategy label; and determine the second probability distribution corresponding to the predictive reasoning path based on the second element probability.

[0128] Exemplarily, the second element probability distribution may be as follows:

[0129] ;

[0130] Where P is the probability distribution of the second element; t i Indicates the generated element (corresponding to token) corresponding to the operation and maintenance strategy label a; w z Represents the search reasoning path; Characterizes each generated element t in the initial reasoning path generated by the fault operation and maintenance large model i The probability in the reference reasoning path z corresponds to the second element probability; among them, the correct control strategy is the predicted reasoning path that can be traced back to the operation and maintenance strategy label.

[0131] Further, according to the second element probability distribution, a second probability distribution is determined. Exemplarily, the second probability distribution can be determined according to the following formula:

[0132] ;

[0133] In the formula, Characterize the key reasoning path for a given operation and maintenance problem sample q and the fault operation knowledge graph G to generate the probability of the operation strategy label a, i.e., the second probability distribution; t i Indicates the token corresponding to the correct control strategy a; w z Represents the search reasoning path; Characterizes each generated element t in the initial reasoning path generated by the fault operation and maintenance large model i The probability in the reference reasoning path z corresponds to the probability of the second element; is a set of search reasoning paths; Indicates the parameters to be adjusted in the fault operation and maintenance model.

[0134] S430: Generate a retrieval reasoning loss according to the second probability distribution.

[0135] Exemplarily, the retrieval reasoning loss may be generated based on the following formula:

[0136] ;

[0137] In the formula, represents the retrieval reasoning loss; Characterize the key reasoning path for a given operation and maintenance problem sample q and the fault operation and maintenance knowledge graph G to generate the probability of the operation and maintenance strategy label a, i.e., the second probability distribution; Indicates the parameters to be adjusted in the fault operation and maintenance model.

[0138] In the above embodiment, a specific method for generating retrieval reasoning loss is given. Since the key reasoning path generated based on the breadth-first algorithm may contain some noise, the set of key reasoning paths and the operation and maintenance problem samples are input into the fault operation and maintenance big model together, and the fault operation and maintenance big model is used to infer the operation and maintenance strategy labels corresponding to the operation and maintenance problem samples, which can reduce the error caused by noise and improve the accuracy of the reasoning path output by the fault operation and maintenance big model, thereby improving the accuracy of the fault operation and maintenance big model.

[0139] Similarly, the process of adjusting the fault operation and maintenance model based on retrieval reasoning loss can be understood as finding the maximum Therefore, the process of optimizing the objective function by the retrieval reasoning module can be further expressed as:

[0140] ;

[0141] In the formula, Characterize the key reasoning path for a given operation and maintenance problem sample q and the fault operation knowledge graph G to generate the probability of the operation strategy label a, i.e., the second probability distribution; t i Indicates the token corresponding to the correct control strategy a; w z Represents the search reasoning path; Characterizes each generated element t in the initial reasoning path generated by the fault operation and maintenance large model i The probability in the reference reasoning path z corresponds to the probability of the second element; is a set of search reasoning paths; Indicates the parameters to be adjusted in the fault operation and maintenance model.

[0142] Based on the technical solutions of the above embodiments, the present application also provides an optional embodiment. In this optional embodiment, the training method of the fault operation and maintenance large model provided by the present application is introduced in detail.

[0143] See also Figure 5From the training structure diagram of the fault operation and maintenance big model shown in the figure, it can be seen that the fault operation and maintenance big model in this embodiment is provided with a path capture module and a retrieval reasoning module. Among them, the path capture module mainly uses the fault operation and maintenance big model to generate reference reasoning paths related to operation and maintenance problem samples from the fault operation and maintenance knowledge graph; the retrieval reasoning module mainly screens from the reference reasoning paths, and infers the operation and maintenance strategy labels corresponding to the fault operation and maintenance samples based on the screened paths.

[0144] Based on this, the fault operation and maintenance big model of this application is combined with the fault operation and maintenance knowledge graph, and the probability of inferring the corresponding operation and maintenance strategy label based on the operation and maintenance problem sample can be summarized as the following formula:

[0145] ;

[0146] In the formula, Characterize the probability of the fault operation and maintenance model inferring the corresponding operation and maintenance strategy label based on the operation and maintenance problem samples; represents the parameters to be adjusted in the fault operation and maintenance model; Z represents the set of reference reasoning paths; It represents the probability of generating a reference reasoning path z based on the fault operation and maintenance knowledge graph when a given operation and maintenance problem sample q is given. It can be calculated by the path capture module. When characterizing a given operation and maintenance problem sample q, a reference reasoning path z, and a fault operation and maintenance knowledge graph G, the probability of the fault operation and maintenance large model generating an operation and maintenance strategy label a corresponding to the operation and maintenance problem sample q can be calculated through the retrieval reasoning module.

[0147] Furthermore, considering that the fault operation and maintenance big model has no prior knowledge of the relationship between entities in the fault operation and maintenance knowledge graph, the fault operation and maintenance big model cannot directly generate a reference reasoning path based on the fault operation and maintenance knowledge graph. In addition, the fault operation and maintenance big model cannot correctly understand the reference reasoning path and make appropriate inferences based on the reference reasoning path to obtain the operation and maintenance strategy label corresponding to the operation and maintenance problem sample. In order to improve the probability of the fault operation and maintenance big model to capture the path and infer the operation and maintenance strategy label, and also to calculate The value of can be maximized by maximizing its variational lower bound. The value of is estimated. This process can be expressed as the following formula:

[0148] ;

[0149] In the formula, Characterize the probability of the fault operation and maintenance model inferring the corresponding operation and maintenance strategy label based on the operation and maintenance problem samples; It represents the probability that the fault operation and maintenance big model generates the operation and maintenance strategy label a corresponding to the operation and maintenance problem sample q when the operation and maintenance problem sample q, the reference reasoning path z, and the fault operation and maintenance knowledge graph G are given; Indicates the preset path distribution, which is an average distribution; Represents the first probability distribution With preset path distribution The KL divergence between It represents the parameter to be adjusted in the fault operation and maintenance model; the right side of the inequality sign represents the variational lower bound; It represents the expected value of the operation and maintenance strategy label corresponding to the operation and maintenance problem sample generated based on the reference reasoning path z and the fault operation and maintenance knowledge graph G.

[0150] Optionally, the KL divergence can be minimized to increase the probability that the fault operation model can infer the corresponding operation strategy label based on the operation problem sample. Accordingly, the value of the path capture loss can be determined by the following formula:

[0151] ;

[0152] In the formula, represents the path capture loss; It means that given the operation and maintenance problem sample q, the operation and maintenance strategy label a and the fault operation and maintenance knowledge graph G, the posterior probability of the reference reasoning path z is generated, corresponding to the preset path distribution; represents the first probability distribution; Z represents the reference reasoning path set; It represents the set of shortest reasoning paths between the entity corresponding to the operation and maintenance problem sample q and the entity corresponding to the operation and maintenance strategy label a; Indicates the parameters to be adjusted in the fault operation and maintenance model.

[0153] The present application also provides a method for determining path capture loss. Exemplarily, a preset path generation template is input into the fault operation and maintenance big model, so that the fault operation and maintenance big model can perform path reasoning according to the preset path generation module and generate a reasoning path. The preset path generation template can be as follows:

[0154] <prompt>Please generate an inference path based on the fault operation and maintenance problem;

[0155] <question> ；

[0156] <response> <path> r1 <sep> r2 <sep> ... <sep> r i < / sep> < / sep> < / sep> < / path> ;

[0157] in, <path> 、 <sep> 、< / sep> < / path> Respectively represent the starting point, separator and end point of the reasoning path; r1, r2...r i They are the entities involved in the reasoning path process.

[0158] In addition, the retrieval reasoning loss can be determined by the following formula:

[0159] ;

[0160] In the formula, represents the retrieval reasoning loss; It represents the expected value of the operation and maintenance strategy label a corresponding to the operation and maintenance problem sample generated based on the reference reasoning path z and the fault operation and maintenance knowledge graph G; Characterize the key reasoning path for a given operation and maintenance problem sample q and the fault operation and maintenance knowledge graph G to generate the probability of the operation and maintenance strategy label a, i.e., the second probability distribution; Indicates the parameters to be adjusted in the fault operation and maintenance model.

[0161] According to the path capture loss and retrieval reasoning loss, the target loss is determined, and according to the target loss, the model parameters of the fault operation and maintenance model are adjusted. For example, the target loss may be as follows:

[0162] ;

[0163] Where L is the target loss; represents the path capture loss; represents the retrieval inference loss.

[0164] In some embodiments, Figure 6 As shown, a method for using a fault operation and maintenance large model is provided, and the method is applied to a terminal as an example. It can be understood that the method can also be applied to a server, and can also be applied to a system including a terminal and a server, and is implemented through the interaction between the terminal and the server. The method includes the following steps:

[0165] S610, obtaining operation and maintenance problem data to be tested.

[0166] The operation and maintenance problem data to be tested may be operation and maintenance problem data for which a fault operation and maintenance strategy is required to be determined. The operation and maintenance problem data to be tested may include problem description data of the operation and maintenance problem to be tested.

[0167] Exemplarily, when there is a need to determine a fault operation and maintenance strategy, the operation and maintenance problem data to be tested is obtained. For example, the problem description data uploaded by the staff through the operation and maintenance problem platform can be obtained, and the problem description data is used as the operation and maintenance problem data to be tested.

[0168] S620: Determine the answer strategy corresponding to the operation and maintenance problem data to be tested based on the trained fault operation and maintenance big model.

[0169] Among them, the fault operation and maintenance big model is trained based on the above-mentioned training method of the fault operation and maintenance big model.

[0170] For example, the operation and maintenance problem data to be tested can be input into the trained fault operation and maintenance large model to obtain the answer strategy corresponding to the operation and maintenance problem data to be tested, and then the operation and maintenance of the corresponding operation and maintenance problem data to be tested can be performed based on the answer strategy.

[0171] In the above embodiment, the aforementioned adjusted fault operation and maintenance big model is directly reused to process the operation and maintenance problem data to be tested, thereby improving the convenience and versatility of operation and maintenance. In addition, since the calculation efficiency and accuracy of the adjusted fault operation and maintenance big model are improved to a certain extent, the processing efficiency of processing the operation and maintenance problem data to be tested and the accuracy of the answer strategy determination result are improved.

[0172] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0173] Based on the same inventive concept, the embodiment of the present application also provides a training device for a fault operation and maintenance large model for implementing the training method for the fault operation and maintenance large model involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in the embodiments of the training device for one or more fault operation and maintenance large models provided below can refer to the limitations of the training method for the fault operation and maintenance large model above, and will not be repeated here.

[0174] In an exemplary embodiment, Figure 7 As shown, a training device for a fault operation and maintenance large model is provided, including: a generation module 710, a first determination module 720 and an adjustment module 730, wherein:

[0175] The generation module 710 is used to generate the path capture loss of the reference reasoning path related to the operation and maintenance problem sample from the fault operation and maintenance knowledge graph based on the pre-trained fault operation and maintenance large model; wherein the entities of the fault operation and maintenance knowledge graph include the problem description, the cause of the problem and the operation and maintenance strategy of the operation and maintenance problem; and the edge relationship between the entities includes the causal relationship;

[0176] The first determination module 720 is used to determine the retrieval reasoning loss of the operation and maintenance problem sample corresponding to the operation and maintenance strategy label generated by the pre-trained fault operation and maintenance large model according to the operation and maintenance problem sample, the reference reasoning path and the fault operation and maintenance knowledge graph;

[0177] The adjustment module 730 is used to adjust the pre-trained fault operation and maintenance large model according to the path capture loss and / or the retrieval reasoning loss.

[0178] In some embodiments, the generation module 710 includes a first generation unit, which is used to generate an initial reasoning path related to the operation and maintenance problem sample based on a pre-trained fault operation and maintenance large model; a first determination unit, which is used to determine that the initial reasoning path is a first probability distribution of a reference reasoning path in the fault operation and maintenance knowledge graph; and a second determination unit, which is used to determine the path capture loss based on the difference between the first probability distribution and the preset path distribution.

[0179] In some embodiments, the generation module 710 also includes a third determination unit, which is used to use at least one shortest reasoning path between the operation and maintenance problem sample and the corresponding operation and maintenance strategy label in the fault operation and maintenance knowledge graph as a reference reasoning path.

[0180] In some embodiments, the second determination subunit includes a first determination subunit, which is used to determine the probability distribution of the first element in the reference reasoning path in the fault operation and maintenance knowledge graph for each generated element in the initial reasoning path; and a second determination subunit, which is used to determine the path capture loss based on the first element probability distribution and the preset path distribution.

[0181] In some embodiments, the first determination module 720 includes a second generation unit for generating at least one key reasoning path based on a reference reasoning path; a third generation unit for generating a second probability distribution of operation and maintenance strategy labels based on a pre-trained fault operation and maintenance large model, given an operation and maintenance problem sample, a key reasoning path and a fault operation and maintenance knowledge graph; and a fourth generation unit for generating a retrieval reasoning loss based on the second probability distribution.

[0182] In some embodiments, the second generation unit includes a first generation sub-unit, which is used to determine the confidence of a reference reasoning path related to the operation and maintenance problem sample based on a pre-trained fault operation and maintenance large model, and select at least one key reasoning path from the reference reasoning path according to the confidence of the reference reasoning path; and / or, the second generation sub-unit is used to search for at least one search reasoning path between the operation and maintenance problem sample and the corresponding operation and maintenance strategy label in the fault operation and maintenance knowledge graph based on a breadth-first algorithm, and generate a key reasoning path including a reference reasoning path and a search reasoning path.

[0183] In some embodiments, the third generation unit includes a third generation sub-unit, which is used to input the key reasoning path and operation and maintenance problem samples into a pre-trained fault operation and maintenance large model to generate a predictive reasoning path; a third determination sub-unit, which is used to determine the second element probability of each generated element in the predictive reasoning path being located in the operation and maintenance strategy label; and a fourth determination sub-unit, which is used to determine the second probability distribution corresponding to the predictive reasoning path based on the second element probability.

[0184] Based on the same inventive concept, the embodiment of the present application also provides a device for using a fault operation and maintenance large model for implementing the method for using the fault operation and maintenance large model involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in the embodiments of the device for using one or more fault operation and maintenance large models provided below can refer to the limitations of the method for using the fault operation and maintenance large model above, and will not be repeated here.

[0185] In an exemplary embodiment, Figure 8 As shown, a device for using a large fault operation and maintenance model is provided, including: an acquisition module 810 and a second determination module 820, wherein:

[0186] The acquisition module 810 is used to acquire the operation and maintenance problem data to be tested;

[0187] The second determination module 820 is used to determine the answer strategy corresponding to the operation and maintenance problem data to be tested based on the trained fault operation and maintenance big model; wherein the fault operation and maintenance big model is trained using the above-mentioned fault operation and maintenance big model training method.

[0188] Each module in the training device of the above-mentioned fault operation and maintenance large model can be implemented in whole or in part by software, hardware and their combination. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0189] In an exemplary embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as shown in FIG. Fig. 9 As shown. The computer device includes a processor, a memory, an input / output interface, a communication interface, a display unit and an input device. The processor, the memory and the input / output interface are connected through a system bus, and the communication interface, the display unit and the input device are connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The input / output interface of the computer device is used to exchange information between the processor and the external device. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be realized through WIFI, a mobile cellular network, near field communication (NFC) or other technologies. When the computer program is executed by the processor, a training method for a large fault operation and maintenance model is realized. The display unit of the computer device is used to form a visually visible picture, which can be a display screen, a projection device or a virtual reality imaging device. The display screen can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covering the display screen, or a button, trackball or touchpad set on the computer device shell, or an external keyboard, touchpad or mouse.

[0190] Those skilled in the art will understand that Fig. 9 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0191] In an exemplary embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above-mentioned method embodiments when executing the computer program.

[0192] In some embodiments, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0193] In some embodiments, a computer program product is provided, including a computer program, which implements the steps in the above-mentioned method embodiments when executed by a processor.

[0194] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, an artificial intelligence (AI) processor, etc., but are not limited to this.

[0195] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0196] The above embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.< / response> < / question> < / prompt>

Claims

1. A method for training a large fault operation and maintenance model, characterized in that: include: Based on the pre-trained fault operation and maintenance big model, the path capture loss of the reference reasoning path related to the operation and maintenance problem sample is generated from the fault operation and maintenance knowledge graph; wherein the entities of the fault operation and maintenance knowledge graph include the problem description, problem cause and operation and maintenance strategy of the operation and maintenance problem; the edge relationship between the entities includes the causal relationship; Determine, according to the operation and maintenance problem sample, the reference reasoning path and the fault operation and maintenance knowledge graph, the retrieval reasoning loss of the operation and maintenance strategy label corresponding to the operation and maintenance problem sample generated by the pre-trained fault operation and maintenance large model; The pre-trained fault operation and maintenance large model is adjusted according to the path capture loss and / or the retrieval reasoning loss.

2. The method according to claim 1, characterized in that The path capture loss of the reference reasoning path related to the operation and maintenance problem sample is generated from the fault operation and maintenance knowledge graph based on the pre-trained fault operation and maintenance large model, including: Generate the initial reasoning path related to the operation and maintenance problem samples based on the pre-trained fault operation and maintenance model; Determine that the initial reasoning path is a first probability distribution of a reference reasoning path in the fault operation and maintenance knowledge graph; The path capture loss is determined according to the difference between the first probability distribution and the preset path distribution.

3. The method according to claim 2, characterized in that The method further comprises: At least one shortest reasoning path between the operation and maintenance problem sample and the corresponding operation and maintenance strategy label in the fault operation and maintenance knowledge graph is used as the reference reasoning path.

4. The method according to claim 2, characterized in that: The determining the path capture loss according to the difference between the first probability distribution and the preset path distribution includes: Determine, for each generated element in the initial reasoning path, a probability distribution of the first element in the reference reasoning path in the fault operation and maintenance knowledge graph; A path capture loss is determined according to the first element probability distribution and the preset path distribution.

5. The method according to any one of claims 1 to 4, characterized in that: The determining, based on the operation and maintenance problem sample, the reference reasoning path and the fault operation and maintenance knowledge graph, of the retrieval reasoning loss of the operation and maintenance strategy label corresponding to the operation and maintenance problem sample generated by the pre-trained fault operation and maintenance large model comprises: generating at least one key reasoning path according to the reference reasoning path; Based on the pre-trained fault operation and maintenance large model, a second probability distribution of the operation and maintenance strategy label is generated given the operation and maintenance problem sample, the key reasoning path and the fault operation and maintenance knowledge graph; The retrieval inference loss is generated according to the second probability distribution.

6. The method according to claim 5, characterized in that Generating at least one key reasoning path according to the reference reasoning path includes: Based on the pre-trained fault operation and maintenance big model, determine the confidence of the reference reasoning path related to the operation and maintenance problem sample, and select at least one key reasoning path from the reference reasoning path according to the confidence of the reference reasoning path; and / or, Based on the breadth-first algorithm, at least one search reasoning path between the operation and maintenance problem sample and the corresponding operation and maintenance strategy label is searched in the fault operation and maintenance knowledge graph, and a key reasoning path including the reference reasoning path and the search reasoning path is generated.

7. The method according to claim 5, characterized in that The generating a second probability distribution of the operation and maintenance strategy label based on the pre-trained fault operation and maintenance big model, given the operation and maintenance problem sample, the key reasoning path and the fault operation and maintenance knowledge graph, comprises: Input the key reasoning path and the operation and maintenance problem sample into a pre-trained fault operation and maintenance large model to generate a predicted reasoning path; Determine the probability of each generated element in the prediction reasoning path being located at the second element of the operation and maintenance strategy label; A second probability distribution corresponding to the predicted reasoning path is determined according to the second element probability.

8. A method for using a large fault operation and maintenance model, characterized in that: include: Obtain the operation and maintenance problem data to be tested; Based on the trained fault operation and maintenance big model, determine the answer strategy corresponding to the operation and maintenance problem data to be tested; Wherein, the fault operation and maintenance large model is trained by the method described in any one of claims 1-7.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images