Vehicle fault diagnosis method, device, computer equipment and storage medium

Abstract

The present application relates to a vehicle fault diagnosis method, device, computer equipment, storage medium and computer program product. The method comprises: obtaining a vehicle fault description text; inputting the vehicle fault description text into a pre-trained cross-expert fine-grained shared neural network model, and outputting a vehicle fault diagnosis result; wherein the cross-expert fine-grained shared neural network model comprises a first input processing module, an encoder module, a second input processing module, a decoder module and an output module. The use of this method can improve the efficiency of vehicle fault diagnosis and reduce the cost of vehicle fault diagnosis.

Description

Technical Field

[0001] The present application relates to the technical field of vehicle fault diagnosis, and in particular to a vehicle fault diagnosis method, apparatus, computer equipment, storage medium and computer program product. Background Art

[0002] With the rapid development of the automobile industry, vehicles have become a common means of transportation in users' daily lives. At the same time, potential safety hazards such as vehicle failures during driving have also emerged. Therefore, it is particularly important to conduct timely and effective automobile fault diagnosis to ensure driving safety.

[0003] However, existing vehicle fault diagnosis technologies have problems such as low vehicle fault diagnosis efficiency or high vehicle fault diagnosis cost due to the high complexity of vehicle systems, uneven quality of vehicle fault data, and non-uniform vehicle fault diagnosis standards. Summary of the invention

[0004] Based on this, it is necessary to provide a vehicle fault diagnosis method, device, computer equipment, storage medium and computer program product that can improve the efficiency of vehicle fault diagnosis and reduce the cost of vehicle fault diagnosis in order to address the above technical problems.

[0005] In a first aspect, a vehicle fault diagnosis method is provided, the method comprising:

[0006] Get the vehicle fault description text;

[0007] The vehicle fault description text is input into a pre-trained cross-expert fine-grained shared neural network model, and the vehicle fault diagnosis result is output; wherein the cross-expert fine-grained shared neural network model includes a first input processing module, an encoder module, a second input processing module, a decoder module and an output module; the first input processing module is used to perform data processing on the vehicle fault description text to obtain the corresponding first feature vector; the encoder module is used to perform feature extraction processing on each first feature vector to obtain the corresponding second feature vector; the second input processing module is used to perform data processing on each second feature vector to obtain the corresponding third feature vector; the decoder module is used to perform classification prediction processing on each third feature vector to obtain the corresponding fourth feature vector; the output module is used to perform feature fusion and classification probability calculation processing on each fourth feature vector to obtain the vehicle fault diagnosis result.

[0008] In one embodiment, a cross-expert fine-grained shared neural network model is constructed, including: obtaining a second preset number of historical vehicle fault description texts; randomly dividing each historical vehicle fault description text to generate a training sample set and a test sample set; training a preset neural network model based on a self-attention mechanism according to the training sample set, testing the neural network model according to the test sample set, and adjusting the model parameters of the neural network model according to the training results, the test results and the loss value calculated by the cross-expert multi-task loss function until the loss value calculated by the cross-expert multi-task loss function is less than the loss threshold, thereby generating a cross-expert fine-grained shared neural network model to output a vehicle fault diagnosis result based on the cross-expert fine-grained shared neural network model.

[0009] In one of the embodiments, determining a cross-expert multi-task loss function includes: determining a target probability distribution model; the target probability distribution model is used to characterize the probability that each expert in the forward feedback network processes the corresponding task feature; determining an interaction intensity calculation model based on a joint distribution probability algorithm and the target probability distribution model; the interaction intensity calculation model is used to calculate the interaction intensity between each expert, the interaction intensity between each expert and the corresponding task feature, and the interaction intensity between each task feature; determining a cross-expert multi-task loss function based on the interaction intensity calculation model and the crowd parameters of each expert.

[0010] In one embodiment, the expression of the cross-expert multi-task loss function is:

[0011]

[0012] Where L represents the cross-expert multi-task loss function; M represents the number of task features; w Ti Indicates the T j The weight value of the task feature in the cross-expert multi-task loss function; w MI is the weight value corresponding to the interaction strength calculation model; I(T; E) represents the interaction strength calculation model.

[0013] In one of the embodiments, a first feature vector is obtained after data processing is performed on a vehicle fault description text, including: performing word embedding processing on the vehicle fault description text to obtain a corresponding word vector; and performing regularization processing on each word vector to obtain a corresponding first feature vector.

[0014] In one embodiment, the encoder module includes a first multi-head attention layer, a first normalization layer, a first feedforward layer and a second normalization layer; each first feature vector is subjected to feature extraction processing to obtain a corresponding second feature vector, including: performing feature diversification processing on each first feature vector through the first multi-head attention layer to obtain a corresponding first target feature vector; performing layer normalization processing on each first target feature vector through the first normalization layer to obtain a corresponding second target feature vector; performing generalization capability enhancement calculation on each second target feature vector through the first feedforward layer to obtain a corresponding third target feature vector; wherein the first feedforward layer includes a feedforward network determined according to a fine-grained expert strategy and a specified shared expert strategy; performing layer normalization processing on each third target feature vector through the second normalization layer to obtain a corresponding second feature vector.

[0015] In one embodiment, performing data processing on each second eigenvector to obtain a corresponding third eigenvector includes: performing regularization processing on each second eigenvector to obtain a corresponding third eigenvector.

[0016] In one embodiment, the decoder module includes a masked multi-head attention layer, a third normalization layer, a second multi-head attention layer, a fourth normalization layer, a second feed-forward layer and a fifth normalization layer; each third eigenvector is subjected to classification prediction processing to obtain a corresponding fourth eigenvector, including: performing eigenvalue filtering processing and feature diversification processing on each third eigenvector through the masked multi-head attention layer to obtain a corresponding fourth target eigenvector; performing layer normalization processing on each fourth target eigenvector through the third normalization layer to obtain a corresponding fifth target eigenvector; performing feature diversification processing on each fifth target eigenvector through the second multi-head attention layer to obtain a corresponding sixth target eigenvector; performing layer normalization processing on each sixth target eigenvector through the fourth normalization layer to obtain a corresponding seventh target eigenvector; performing generalization capability enhancement calculation on each seventh target eigenvector through the second feed-forward layer to obtain a corresponding eighth target eigenvector; wherein the second feed-forward layer includes a feed-forward network determined according to a fine-grained expert strategy and a specified shared expert strategy; performing layer normalization processing on each eighth target eigenvector through the fifth normalization layer to obtain a corresponding fourth eigenvector.

[0017] In one embodiment, determining the forward feedback network includes: based on the fine-grained expert strategy, performing expert fine-graining processing on the initial forward feedback network to obtain the fine-grained initial forward feedback network; based on the specified shared expert strategy, performing expert shared processing on the fine-grained initial forward feedback network to obtain the forward feedback network.

[0018] In one embodiment, the expression of the forward feedback network is:

[0019]

[0020] Among them, the expression of the gate value of the i-th expert is:

[0021]

[0022] Among them, the expression of the correlation score between the expert and the corresponding feature vector is:

[0023]

[0024] Among them, K s is the number of shared experts; mN-K s It represents the number of experts after fine-graining except shared experts; N represents the number of experts before fine-graining; mN represents the number of experts after fine-graining; FFN i (·) represents the process of the i-th expert processing the feature vector; g i,t represents the gate value of the i-th expert; s i,t represents the correlation score between the expert and the corresponding feature vector; Topk(.,K) represents the set of the highest K correlation scores between the t-th feature vector and mN fine-grained experts; represents the i-th expert in the l-th layer; represents the output feature vector of the forward feedback network; Represents the input feature vector of the feedforward network.

[0025] In one embodiment, the output module includes a linear layer and a softmax layer; after performing feature fusion and classification probability calculation processing on each fourth eigenvector, a cross-expert fine-grained shared neural network sub-model is obtained, including: performing feature fusion processing on each fourth eigenvector through a linear layer to obtain a corresponding ninth target feature vector; performing classification probability calculation processing on each ninth target feature vector through a softmax layer to obtain a corresponding vehicle fault diagnosis result.

[0026] In one embodiment, the expression of the cross-expert fine-grained shared neural network model is:

[0027]

[0028] Where x represents the input of the cross-expert fine-grained shared neural network model; E k (x) represents the fault feature vector mined by the cross-expert fine-grained shared neural network model; G k (x) represents the weight value corresponding to the input of the cross-expert fine-grained shared neural network model; y represents the output of the cross-expert fine-grained shared module.

[0029] In one embodiment, obtaining a new energy vehicle fault description text includes:

[0030] Obtain unstructured vehicle production record data and unstructured vehicle fault manual data;

[0031] The unstructured vehicle production record data and unstructured vehicle fault manual data are cleaned to obtain the vehicle fault description text.

[0032] In one of the embodiments, the vehicle fault diagnosis result includes a fault keyword extraction result, a fault cause analysis result and a fault problem strategy result.

[0033] In one of the embodiments, the cross-expert fine-grained shared neural network model adopts a Transformer model.

[0034] In a second aspect, a vehicle fault diagnosis device is provided, the device comprising: a text acquisition module and a fault diagnosis module.

[0035] Wherein, the text acquisition module is used to obtain the vehicle fault description text;

[0036] The fault diagnosis module is used to input the vehicle fault description text into a pre-trained cross-expert fine-grained shared neural network model, and output the vehicle fault diagnosis result; wherein the cross-expert fine-grained shared neural network model includes a first input processing module, an encoder module, a second input processing module, a decoder module and an output module; the first input processing module is used to perform data processing on the vehicle fault description text to obtain the corresponding first feature vector; the encoder module is used to perform feature extraction processing on each first feature vector to obtain the corresponding second feature vector; the second input processing module is used to perform data processing on each second feature vector to obtain the corresponding third feature vector; the decoder module is used to perform classification prediction processing on each third feature vector to obtain the corresponding fourth feature vector; the output module is used to perform feature fusion and classification probability calculation processing on each fourth feature vector to obtain the vehicle fault diagnosis result.

[0037] In a third aspect, a computer device is provided, which includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of any method in the above method embodiments are implemented.

[0038] In a fourth aspect, 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 of any method in the above method embodiments are implemented.

[0039] In a fifth aspect, a computer program product is provided, which includes a computer program, and when the computer program is executed by a processor, the steps of any method in the above method embodiments are implemented.

[0040] The above-mentioned vehicle fault diagnosis method, device, computer equipment, storage medium and computer program product obtain a vehicle fault description text; then, the vehicle fault description text is input into a pre-trained cross-expert fine-grained shared neural network model, and a vehicle fault diagnosis result is output; wherein the cross-expert fine-grained shared neural network model includes a first input processing module, an encoder module, a second input processing module, a decoder module and an output module; the first input processing module is used to perform data processing on the vehicle fault description text to obtain a corresponding first feature vector; the encoder module is used to perform feature extraction processing on each first feature vector to obtain a corresponding second feature vector; the second input processing module is used to perform data processing on each second feature vector to obtain a corresponding third feature vector; the decoder module is used to perform classification prediction processing on each third feature vector to obtain a corresponding fourth feature vector; the output module is used to perform feature fusion and classification probability calculation processing on each fourth feature vector to obtain a vehicle fault diagnosis result, thereby avoiding vehicle fault diagnosis through manual experience due to the high complexity of the vehicle system, uneven quality of vehicle fault data and inconsistent vehicle fault diagnosis standards, improving the efficiency and accuracy of vehicle fault diagnosis and reducing the cost of vehicle fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 A diagram showing an application environment of a vehicle fault diagnosis method in an embodiment;

[0042] Figure 2 A first flow chart of a vehicle fault diagnosis method in one embodiment;

[0043] Figure 3 A schematic diagram of a process for obtaining a fault description text of a new energy vehicle in one embodiment;

[0044] Figure 4 A structural block diagram of a cross-expert fine-grained shared neural network model in one embodiment;

[0045] Figure 5 It is a schematic diagram of a flow chart of obtaining a first feature vector after data processing of a vehicle fault description text in one embodiment;

[0046] Figure 6 It is a schematic diagram of a process of obtaining a corresponding second feature vector after performing feature extraction processing on each first feature vector in an embodiment;

[0047] Figure 7 A schematic diagram of a flow chart for determining a forward feedback network in one embodiment;

[0048] Figure 8 is a schematic diagram of an initial forward feedback network in one embodiment;

[0049] Fig. 9 is a schematic diagram of an initial forward feedback network after fine-graining processing in one embodiment;

[0050] Fig.10 is a schematic diagram of a forward feedback network in one embodiment;

[0051] Fig.11 It is a schematic diagram of a process of obtaining a corresponding fourth eigenvector after performing classification prediction processing on each third eigenvector in an embodiment;

[0052] Fig.12 It is a flowchart of obtaining a cross-expert fine-grained shared neural network sub-model after performing feature fusion and classification probability calculation processing on each fourth feature vector in one embodiment;

[0053] Fig.13 A schematic diagram of a process for building a cross-expert fine-grained shared neural network model in one embodiment;

[0054] Fig.14 A schematic diagram of a process for determining a cross-expert multi-task loss function in one embodiment;

[0055] Fig.15 is a structural block diagram of a vehicle fault diagnosis device in one embodiment;

[0056] Fig.16 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0057] 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.

[0058] In order to facilitate understanding of the present application, the present application will be described more fully below with reference to the relevant drawings. Embodiments of the present application are provided in the drawings. However, the present application can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present application more thorough and comprehensive.

[0059] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein in the specification of this application are only for the purpose of describing specific embodiments and are not intended to limit this application.

[0060] It is understood that the terms "first", "second", etc. used in this application may be used herein to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish a first element from another element. For example, without departing from the scope of this application, a first resistor may be referred to as a second resistor, and similarly, a second resistor may be referred to as a first resistor. Both the first resistor and the second resistor are resistors, but they are not the same resistor.

[0061] It can be understood that the “connection” in the following embodiments should be understood as “electrical connection”, “communication connection”, etc. if the connected circuits, modules, units, etc. have electrical signals or data transmission between each other.

[0062] When used herein, the singular forms "a", "an", and "said / the" may also include plural forms, unless the context clearly indicates otherwise. It should also be understood that the terms "include / comprise" or "have" etc. specify the presence of stated features, wholes, steps, operations, components, parts or combinations thereof, but do not exclude the possibility of the presence or addition of one or more other features, wholes, steps, operations, components, parts or combinations thereof.

[0063] The vehicle fault diagnosis method provided by this application can be applied to Figure 1 In the application environment shown, the terminal 102 communicates with the server 104 through the network. The terminal 102 can be, but is not limited to, various vehicle-mounted terminals, and the server 104 can be implemented by an independent server or a server cluster composed of multiple servers.

[0064] In a specific example, the vehicle using the above vehicle fault diagnosis method can be but is not limited to a new energy vehicle. The above is only a specific example. In actual application, it can be flexibly set according to needs and is not limited here.

[0065] First, as Figure 2 As shown, a vehicle fault diagnosis method is provided, which is applied to Figure 1 The server 104 in the example is used for explanation, and the steps include the following steps 201 and 202.

[0066] Step 201, obtaining vehicle fault description text.

[0067] The vehicle fault description text is used to represent the text describing the vehicle fault in natural language. The server 104 can obtain the vehicle fault description text.

[0068] In one embodiment, if Figure 3As shown, obtaining the new energy vehicle fault description text includes step 301 and step 302.

[0069] Step 301, obtaining unstructured vehicle production record data and unstructured vehicle fault manual data.

[0070] Step 302 , performing data cleaning processing on the unstructured vehicle production record data and the unstructured vehicle fault manual data to obtain a vehicle fault description text.

[0071] Specifically, the server 104 may obtain unstructured vehicle production record data and unstructured vehicle fault manual data; then, perform data cleaning on the unstructured vehicle production record data and the unstructured vehicle fault manual data to obtain a vehicle fault description text.

[0072] In this embodiment, unstructured vehicle production record data and unstructured vehicle fault manual data are obtained; then, the unstructured vehicle production record data and the unstructured vehicle fault manual data are cleansed to obtain a vehicle fault description text, thereby removing errors and noises in the unstructured vehicle production record data and the unstructured vehicle fault manual data, ensuring the accuracy and consistency of the unstructured vehicle production record data and the unstructured vehicle fault manual data, reducing a large amount of redundant and irrelevant information, improving the quality of the unstructured vehicle production record data and the unstructured vehicle fault manual data, improving the convenience and accuracy of obtaining new energy vehicle fault description texts, and improving the accuracy of vehicle fault diagnosis at the source.

[0073] Step 202 , input the vehicle fault description text into a pre-trained cross-expert fine-grained shared neural network model, and output the vehicle fault diagnosis result.

[0074] Among them, Figure 4 As shown, the cross-expert fine-grained shared neural network model 4000 includes a first input processing module 4100, an encoder module 4200, a second input processing module 4300, a decoder module 4400 and an output module 4500; the first input processing module 4100 is used to perform data processing on the vehicle fault description text to obtain the corresponding first feature vector; the encoder module 4200 is used to perform feature extraction processing on each first feature vector to obtain the corresponding second feature vector; the second input processing module 4300 is used to perform data processing on each second feature vector to obtain the corresponding third feature vector; the decoder module 4400 is used to perform classification prediction processing on each third feature vector to obtain the corresponding fourth feature vector; the output module 4500 is used to perform feature fusion and classification probability calculation processing on each fourth feature vector to obtain the vehicle fault diagnosis result.

[0075] Specifically, the server 104 inputs the vehicle fault description text into the pre-trained cross-expert fine-grained shared neural network model 4000, and then outputs the vehicle fault diagnosis result. It can be understood that the server 104 obtains the corresponding first feature vector after data processing the vehicle fault description text through the first input processing module 4100 of the cross-expert fine-grained shared neural network model; then, the encoder module 4200 performs feature extraction processing on each first feature vector to obtain the corresponding second feature vector; then, the second input processing module 4300 performs data processing on each second feature vector to obtain the corresponding third feature vector; then, the decoder module 4400 performs classification prediction processing on each third feature vector to obtain the corresponding fourth feature vector; finally, the output module 4500 is used to perform feature fusion and classification probability calculation processing on each fourth feature vector to obtain the vehicle fault diagnosis result, thereby avoiding the vehicle fault diagnosis through manual experience due to the high complexity of the vehicle system, the uneven quality of vehicle fault data and the non-uniform vehicle fault diagnosis standards, improving the efficiency and accuracy of vehicle fault diagnosis, and reducing the cost of vehicle fault diagnosis.

[0076] In one of the embodiments, the vehicle fault diagnosis result includes fault keyword extraction results, fault cause analysis results and fault problem strategy results, which improves the comprehensiveness of the vehicle fault diagnosis result.

[0077] In one embodiment, the cross-expert fine-grained shared neural network model adopts the Transformer model. It can be understood that the Transformer model is an efficient sequence processing tool, which has a wide range of applications in the field of natural language processing, and provides better performance and flexibility than the traditional RNN model through its self-attention mechanism and structural design.

[0078] In one embodiment, if Figure 5 As shown, after data processing is performed on the vehicle fault description text to obtain the first feature vector, the process includes step 501 and step 502.

[0079] Step 501, perform word embedding processing on the vehicle fault description text to obtain the corresponding word vector;

[0080] Step 502: perform regularization processing on each word vector to obtain a corresponding first feature vector.

[0081] Among them, the server 104 performs word embedding processing on the vehicle fault description text to obtain the corresponding word vector, so that the words described in natural language in the vehicle fault description text are converted into a fixed-dimensional word vector representation, which plays a role in dimensionality reduction; then, each word vector is regularized to obtain the corresponding first eigenvector, thereby improving the convergence speed and calculation speed of the algorithm and preventing overfitting.

[0082] In one embodiment, if Figure 4 and 6 As shown, the encoder module 4200 includes a first multi-head attention layer 4210, a first normalization layer 4220, a first pre-feedback layer 4230 and a second normalization layer 4240. Among them, after performing feature extraction processing on each first feature vector to obtain the corresponding second feature vector, including steps 601 to 604.

[0083] Step 601, perform feature diversification processing on each first feature vector through a first multi-head attention layer to obtain a corresponding first target feature vector.

[0084] Step 602 , performing layer normalization processing on each first target feature vector through a first normalization layer to obtain a corresponding second target feature vector.

[0085] Step 603: Perform a generalization capability enhancement calculation on each second target feature vector through the first pre-feedback layer to obtain a corresponding third target feature vector.

[0086] Step 604: perform layer normalization processing on each third target feature vector through a second normalization layer to obtain a corresponding second feature vector.

[0087] Among them, the first pre-feedback layer 4230 includes a feed-forward network determined according to the fine-grained expert strategy and the specified sharing expert strategy. It can be understood that the fine-grained expert strategy is a strategy for adjusting the expert fine-grainedness of the feed-forward network of the feed-forward layer in the cross-expert fine-grained shared neural network model. The specified sharing expert strategy is a strategy for adjusting the expert specified sharing degree of the feed-forward network of the feed-forward layer in the expert fine-grained shared neural network model. In addition, the fine-grained expert strategy and the specified sharing expert strategy can be flexibly set according to user needs and are not limited here. The pre-feedback layer includes the first pre-feedback layer 4230 in the encoder module 4200 and the second pre-feedback layer 4450 in the decoder module 4400.

[0088] Specifically, the server 104 performs feature diversification processing on each first eigenvector through the first multi-head attention layer 4210 to obtain a corresponding first target eigenvector, so that the model can pay attention to information of different dimensions, thereby improving the expressiveness and attention allocation of the model; then, the first normalization layer 4220 performs layer normalization processing on each first target eigenvector to obtain a corresponding second target eigenvector, thereby improving the generalization ability, improving gradient propagation, eliminating internal covariate shift, and improving the training effect and performance of the model; then, the first pre-feedback layer 4230 performs generalization ability enhancement calculation on each second target eigenvector to obtain a corresponding third target eigenvector, thereby enhancing the generalization ability and facilitating the capture of higher-level features; and, the second normalization layer 4240 performs layer normalization processing on each third target eigenvector to obtain a corresponding second eigenvector, thereby improving the generalization ability, improving gradient propagation, eliminating internal covariate shift, and further improving the training effect and performance of the model.

[0089] In one embodiment, if Figure 7 As shown, determining a forward feedback network includes step 701 and step 702.

[0090] Step 701, based on the fine-grained expert strategy, the initial forward feedback network is subjected to expert fine-grained processing to obtain the fine-grained initial forward feedback network.

[0091] Step 702: Based on the specified shared expert strategy, the initial forward feedback network after fine-graining processing is subjected to expert sharing processing to obtain a forward feedback network.

[0092] Among them, based on the fine-grained expert strategy, the server 104 performs expert fine-graining processing on the initial forward feedback network to obtain the fine-grained initial forward feedback network, thereby improving the fine-grainedness of each expert in the initial forward feedback network, reducing the computational amount of the model, improving the model computational efficiency, and enhancing the collaborative flexibility and adaptability of each expert in the initial forward feedback network; then, based on the specified shared expert strategy, the fine-grained initial forward feedback network is subjected to expert sharing processing to obtain the forward feedback network, ensuring that certain knowledge can be shared between the experts in the fine-grained initial forward feedback network, so as to make the model converge faster, reduce the model computational amount, and alleviate parameter redundancy. In addition, the forward feedback network includes the forward feedback network of the first forward feedback layer 4230 in the encoder module 4200 and the forward feedback network of the second forward feedback layer 4450 in the decoder module 4400.

[0093] In a specific example, the expressions of the first multi-head attention layer 4210, the masked multi-head attention layer, and the second multi-head attention layer are:

[0094]

[0095] Where T represents the sequence length of the input feature vector of the multi-head attention layer, Att(·) is the attention mechanism, Represents the input feature vector of the multi-head attention layer, that is, the t-th feature vector after the l-1-th layer of attention, that is, the hidden state of the t-th token after the l-1-th layer of attention; The output feature vector representing the input feature vector of the multi-head attention layer, that is, all feature vectors after the l-th layer of attention, that is, the hidden state of all tokens after the l-th layer of attention. The input feature vector of the first multi-head attention layer 4210 is the first feature vector, and the input feature vector of the masked multi-head attention layer 4410 is the third feature vector; the input feature vector of the second multi-head attention layer 4430 is the fifth target feature vector; the output feature vector of the first multi-head attention layer 4210 is the first target feature vector, and the input feature vector of the masked multi-head attention layer 3310 is the fourth target feature vector; the input feature vector of the second multi-head attention layer 4430 is the sixth target feature vector. The above are only specific examples. In actual applications, they can be flexibly set according to needs and are not limited here.

[0096] In a specific example, the expression of the initial forward feedback network is:

[0097]

[0098] Among them, FFN(·) is an expert, Represents the output feature vector of the initial forward feedback network, that is, all feature vectors after the l-th layer of forward feedback, and also the hidden state of all tokens after the l-th layer of forward feedback. It represents the input feature vector of the initial forward feedback network, that is, it represents the t-th feature vector after the l-th layer of forward feedback, that is, the hidden state of the t-th token after the l-th layer of forward feedback.

[0099] Specifically, the cross-expert fine-grained shared neural network model replaces all FFN(·) in the initial forward feedback network with N FFN(·) with smaller dimensions, where each small FFN(·) represents an expert. Then, the fine-grained expert strategy is used to perform expert fine-graining on the initial forward feedback network to obtain the fine-grained initial forward feedback network, thereby improving the specificity of the experts in the model. When the number of experts is limited, each expert will learn a large amount of different knowledge, resulting in excessive model computation and ineffective use of knowledge. However, if each feature vector is assigned to multiple experts, the diverse knowledge will be decomposed and easy to learn. Therefore, by performing expert fine-graining on the initial forward feedback network, the experts are divided into finer granularities, enhancing the flexibility and adaptability of expert collaboration, thereby keeping the expert parameters and model computation consistent.

[0100] like Figure 8 and 9 As shown in the figure, the present invention reduces the dimension of the expert FFN(·) to the original This will result in mN experts with smaller granularity, and improve the model's ability to learn diverse knowledge.

[0101] Specifically, the expression of the initial forward feedback network after fine-graining is:

[0102]

[0103] Among them, the expression of the gate value of the i-th expert is:

[0104]

[0105] Among them, the expression of the correlation score between the expert and the corresponding feature vector is:

[0106]

[0107] Where N is the number of experts before fine-graining, and mN is the number of experts after fine-graining. If m = 2, because Figure 8 Only the top K=2 non-fine-grained experts with the highest feature vectors among mN fine-grained experts are selected, so Fig. 9 The top mK=4 fine-grained experts should be selected. i (·) represents the process of the i-th expert FFN (·) processing the feature vector; g i,t represents the gate value of the i-th expert, s i,t represents the correlation score between the expert and the corresponding feature vector; Topk(.,K) represents the set of the highest K correlation scores between the t-th feature vector and mN fine-grained experts; mK represents the experts with the top mK correlation scores among the mN fine-grained experts; represents the i-th expert in the l-th layer; It represents the output feature vector of the initial forward feedback network after fine-grained processing, all feature vectors after forward feedback, and also the hidden state of all tokens after the l-th layer of forward feedback. It represents the input feature vector of the initial forward feedback network after fine-graining processing, that is, it represents the t-th feature vector after the l-th layer of forward feedback, that is, the hidden state of the t-th token after the l-th layer of forward feedback. However, when each feature vector is assigned to multiple experts, experts need to share certain knowledge to ensure that the model converges faster, but this will also cause redundancy in the model parameters, so it is necessary to perform expert sharing processing on the initial forward feedback network after fine-graining processing based on the specified shared expert strategy to obtain the forward feedback network.

[0108] Specifically, the expression of the forward feedback network is:

[0109]

[0110] Among them, the expression of the gate value of the i-th expert is:

[0111]

[0112] Among them, the correlation score between the expert and the corresponding feature vector is:

[0113]

[0114] Among them, K s is the number of shared experts, mN-K s It is represented by the number of experts after fine-graining processing other than shared experts, N is the number of experts before fine-graining processing, mN is the number of experts after fine-graining processing, FFN i (·) represents the process of the i-th expert FFN (·) processing the feature vector; g i,t represents the gate value of the i-th expert, s i,t represents the correlation score between the expert and the corresponding feature vector; Topk(.,K) represents the set of the highest K correlation scores between the t-th feature vector and mN fine-grained experts; represents the i-th expert in the l-th layer; Represents the output feature vector of the forward feedback network, all feature vectors after the forward feedback, and also the hidden state of all tokens after the l-th layer of forward feedback. represents the input feature vector of the forward feedback network, that is, the t-th feature vector after the l-th layer of forward feedback, that is, the hidden state of the t-th token after the l-th layer of forward feedback. Fig.10 As shown, K s=1, mN-K s The above is only a specific example, and it can be flexibly set according to the actual application, and is not limited here.

[0115] In this embodiment, based on the fine-grained expert strategy, the initial feedforward network is subjected to expert fine-graining processing to obtain the fine-grained initial feedforward network, thereby improving the fine-graining of each expert in the initial feedforward network, reducing the computational complexity of the model, improving the model computational efficiency, and enhancing the collaborative flexibility and adaptability of each expert in the initial feedforward network; then, based on the specified shared expert strategy, the fine-grained initial feedforward network is subjected to expert sharing processing to obtain the feedforward network, ensuring that certain knowledge can be shared between the experts in the fine-grained initial feedforward network, so as to make the model converge faster, reduce the computational complexity of the model, and alleviate parameter redundancy.

[0116] In this embodiment, the first multi-head attention layer 4210 performs feature diversification processing on each first eigenvector to obtain a corresponding first target feature vector; then, the first normalization layer 4220 performs layer normalization processing on each first target feature vector to obtain a corresponding second target feature vector; then, the first pre-feedback layer 4230 performs generalization capability enhancement calculation on each second target feature vector to obtain a corresponding third target feature vector; and, the second normalization layer 4240 performs layer normalization processing on each third target feature vector to obtain a corresponding second feature vector, so as to facilitate the feature extraction processing of the first eigenvector by the encoder module 4200 of the cross-expert fine-grained shared neural network model 4000, thereby improving the training effect and performance of the model.

[0117] In one embodiment, the data processing of each second eigenvector is performed to obtain the corresponding third eigenvector, including:

[0118] Regularization is performed on each second eigenvector to obtain a corresponding third eigenvector.

[0119] Specifically, the server 104 performs regularization processing on each second eigenvector to obtain a corresponding third eigenvector, thereby improving the convergence speed and calculation speed of the algorithm and preventing overfitting.

[0120] In this embodiment, regularization is performed on each second eigenvector to obtain a corresponding third eigenvector, thereby improving the convergence speed and calculation speed of the algorithm and preventing overfitting.

[0121] In one embodiment, if Figure 4 and 11As shown, the decoder module 4400 includes a masked multi-head attention layer 4410, a third normalization layer 4420, a second multi-head attention layer 4430, a fourth normalization layer 4440, a second pre-feedback layer 4450, and a fifth normalization layer 4460. After performing classification prediction processing on each third eigenvector, the corresponding fourth eigenvector is obtained, including steps 1101 to 1106.

[0122] Step 1101, performing eigenvalue filtering and feature diversification processing on each third eigenvector through a masked multi-head attention layer to obtain a corresponding fourth target eigenvector;

[0123] Step 1102, performing layer normalization processing on each fourth target feature vector through a third normalization layer to obtain a corresponding fifth target feature vector;

[0124] Step 1103, performing feature diversification processing on each fifth target feature vector through a second multi-head attention layer to obtain a corresponding sixth target feature vector;

[0125] Step 1104, performing layer normalization processing on each sixth target feature vector through a fourth normalization layer to obtain a corresponding seventh target feature vector;

[0126] Step 1105, performing a generalization capability enhancement calculation on each seventh target feature vector through the second front feedback layer to obtain a corresponding eighth target feature vector;

[0127] Step 1106 , performing layer normalization processing on each eighth target eigenvector through a fifth normalization layer to obtain a corresponding fourth eigenvector.

[0128] Among them, the second forward feedback layer includes a forward feedback network determined according to the fine-grained expert strategy and the specified sharing expert strategy; it can be understood that the fine-grained expert strategy is a strategy for adjusting the expert fine-grainedness of the forward feedback network of the forward feedback layer in the cross-expert fine-grained shared neural network model. The specified sharing expert strategy is a strategy for adjusting the expert specified sharing degree of the forward feedback network of the forward feedback layer in the expert fine-grained shared neural network model. In addition, the fine-grained expert strategy and the specified sharing expert strategy can be flexibly set according to user needs and are not limited here. The forward feedback layer includes a first forward feedback layer 4230 in the encoder module 4200 and a second forward feedback layer 4450 in the decoder module 4400.

[0129] Specifically, the server 104 performs eigenvalue filtering and feature diversification processing on each third eigenvector through a masked multi-head attention layer 4410 to obtain a corresponding fourth target eigenvector, so that the model completes the masking operation and pays attention to information of different dimensions, thereby improving the expressiveness and attention allocation of the model; then, the server 104 performs layer normalization processing on each fourth target eigenvector through a third normalization layer 4420 to obtain a corresponding fifth target eigenvector, thereby improving the generalization ability, improving the gradient propagation, eliminating the internal covariate shift, and improving the training effect and performance of the model; then, the server 104 performs feature diversification processing on each fifth target eigenvector through a second multi-head attention layer 4430 to obtain a corresponding sixth target eigenvector, so that the model can pay attention to information of different dimensions, thereby improving the expressiveness and attention allocation of the model; The expressive power and attention allocation of the model are improved; then, each sixth target feature vector is layer-normalized through the fourth normalization layer 4440 to obtain the corresponding seventh target feature vector, thereby improving the generalization ability, improving the gradient propagation, eliminating the internal covariate shift, and improving the training effect and performance of the model; and, the generalization ability enhancement calculation is performed on each seventh target feature vector through the second pre-feedback layer 4450 to obtain the corresponding eighth target feature vector, thereby enhancing the generalization ability and facilitating the capture of more advanced features; finally, each eighth target feature vector is layer-normalized through the fifth normalization layer to obtain the corresponding fourth feature vector, thereby further improving the generalization ability, improving the gradient propagation, eliminating the internal covariate shift, and improving the training effect and performance of the model.

[0130] In this embodiment, the masked multi-head attention layer 4410 performs eigenvalue filtering and feature diversification processing on each third eigenvector to obtain a corresponding fourth target eigenvector, so that the model completes the masking operation and pays attention to information of different dimensions, thereby improving the model's expression ability and attention allocation; then, the third normalization layer 4420 performs layer normalization processing on each fourth target eigenvector to obtain a corresponding fifth target eigenvector, thereby improving the generalization ability, improving the gradient propagation, eliminating the internal covariate shift, and improving the training effect and performance of the model; then, the second multi-head attention layer 4430 performs feature diversification processing on each fifth target eigenvector to obtain a corresponding sixth target eigenvector, so that the model can pay attention to information of different dimensions, thereby improving the model's expression ability and attention allocation; then, the fourth normalization layer 4420 performs layer normalization processing on each fourth target eigenvector to obtain a corresponding sixth target eigenvector, thereby improving the model's expression ability and attention allocation. 440 performs layer normalization processing on each sixth target feature vector to obtain the corresponding seventh target feature vector, thereby improving the generalization ability, improving gradient propagation, eliminating internal covariate shift, and improving the training effect and performance of the model; and, through the second pre-feedback layer 4450, generalization ability enhancement calculation is performed on each seventh target feature vector to obtain the corresponding eighth target feature vector, thereby enhancing the generalization ability and facilitating the capture of more advanced features; finally, layer normalization processing is performed on each eighth target feature vector through the fifth normalization layer to obtain the corresponding fourth feature vector, thereby further improving the generalization ability, improving gradient propagation, eliminating internal covariate shift, and improving the training effect and performance of the model, so that the decoder module 4400 obtains the corresponding fourth feature vector after classification prediction processing on each third feature vector, thereby improving the training effect and performance of the model.

[0131] In one embodiment, if Figure 4 and 12 As shown, the output module 4500 includes a linear layer 4510 and a softmax layer 4520. After performing feature fusion and classification probability calculation processing on each fourth feature vector, a cross-expert fine-grained shared neural network sub-model is obtained, including step 1201 and step 1202.

[0132] Step 1201, performing feature fusion processing on each fourth eigenvector through a linear layer to obtain a corresponding ninth target eigenvector.

[0133] Step 1202, performing classification probability calculation processing on each ninth target feature vector through a softmax layer to obtain a corresponding vehicle fault diagnosis result.

[0134] Specifically, the server 104 performs feature fusion processing on each fourth eigenvector through the linear layer 4510 to obtain the corresponding ninth target feature vector, thereby realizing dimensional conversion; then, the softmax layer 4520 performs classification probability calculation processing on each ninth target feature vector to obtain the corresponding vehicle fault diagnosis result, thereby avoiding vehicle fault diagnosis through manual experience due to the high complexity of the vehicle system, the uneven quality of vehicle fault data and the inconsistent vehicle fault diagnosis standards, thereby improving the efficiency and accuracy of vehicle fault diagnosis and reducing the cost of vehicle fault diagnosis.

[0135] In this embodiment, feature fusion processing is performed on each fourth eigenvector through the linear layer 4510 to obtain the corresponding ninth target feature vector; then, classification probability calculation processing is performed on each ninth target feature vector through the softmax layer 4520 to obtain the corresponding vehicle fault diagnosis result, thereby avoiding vehicle fault diagnosis through manual experience due to the high complexity of the vehicle system, uneven quality of vehicle fault data and inconsistent vehicle fault diagnosis standards, improving the efficiency and accuracy of vehicle fault diagnosis, reducing the cost of vehicle fault diagnosis, and facilitating the output module 4500 to perform feature fusion and classification probability calculation processing on each fourth eigenvector to obtain a cross-expert fine-grained shared neural network sub-model.

[0136] In a specific example, the expression of the cross-expert fine-grained shared neural network model is:

[0137]

[0138] Among them, x represents the input of the cross-expert fine-grained shared neural network model, that is, the expert input, which is the vehicle fault description text; E k (x) represents the fault feature vector mined by the cross-expert fine-grained shared neural network model, that is, the corresponding fault feature vector mined by each expert; G k (x) represents the weight value corresponding to the input of the cross-expert fine-grained shared neural network model; y represents the output of the cross-expert fine-grained shared module, that is, the vehicle fault diagnosis result.

[0139] Specifically, the expression of the weight value corresponding to the input of the cross-expert fine-grained shared neural network model is:

[0140] G(x)=TopK(Softmax(xW g +Ν(0,1)Softplus(xW noise )))

[0141] Among them, G(x) represents the weight value corresponding to the input of the cross-expert fine-grained shared neural network model. G(x) is a parameter with W g and W noiseTop-K network; Topk(.,K) represents the set of the highest K correlation scores between the t-th feature vector and mN fine-grained experts, where all elements of the feature vector except the largest K values ​​are set to 0, Softmax(·) represents the Softmax function, and Softplus(·) represents the smoothed estimate of the Relu function; W g represents the path weight; W noise represents the noise weight matrix;

[0142] Specifically, Softplus(x)=log(1+exp(x))

[0143] In order to make each expert aware of the target task, a learnable embedding vector e is also added to the cross-expert fine-grained sharing module. i To the hidden layer, the expression of the cross-expert fine-grained shared neural network model can also be:

[0144]

[0145] Among them, i is the identity feature of the task, e i represents the target task embedding vector, N represents the total number of experts, and k represents the kth expert. The above are only specific examples. In actual applications, they can be flexibly set according to needs and are not limited here.

[0146] In the above-mentioned vehicle fault diagnosis method, a vehicle fault description text is obtained; then, the vehicle fault description text is input into a pre-trained cross-expert fine-grained shared neural network model, and a vehicle fault diagnosis result is output; wherein the cross-expert fine-grained shared neural network model includes a first input processing module, an encoder module, a second input processing module, a decoder module and an output module; the first input processing module is used to perform data processing on the vehicle fault description text to obtain a corresponding first feature vector; the encoder module is used to perform feature extraction processing on each first feature vector to obtain a corresponding second feature vector; the second input processing module is used to perform data processing on each second feature vector to obtain a corresponding third feature vector; the decoder module is used to perform classification prediction processing on each third feature vector to obtain a corresponding fourth feature vector; the output module is used to perform feature fusion and classification probability calculation processing on each fourth feature vector to obtain a vehicle fault diagnosis result, thereby avoiding vehicle fault diagnosis through manual experience due to the high complexity of the vehicle system, uneven quality of vehicle fault data and inconsistent vehicle fault diagnosis standards, thereby improving the efficiency and accuracy of vehicle fault diagnosis and reducing the cost of vehicle fault diagnosis.

[0147] In a specific example, the vehicle fault description text is: "The vehicle's mileage after a single charge is significantly reduced. Compared with a new car or when the battery is in good condition, it needs to be charged more frequently under the same conditions, affecting daily use and driving plans." After inputting the above vehicle fault description text into a pre-trained cross-expert fine-grained shared neural network model, the output vehicle fault diagnosis results include fault keyword extraction results, fault cause analysis results, and fault problem strategy results. Among them, the fault keyword extraction results are reduced mileage and frequent charging; the fault cause analysis results are bad driving habits and reduced battery performance; the fault problem strategy results are to optimize the owner's driving habits and regularly inspect and maintain the vehicle. The above are only specific examples. In actual applications, they can be flexibly set according to needs and are not limited here.

[0148] In one embodiment, if Fig.13 As shown, constructing a cross-expert fine-grained shared neural network model includes steps 1301 to 1302.

[0149] Step 1301, obtaining a second preset number of historical vehicle fault description texts;

[0150] Step 1302, randomly divide each historical vehicle fault description text into a training sample set and a test sample set;

[0151] Step 1303: train the preset neural network model based on the self-attention mechanism according to the training sample set, test the neural network model according to the test sample set, and adjust the model parameters of the neural network model according to the training results, the test results and the loss value calculated by the cross-expert multi-task loss function until the loss value calculated by the cross-expert multi-task loss function is less than the loss threshold, and generate a cross-expert fine-grained shared neural network model to output the vehicle fault diagnosis result based on the cross-expert fine-grained shared neural network model.

[0152] Specifically, the server 104 obtains a second preset number of historical vehicle fault description texts; then, each historical vehicle fault description text is randomly divided to generate a training sample set and a test sample set; then, a preset neural network model based on the self-attention mechanism is trained according to the training sample set, and the neural network model is tested according to the test sample set, and the model parameters of the neural network model are adjusted according to the training results, the test results and the loss value calculated by the cross-expert multi-task loss function, until the loss value calculated by the cross-expert multi-task loss function is less than the loss threshold, and a cross-expert fine-grained shared neural network model is generated to output the vehicle fault diagnosis result based on the cross-expert fine-grained shared neural network model, thereby realizing the rapid construction of a cross-expert fine-grained shared neural network model, thereby avoiding vehicle fault diagnosis through manual experience due to the high complexity of the vehicle system, the uneven quality of vehicle fault data and the inconsistent vehicle fault diagnosis standards, thereby improving the efficiency and accuracy of vehicle fault diagnosis and reducing the cost of vehicle fault diagnosis. In addition, the cross-expert multi-task loss function can ensure that processing multiple tasks is an effective interaction between different experts, so that the task gradient direction and convergence speed remain consistent, improve the model convergence speed and accuracy, and reduce the number of errors in the cross-expert fine-grained shared neural network model.

[0153] In this embodiment, a second preset number of historical vehicle fault description texts are obtained; then, each historical vehicle fault description text is randomly divided to generate a training sample set and a test sample set; then, a preset neural network model based on a self-attention mechanism is trained according to the training sample set, and the neural network model is tested according to the test sample set, and the model parameters of the neural network model are adjusted according to the training results, the test results and the loss value calculated by the cross-expert multi-task loss function, until the loss value calculated by the cross-expert multi-task loss function is less than the loss threshold, and a cross-expert fine-grained shared neural network model is generated to output the vehicle fault diagnosis result based on the cross-expert fine-grained shared neural network model, thereby realizing the rapid construction of a cross-expert fine-grained shared neural network model, thereby avoiding the vehicle fault diagnosis through manual experience due to the high complexity of the vehicle system, the uneven quality of vehicle fault data and the non-uniform vehicle fault diagnosis standards, improving the efficiency and accuracy of vehicle fault diagnosis, and reducing the cost of vehicle fault diagnosis. In addition, the cross-expert multi-task loss function can ensure that processing multiple tasks is an effective interaction between different experts, so that the task gradient direction and convergence speed are consistent, the model convergence speed and accuracy are improved, and the number of errors in the cross-expert fine-grained shared neural network model is reduced.

[0154] In one embodiment, if Fig.14 As shown, determining the cross-expert multi-task loss function includes steps 1401 to 1403.

[0155] Step 1401, determine the target probability distribution model.

[0156] Step 1402: Determine an interaction strength calculation model according to the joint distribution probability algorithm and the target probability distribution model.

[0157] Step 1403 , determining a cross-expert multi-task loss function according to the interaction intensity calculation model and the crowd parameters of each expert.

[0158] Among them, the target probability distribution model is used to characterize the probability that each expert in the forward feedback network processes the corresponding task features; the interaction intensity calculation model is used to calculate the interaction intensity between experts, the interaction intensity between each expert and the corresponding task features, and the interaction intensity between task features.

[0159] Specifically, server 104 determines the target probability distribution model; then, determines the interaction intensity calculation model based on the joint distribution probability algorithm and the target probability distribution model; finally, determines the cross-expert multi-task loss function based on the interaction intensity calculation model and the crowd parameters of each expert. The cross-expert multi-task loss function can ensure that processing multiple tasks is an effective interaction between different experts, so that the task gradient direction and convergence speed are consistent, thereby improving the model convergence speed and accuracy, and reducing the number of errors in the cross-expert fine-grained shared neural network model.

[0160] In this embodiment, the target probability distribution model is determined; then, the interaction intensity calculation model is determined according to the joint distribution probability algorithm and the target probability distribution model; finally, the cross-expert multi-task loss function is determined according to the interaction intensity calculation model and the crowd parameters of each expert. The cross-expert multi-task loss function can ensure that processing multiple tasks is an effective interaction between different experts, so that the task gradient direction and convergence speed are consistent, thereby improving the model convergence speed and accuracy, and reducing the number of errors in the cross-expert fine-grained shared neural network model.

[0161] In a specific example, the expression of the target probability distribution model is:

[0162]

[0163] Among them, P(E i |T j ) represents the target probability distribution model; Represents each expert E in the forward feedback network i Process the corresponding task characteristics T j The probability of Ti Represents the task characteristics T j The number of corresponding vehicle fault description texts.

[0164] Specifically, Expression 1 of the joint distribution probability algorithm is:

[0165] P(E,T)=P(E|T)P(T)

[0166] It can be changed into Expression 2 of the joint distribution probability algorithm. Expression 2 of the joint distribution probability algorithm is:

[0167] P(E)=∑ T P(E,T)

[0168] Taking into account that experts should have a certain degree of expertise in a designated task, an interaction strength calculation model is determined based on the joint distribution probability algorithm and the target probability distribution model, and the interaction strength calculation model is used to calculate the interaction strength between the experts, the interaction strength between the experts and the corresponding task features, and the interaction strength between the task features.

[0169] Specifically, the expression of the interaction strength calculation model is:

[0170]

[0171] Where I(T; E) represents the interaction intensity calculation model; M represents the number of task features; N represents the number of experts; P(T i ,E j ) represents the negative entropy of the target probability distribution model; P(T i ) represents the task feature T in the forward feedback network j probability; Represents the expert E in the forward feedback network i probability.

[0172] If the expert E i For all task features T j If we specify the same frequency, then I(T; E) is 0; if each expert specifies a task feature T j , then the experts E i The mutual communication between them will be captured to the maximum, and I(T; E) will reach the maximum value of 1. In order to better understand the interaction intensity calculation model, the above expression of I(T; E) can be decomposed into the second expression of I(T; E):

[0173]

[0174] It can be understood that the first term of the second expression represents P(T i ,E j )=P(E i |T j )P(T j ), maximizing the first term of the second expression favors the stimulus condition distribution P(E i |T j ) is sharp, because P(Tj ) is a constant determined by the data distribution and is not affected by the model parameters. The second term of the second expression represents the i ) is a constant determined by , which can be ignored. The third term represents P(E j ) determines the entropy, maximizing this allows each expert E i Evenly use the data set. Therefore, use -I(T;E Y ) and each expert weight parameter w MI Thus, a cross-expert multi-task loss function is constructed to ensure that different experts E i The effective interaction between them alleviates the j The gradient contradiction between them can effectively improve the performance of the model, so the expression of the cross-expert multi-task loss function is determined as:

[0175]

[0176] Where L represents the cross-expert multi-task loss function; Indicates the T j The loss value of a task; M represents the number of task features; Indicates the T j The weight value of a task in the cross-expert multi-task loss function; w MI is the weight value corresponding to the interaction strength calculation model; I(T; E) represents the interaction strength calculation model; the above are only specific examples, which can be flexibly set according to needs in actual applications and are not limited here.

[0177] It should be understood that although Figure 2-Figure 3 , Figure 5-Figure 7 and Figure 11-Figure 14 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 2-Figure 3 , Figure 5-Figure 7 and Figure 11-Figure 14 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0178] Second, as Fig.15 As shown, a vehicle fault diagnosis device is provided, which includes: a text acquisition module 1510 and a fault diagnosis module 1520.

[0179] Among them, the text acquisition module 1510 is used to obtain the vehicle fault description text; the fault diagnosis module 1520 is used to input the vehicle fault description text into a pre-trained cross-expert fine-grained shared neural network model, and output the vehicle fault diagnosis result; wherein the cross-expert fine-grained shared neural network model includes a first input processing module, an encoder module, a second input processing module, a decoder module and an output module; the first input processing module is used to perform data processing on the vehicle fault description text to obtain the corresponding first feature vector; the encoder module is used to perform feature extraction processing on each first feature vector to obtain the corresponding second feature vector; the second input processing module is used to perform data processing on each second feature vector to obtain the corresponding third feature vector; the decoder module is used to perform classification prediction processing on each third feature vector to obtain the corresponding fourth feature vector; the output module is used to perform feature fusion and classification probability calculation processing on each fourth feature vector to obtain the vehicle fault diagnosis result.

[0180] In one of the embodiments, the vehicle fault diagnosis device further includes a model building module.

[0181] Among them, the model construction module is used to obtain a second preset number of historical vehicle fault description texts; the model construction module is used to randomly divide each historical vehicle fault description text to generate a training sample set and a test sample set; the model construction module is used to train a preset neural network model based on the self-attention mechanism according to the training sample set, test the neural network model according to the test sample set, and adjust the model parameters of the neural network model according to the training results, test results and the loss value calculated by the cross-expert multi-task loss function until the loss value calculated by the cross-expert multi-task loss function is less than the loss threshold, and generate a cross-expert fine-grained shared neural network model to output the vehicle fault diagnosis result based on the cross-expert fine-grained shared neural network model.

[0182] In one of the embodiments, the model building module includes a loss function determination unit.

[0183] Among them, the loss function determination unit is used to determine the target probability distribution model; the target probability distribution model is used to characterize the probability of each expert in the forward feedback network processing the corresponding task features; the loss function determination unit is used to determine the interaction intensity calculation model according to the joint distribution probability algorithm and the target probability distribution model; the interaction intensity calculation model is used to calculate the interaction intensity between each expert, the interaction intensity between each expert and the corresponding task features, and the interaction intensity between each task features; the loss function determination unit is used to determine the cross-expert multi-task loss function according to the interaction intensity calculation model and the crowd parameters of each expert.

[0184] In one embodiment, the expression of the cross-expert multi-task loss function is:

[0185]

[0186] Where L represents the cross-expert multi-task loss function; M represents the number of task features; Indicates the T j The weight value of the task feature in the cross-expert multi-task loss function; w MI is the weight value corresponding to the interaction strength calculation model; I(T; E) represents the interaction strength calculation model.

[0187] In one of the embodiments, the first input processing module includes a word embedding unit and a first regularization processing unit.

[0188] Among them, the word embedding unit is used to perform word embedding processing on the vehicle fault description text to obtain the corresponding word vector; the first regularization processing unit is used to perform regularization processing on each word vector to obtain the corresponding first feature vector.

[0189] In one embodiment, the encoder module includes a first multi-head attention layer, a first normalization layer, a first feed-forward layer, and a second normalization layer;

[0190] Among them, the first multi-head attention layer is used to perform feature diversification processing on each first eigenvector to obtain the corresponding first target feature vector; the first normalization layer is used to perform layer normalization processing on each first target feature vector to obtain the corresponding second target feature vector; the first forward feedback layer is used to perform generalization capability enhancement calculation on each second target feature vector to obtain the corresponding third target feature vector; wherein the first forward feedback layer includes a forward feedback network determined according to the fine-grained expert strategy and the specified shared expert strategy. The second normalization layer is used to perform layer normalization processing on each third target feature vector to obtain the corresponding second feature vector.

[0191] In one embodiment, the second input processing module includes a second input processing unit.

[0192] The second input processing unit is used to perform regularization processing on each second eigenvector to obtain a corresponding third eigenvector.

[0193] In one of the embodiments, the decoder module includes a masked multi-head attention layer, a third normalization layer, a second multi-head attention layer, a fourth normalization layer, a second forward feedback layer, and a fifth normalization layer.

[0194] Among them, the masked multi-head attention layer is used to perform eigenvalue filtering and feature diversification processing on each third eigenvector to obtain the corresponding fourth target eigenvector; the third normalization layer is used to perform layer normalization processing on each fourth target eigenvector to obtain the corresponding fifth target eigenvector; the second multi-head attention layer is used to perform feature diversification processing on each fifth target eigenvector to obtain the corresponding sixth target eigenvector; the fourth normalization layer is used to perform layer normalization processing on each sixth target eigenvector to obtain the corresponding seventh target eigenvector; the second front feedback layer is used to perform generalization capability enhancement calculation on each seventh target eigenvector to obtain the corresponding eighth target eigenvector; wherein the second front feedback layer includes a forward feedback network determined according to the fine-grained expert strategy and the specified shared expert strategy; the fifth normalization layer is used to perform layer normalization processing on each eighth target eigenvector to obtain the corresponding fourth eigenvector.

[0195] In one of the embodiments, the vehicle fault diagnosis device includes a forward feedback network determination module.

[0196] Among them, the feedback forward network determination module is used to perform expert fine-graining processing on the initial feedback forward network based on the fine-graining expert strategy to obtain the initial feedback forward network after fine-graining processing; the feedback forward network determination module is used to perform expert sharing processing on the initial feedback forward network after fine-graining processing based on the specified shared expert strategy to obtain the feedback forward network.

[0197] In one embodiment, the expression of the forward feedback network is:

[0198]

[0199] Among them, the expression of the gate value of the i-th expert is:

[0200]

[0201] Among them, the expression of the correlation score between the expert and the corresponding feature vector is:

[0202]

[0203] Among them, K s is the number of shared experts; mN-K s It is represented by the number of experts other than the shared experts; N is the number of experts before fine-grained processing; mN is the number of experts after fine-grained processing; FFN i (·) represents the process of the i-th expert processing the feature vector; g i,t represents the gate value of the i-th expert; s i,trepresents the correlation score between the expert and the corresponding feature vector; Topk(.,K) represents the set of the highest K correlation scores between the t-th feature vector and mN fine-grained experts; represents the i-th expert in the l-th layer; represents the output feature vector of the forward feedback network; Represents the input feature vector of the feedforward network.

[0204] In one embodiment, the output module includes a linear layer and a softmax.

[0205] Among them, the linear layer is used to perform feature fusion processing on each fourth eigenvector to obtain the corresponding ninth target eigenvector; the softmax layer is used to perform classification probability calculation processing on each ninth target eigenvector to obtain the corresponding vehicle fault diagnosis result.

[0206] In one embodiment, the expression of the cross-expert fine-grained shared neural network model is:

[0207]

[0208] Where x represents the input of the cross-expert fine-grained shared neural network model; E k (x) represents the fault feature vector mined by the cross-expert fine-grained shared neural network model; G k (x) represents the weight value corresponding to the input of the cross-expert fine-grained shared neural network model; y represents the output of the cross-expert fine-grained shared module.

[0209] In one embodiment, the text acquisition module 1510 includes a text acquisition unit.

[0210] Among them, the text acquisition unit is used to acquire unstructured vehicle production record data and unstructured vehicle fault manual data; the text acquisition unit is used to perform data cleaning processing on the unstructured vehicle production record data and the unstructured vehicle fault manual data to obtain the vehicle fault description text.

[0211] In one of the embodiments, the vehicle fault diagnosis result includes a fault keyword extraction result, a fault cause analysis result and a fault problem strategy result.

[0212] In one of the embodiments, the cross-expert fine-grained shared neural network model adopts a Transformer model.

[0213] The specific definition of the vehicle fault diagnosis device can be found in the definition of the vehicle fault diagnosis method above, which will not be repeated here. Each module in the above-mentioned vehicle fault diagnosis device can be implemented in whole or in part by software, hardware and a combination thereof. 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.

[0214] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Fig.16 As shown. The computer device includes a processor, a memory, a network interface and a database connected through a system bus. Among them, 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, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store and obtain vehicle fault description text data. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a vehicle fault diagnosis method is implemented.

[0215] Those skilled in the art will understand that Fig.16 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.

[0216] In a third aspect, a computer device is provided, which includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of any method in the above method embodiments are implemented.

[0217] In a fourth aspect, 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 of any method in the above method embodiments are implemented.

[0218] In a fifth aspect, a computer program product is provided, which includes a computer program, and when the computer program is executed by a processor, the steps of any method in the above method embodiments are implemented.

[0219] In this embodiment, the computer program product may be, but is not limited to, a vehicle fault diagnosis platform. Through the above-mentioned computer program product, cross-regional and cross-industry communication can be achieved, which not only broadens the horizons of vehicle users, but also brings new development directions to the vehicle fault diagnosis industry.

[0220] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods 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 memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0221] 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 specification.

[0222] The above-mentioned 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 invention patent. 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 patent of the present application shall be subject to the attached claims.

Claims

1. A vehicle fault diagnosis method, the method comprising: Get the vehicle fault description text; The vehicle fault description text is input into a pre-trained cross-expert fine-grained shared neural network model, and a vehicle fault diagnosis result is output; wherein the cross-expert fine-grained shared neural network model includes a first input processing module, an encoder module, a second input processing module, a decoder module and an output module; the first input processing module is used to perform data processing on the vehicle fault description text to obtain a corresponding first feature vector; the encoder module is used to perform feature extraction processing on each of the first feature vectors to obtain a corresponding second feature vector; the second input processing module is used to perform data processing on each of the second feature vectors to obtain a corresponding third feature vector; the decoder module is used to perform classification prediction processing on each of the third feature vectors to obtain a corresponding fourth feature vector; the output module is used to perform feature fusion and classification probability calculation processing on each of the fourth feature vectors to obtain the vehicle fault diagnosis result; The encoder module includes a first multi-head attention layer, a first normalization layer, a first pre-feedback layer and a second normalization layer; the step of performing feature extraction on each of the first feature vectors to obtain a corresponding second feature vector includes: Performing feature diversification processing on each of the first feature vectors through the first multi-head attention layer to obtain a corresponding first target feature vector; Performing layer normalization processing on each of the first target feature vectors through the first normalization layer to obtain a corresponding second target feature vector; Performing a generalization capability enhancement calculation on each of the second target feature vectors through the first forward feedback layer to obtain a corresponding third target feature vector; wherein the first forward feedback layer includes a forward feedback network determined according to a fine-grained expert strategy and a specified shared expert strategy; Performing layer normalization processing on each of the third target feature vectors through the second normalization layer to obtain a corresponding second feature vector; The decoder module includes a masked multi-head attention layer, a third normalization layer, a second multi-head attention layer, a fourth normalization layer, a second forward feedback layer and a fifth normalization layer; the classification prediction processing of each of the third feature vectors to obtain the corresponding fourth feature vector includes: Performing eigenvalue filtering and feature diversification processing on each of the third eigenvectors through the masked multi-head attention layer to obtain a corresponding fourth target eigenvector; Performing layer normalization processing on each of the fourth target feature vectors through the third normalization layer to obtain a corresponding fifth target feature vector; Performing feature diversification processing on each of the fifth target feature vectors through the second multi-head attention layer to obtain a corresponding sixth target feature vector; Performing layer normalization processing on each of the sixth target feature vectors through the fourth normalization layer to obtain a corresponding seventh target feature vector; Performing a generalization capability enhancement calculation on each of the seventh target feature vectors through the second forward feedback layer to obtain a corresponding eighth target feature vector; wherein the second forward feedback layer includes a forward feedback network determined according to the fine-grained expert strategy and the specified shared expert strategy; Performing layer normalization processing on each of the eighth target feature vectors through the fifth normalization layer to obtain the corresponding fourth feature vector; Wherein, determining the forward feedback network includes: Based on the fine-grained expert strategy, the initial forward feedback network is subjected to expert fine-grained processing to obtain the initial forward feedback network after fine-grained processing; Based on the designated shared expert strategy, the initial forward feedback network after the fine-grained processing is subjected to expert sharing processing to obtain the forward feedback network.

2. The method according to claim 1, characterized in that Constructing the cross-expert fine-grained shared neural network model includes: Acquire a second preset number of historical vehicle fault description texts; Randomly divide the historical vehicle fault description texts to generate a training sample set and a test sample set; The preset neural network model based on the self-attention mechanism is trained according to the training sample set, the neural network model is tested according to the test sample set, and the model parameters of the neural network model are adjusted according to the training results, the test results and the loss value calculated by the cross-expert multi-task loss function until the loss value calculated by the cross-expert multi-task loss function is less than the loss threshold, and the cross-expert fine-grained shared neural network model is generated to output the vehicle fault diagnosis result based on the cross-expert fine-grained shared neural network model.

3. The method according to claim 2, characterized in that Determining the cross-expert multi-task loss function includes: Determine a target probability distribution model; the target probability distribution model is used to characterize the probability of each expert in the forward feedback network processing the corresponding task feature; Determine an interaction strength calculation model according to the joint distribution probability algorithm and the target probability distribution model; the interaction strength calculation model is used to calculate the interaction strength between the experts, the interaction strength between the experts and the corresponding task features, and the interaction strength between the task features; The cross-expert multi-task loss function is determined according to the interaction intensity calculation model and the crowd parameters of each of the experts.

4. The method according to claim 1, characterized in that The first feature vector is obtained after data processing of the vehicle fault description text, including: Performing word embedding processing on the vehicle fault description text to obtain a corresponding word vector; Regularization is performed on each of the word vectors to obtain the corresponding first feature vector.

5. The method according to claim 1, characterized in that The step of performing data processing on each of the second feature vectors to obtain a corresponding third feature vector comprises: Regularization is performed on each of the second eigenvectors to obtain the corresponding third eigenvector.

6. The method according to claim 1, characterized in that The obtaining of the new energy vehicle fault description text includes: Obtain unstructured vehicle production record data and unstructured vehicle fault manual data; The unstructured vehicle production record data and the unstructured vehicle fault manual data are cleansed to obtain the vehicle fault description text.

7. The method according to claim 1, characterized in that The output module includes a linear layer and a softmax layer; the cross-expert fine-grained shared neural network sub-model is obtained after performing feature fusion and classification probability calculation processing on each of the fourth feature vectors, including: Performing feature fusion processing on each of the fourth eigenvectors through the linear layer to obtain a corresponding ninth target eigenvector; The classification probability calculation process is performed on each of the ninth target feature vectors through the softmax layer to obtain the corresponding vehicle fault diagnosis result.

8. A vehicle fault diagnosis device, characterized in that: The device comprises: A text acquisition module is used to acquire vehicle fault description text; A fault diagnosis module, used for inputting the vehicle fault description text into a pre-trained cross-expert fine-grained shared neural network model, and outputting a vehicle fault diagnosis result; wherein the cross-expert fine-grained shared neural network model includes a first input processing module, an encoder module, a second input processing module, a decoder module and an output module; the first input processing module is used for performing data processing on the vehicle fault description text to obtain a corresponding first feature vector; the encoder module is used for performing feature extraction processing on each of the first feature vectors to obtain a corresponding second feature vector; the second input processing module is used for performing data processing on each of the second feature vectors to obtain a corresponding third feature vector; the decoder module is used for performing classification prediction processing on each of the third feature vectors to obtain a corresponding fourth feature vector; the output module is used for performing feature fusion and classification probability calculation processing on each of the fourth feature vectors to obtain the vehicle fault diagnosis result; Among them, the encoder module includes a first multi-head attention layer, a first normalization layer, a first forward feedback layer and a second normalization layer; the feature extraction processing of each of the first feature vectors obtains the corresponding second feature vector, including: the first multi-head attention layer is used to perform feature diversification processing on each of the first feature vectors to obtain the corresponding first target feature vector; the first normalization layer is used to perform layer normalization processing on each of the first target feature vectors to obtain the corresponding second target feature vector; the first forward feedback layer is used to perform generalization capability enhancement calculation on each of the second target feature vectors to obtain the corresponding third target feature vector; wherein the first forward feedback layer includes a forward feedback network determined according to the fine-grained expert strategy and the specified shared expert strategy; the second normalization layer is used to perform layer normalization processing on each of the third target feature vectors to obtain the corresponding second feature vector; Among them, the decoder module includes a masked multi-head attention layer, a third normalization layer, a second multi-head attention layer, a fourth normalization layer, a second pre-feedback layer and a fifth normalization layer; the classification prediction processing of each of the third eigenvectors obtains the corresponding fourth eigenvector, including: the masked multi-head attention layer is used to perform eigenvalue filtering and feature diversification processing on each of the third eigenvectors to obtain the corresponding fourth target eigenvector; the third normalization layer is used to perform layer normalization processing on each of the fourth target eigenvectors to obtain the corresponding fifth target eigenvector; the second multi-head attention layer is used to perform classification prediction processing on each of the third eigenvectors to obtain the corresponding fourth target eigenvector The fifth target feature vector is subjected to feature diversification processing to obtain a corresponding sixth target feature vector; the fourth normalization layer is used to perform layer normalization processing on each of the sixth target feature vectors to obtain a corresponding seventh target feature vector; the second front feedback layer is used to perform generalization capability enhancement calculation on each of the seventh target feature vectors to obtain a corresponding eighth target feature vector; wherein the second front feedback layer includes a forward feedback network determined according to a fine-grained expert strategy and a specified shared expert strategy; the fifth normalization layer is used to perform layer normalization processing on each of the eighth target feature vectors to obtain a corresponding fourth feature vector; Wherein, determining the feedforward network includes: based on the fine-grained expert strategy, performing expert fine-graining processing on the initial feedforward network to obtain the fine-grained initial feedforward network; based on the specified shared expert strategy, performing expert sharing processing on the fine-grained initial feedforward network to obtain the feedforward network.

9. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 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 7 are implemented.

Images