An internal combustion engine fault diagnosis method, device and equipment

Abstract

The present disclosure provides an internal combustion engine fault diagnosis method, device and equipment, wherein the method comprises: acquiring a plurality of state parameters of the internal combustion engine in a current time step, and grouping the plurality of state parameters to obtain a plurality of parameter groups; extracting first deep features of each parameter group through a denoising encoder; processing the first deep features through an attention mechanism module to obtain attention data; wherein the attention data is used to indicate the contribution of the first deep features to fault identification; processing the attention data through a bidirectional gated recurrent unit to obtain a hidden state parameter of the current time step, and identifying a fault of the internal combustion engine according to the hidden state parameter; wherein the hidden state parameter is used to indicate the running state of the internal combustion engine.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of internal combustion engines, in particular to an internal combustion engine fault diagnosis method, device and equipment. BACKGROUND

[0002] Internal combustion engines have complex structures and often operate in harsh conditions with complex and variable working conditions, which are prone to performance decline and even frequent faults. Health state diagnosis can timely detect faults and prevent fault deterioration, improve operation reliability and safety, and is a prerequisite for predictive maintenance and health management. Current internal combustion engine diagnosis methods based on vibration, acoustic emission and other methods are research hotspots and have achieved fruitful results, but the hardware cost of vibration monitoring and other methods is high, and the model is also relatively complex, which is not suitable for small vehicle internal combustion engines. Internal combustion engines are equipped with key performance parameter monitoring sensors, which can monitor the operation state of the internal combustion engine to a certain extent without adding new sensors. However, the current health state diagnosis method based on the parameters of the internal combustion engine mainly judges the abnormality by independently setting the threshold range of each monitoring parameter, and has not realized the deep fusion analysis of multiple parameter correlations. The existing monitoring parameters are not fully utilized, and the health monitoring function needs to be greatly improved. SUMMARY

[0003] The embodiments of the present disclosure at least provide an internal combustion engine fault diagnosis method, device and equipment.

[0004] In a first aspect, the embodiments of the present disclosure provide an internal combustion engine fault identification method, comprising: acquiring a plurality of state parameters of an internal combustion engine in a current time step, and grouping the plurality of state parameters to obtain a plurality of parameter groups; extracting first deep features of each of the parameter groups through a denoising encoder; processing the first deep features through an attention mechanism module to obtain attention data; wherein the attention data is used to indicate the contribution of the first deep features to fault identification; processing the attention data through a bidirectional gated recurrent unit to obtain a hidden state parameter of the current time step, and identifying a fault of the internal combustion engine according to the hidden state parameter; wherein the hidden state parameter is used to indicate the running state of the internal combustion engine.

[0005] In an optional implementation, the denoising encoder comprises a plurality of encoders; and the extracting the first deep features of each of the parameter groups through the denoising encoder comprises: determining an encoder corresponding to each of the parameter groups; extracting the first deep features of the corresponding parameter groups through the encoder to obtain the first deep features of each of the parameter groups, and fusing the first deep features of the parameter groups to process the fused first deep features through the attention mechanism module.

[0006] In an optional implementation, the attention mechanism module comprises a pooling layer, a first fully connected layer and a second fully connected layer; the processing of the first deep layer feature by the attention mechanism module to obtain attention data comprises: performing average pooling processing on the first deep layer feature by the pooling layer to obtain a first pooling result, and performing maximum pooling processing on the first deep layer feature to obtain a second pooling result; performing fully connected processing on the first pooling result by the first fully connected layer to obtain a first fully connected result, and performing fully connected processing on the second pooling result by the second fully connected layer to obtain a second fully connected result; merging the first fully connected result and the second fully connected result to obtain a merged fully connected result; processing the merged fully connected result by a Sigmoid function to obtain a feature weight of the first deep layer feature, and determining the attention data based on the feature weight.

[0007] In an optional implementation, the processing of the attention data by the bidirectional gated recurrent unit to obtain the hidden state parameter of the current time step comprises: determining reset gate data of the current time step based on the attention data, and determining update gate data of the current time step based on the attention data; determining a candidate hidden state parameter of the current time step based on the reset gate data; updating the candidate hidden state parameter based on the update gate data to obtain the hidden state parameter.

[0008] In an optional implementation, the method further comprises training the denoising encoder by: extracting, by an initial encoder, second deep layer features of corresponding training parameter groups in a training sample; wherein the training sample comprises a plurality of training parameter groups, each training parameter group comprising a plurality of parameter sets, and a same parameter set comprising state parameters of parameters of a same type under different working conditions; decoding, by an initial decoder, the second deep layer features extracted by the initial encoder to obtain a decoding result; and self-training the initial decoder and the initial encoder based on a difference between the decoding result and the corresponding training parameter groups to obtain the denoising encoder after training.

[0009] In an optional implementation, the method further comprises training the bidirectional gated recurrent unit by: obtaining training attention data obtained by processing the second deep feature by the attention mechanism module; determining a forward transmission hidden state parameter of a current training time step based on the training attention data and a hidden state parameter of a previous training time step; determining a backward transmission hidden state parameter of the current training time step based on the training attention data and a hidden state parameter of a next training time step; determining a bidirectional transmission hidden state parameter of the current training time step based on the forward transmission hidden state parameter and the backward transmission hidden state parameter, and training the bidirectional gated recurrent unit based on the bidirectional transmission hidden state parameter.

[0010] In an optional implementation, the plurality of training parameter groups are determined by: determining mutual information values between the plurality of parameter sets; performing parameter classification processing on the plurality of parameter sets based on the mutual information values to obtain a classification result; wherein the classification result includes a plurality of classification groups, and mutual information values between state parameters in a same classification group satisfy grouping requirements; and determining the plurality of training parameter groups based on the plurality of classification groups in the classification result; wherein each classification group corresponds to a training parameter group.

[0011] In an optional implementation, the mutual information values between the plurality of parameter sets are determined by: determining a target information entropy of each parameter set; determining a joint information entropy between any two parameter sets; and determining the mutual information values between the plurality of parameter sets based on the target information entropy and the joint information entropy.

[0012] In a second aspect, the embodiments of the present disclosure provide an internal combustion engine fault identification device, comprising: a grouping unit configured to obtain a plurality of state parameters of an internal combustion engine at a current time step, and group the plurality of state parameters to obtain a plurality of parameter groups; an extraction unit configured to extract a first deep feature of each parameter group by a denoising encoder; a feature processing unit configured to process the first deep feature by an attention mechanism module to obtain attention data; wherein the attention data is used to indicate a contribution degree of the first deep feature to fault identification; and an identification unit configured to process the attention data by a bidirectional gated recurrent unit to obtain a hidden state parameter of the current time step, and identify a fault of the internal combustion engine according to the hidden state parameter; wherein the hidden state parameter is used to indicate an operating state of the internal combustion engine.

[0013] In a third aspect, the embodiments of the present disclosure further provide a computer device, comprising a processor, a memory and a bus, the memory stores machine readable instructions executable by the processor, when the computer device is running, the processor and the memory communicate through the bus, and the machine readable instructions are executed by the processor to perform the steps of the first aspect or any possible implementation manner of the first aspect.

[0014] In a fourth aspect, the embodiments of the present disclosure further provide a computer readable storage medium, the computer readable storage medium stores a computer program, when the computer program is executed by a processor, the steps of the first aspect or any possible implementation manner of the first aspect are performed.

[0015] The embodiments of the present disclosure provide an internal combustion engine fault diagnosis method, device and equipment. In the embodiments of the present disclosure, first, a plurality of state parameters of the internal combustion engine in a current time step are obtained, and the plurality of state parameters are grouped to obtain a plurality of parameter groups. Then, a first deep feature of each parameter group is extracted by a denoising encoder, the first deep feature is processed by an attention mechanism module to obtain attention data, and the attention data is processed by a bidirectional gated recurrent unit to obtain a hidden state parameter of the current time step. Finally, the internal combustion engine is fault-recognized according to the hidden state parameter.

[0016] In the above embodiments, by dividing the plurality of state parameters into a plurality of parameter groups, the complexity of the model can be reduced. By extracting the first deep feature of each parameter group through the denoising encoder, processing the first deep feature through the attention mechanism module to obtain the attention data, processing the attention data through the bidirectional gated recurrent unit to obtain the hidden state parameter of the current time step, and fault-recognizing the internal combustion engine according to the hidden state parameter, the automatic diagnosis and evaluation of the health state of the internal combustion engine can be realized. Compared with the diagnosis method by threshold comparison in the prior art, the method provided by the embodiments of the present disclosure can improve the diagnosis accuracy of fault diagnosis, and has important theoretical value and engineering application prospect for improving the operation and maintenance level of the internal combustion engine.

[0017] In order to make the above objectives, characteristics and advantages of the present disclosure more apparent and easy to understand, the following preferred embodiments are specifically described below with reference to the accompanying drawings. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the drawings needed to be used in the embodiments will be briefly introduced as follows. The drawings incorporated into the specification and form a part of the specification, which show the embodiments consistent with the present disclosure, and are used to explain the technical solutions of the present disclosure together with the specification. It should be understood that the following drawings only show some embodiments of the present disclosure, and therefore should not be regarded as a limitation on the scope, and other related drawings can also be obtained by those of ordinary skill in the art without creative labor.

[0019] Figure 1 A flow chart of a method for diagnosing faults of an internal combustion engine is shown;

[0020] Figure 2 A grouping structure schematic diagram of a state parameter after being grouped by a condensed independent variable grouping analysis method is shown;

[0021] Figure 3 A network structure diagram of a denoising autoencoder module is shown;

[0022] Figure 4 A flowchart of a denoising autoencoder is shown;

[0023] Figure 5 A processing flowchart of an attention mechanism module is shown;

[0024] Figure 6 A network structure diagram of a bidirectional gated recurrent unit is shown;

[0025] Figure 7 A loss function curve diagram of an internal combustion engine fault identification model is shown;

[0026] Figure 8 A curve diagram of the accuracy of an internal combustion engine fault identification model is shown;

[0027] Figure 9 A flowchart of another method for diagnosing faults of an internal combustion engine is shown;

[0028] Figure 10 A schematic diagram of an internal combustion engine fault diagnosis device is shown;

[0029] Figure 11 A schematic diagram of a computer device is shown. DETAILED DESCRIPTION

[0030] So that the purposes, technical solutions and advantages of the embodiments of the present disclosure are more apparent, the following will describe the technical solutions in the embodiments of the present disclosure clearly and completely with reference to the accompanying drawings in the embodiments of the present disclosure. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, but not all the embodiments. The components of the embodiments of the present disclosure generally described and shown in the accompanying drawings can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present disclosure provided in the accompanying drawings is not intended to limit the scope of the claimed present disclosure, but only represents selected embodiments of the present disclosure. Based on the embodiments of the present disclosure, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present disclosure.

[0031] It is found through research that the internal combustion engine has a complex structure and is often operated in harsh conditions of complex and variable working conditions, which is prone to performance decline and even frequent failures. Health state diagnosis can timely find faults and prevent fault deterioration, improve operation reliability and safety, and is a prerequisite for predictive maintenance and health management and a problem to be solved. The current internal combustion engine diagnosis method based on vibration, acoustic emission and the like is a research hotspot and has achieved fruitful results, but the hardware cost of vibration and other monitoring methods is high, and the model is also relatively complex, which is not suitable for small vehicle internal combustion engines. The internal combustion engine is equipped with key performance parameter monitoring sensors, and does not need to add new sensors to monitor the internal combustion engine to a certain extent. However, the current health state diagnosis method based on the parameters of the engine mainly judges the abnormality by independently setting the threshold range of each monitoring parameter, and has not realized the deep fusion analysis of the correlation of multiple parameters, and the mining and utilization of existing monitoring parameters are insufficient, and the construction of health monitoring function needs to be greatly improved.

[0032] Based on the above research, the present disclosure provides an internal combustion engine fault diagnosis method, device and equipment. In the embodiments of the present disclosure, by dividing a plurality of state parameters into a plurality of parameter groups, the complexity of the model can be reduced; by extracting first deep features of each parameter group through a denoising encoder, and processing the first deep features through an attention mechanism module to obtain attention data, and processing the attention data through a bidirectional gated recurrent unit to obtain a hidden state parameter of a current time step, and identifying faults of the internal combustion engine according to the hidden state parameter, the automatic diagnosis and evaluation of the health state of the internal combustion engine can be realized, and compared with the diagnosis method through threshold comparison in the prior art, the method provided in the embodiments of the present disclosure can improve the diagnosis accuracy of fault diagnosis, and has important theoretical value and engineering application prospect for improving the operation and maintenance level of the internal combustion engine.

[0033] The above-mentioned defects are the results of the inventors' practice and careful research, and thus the discovery process of the above-mentioned problems and the solutions proposed by the present disclosure to solve the above-mentioned problems should be the contributions of the inventors to the present disclosure.

[0034] It should be noted that similar reference numerals and letters represent similar items in the following drawings, and thus, once an item is defined in one drawing, it need not be further defined and explained in subsequent drawings.

[0035] To facilitate the understanding of the present embodiment, first, a kind of internal combustion engine fault identification method disclosed by the present embodiment will be introduced in detail, and the execution subject of the internal combustion engine fault identification method provided by the present embodiment is generally a computer device with certain computing power. In some possible implementation manners, the internal combustion engine fault identification method can be realized by the way of processor calling computer readable instructions stored in the memory.

[0036] The internal combustion engine fault identification method provided by the present embodiment will be described below.

[0037] Embodiment One

[0038] Referring to Figure 1 Fig. 1 is a flowchart of an internal combustion engine fault identification method provided by the present embodiment, and the method comprises steps S101-S104, wherein:

[0039] S101: Obtain a plurality of state parameters of the internal combustion engine in the current time step, and group the plurality of state parameters to obtain a plurality of parameter groups.

[0040] In the present embodiment, a plurality of sensors can be installed in different positions of the internal combustion engine in advance, and the state parameters of the internal combustion engine under various working conditions can be collected through the sensors.

[0041] After obtaining the plurality of state parameters, the plurality of state parameters can be grouped to obtain a plurality of parameter groups, wherein each parameter group contains at least one type of state parameter.

[0042] After obtaining the plurality of parameter groups, each parameter group can be processed by a fault identification model to obtain a fault identification result of the internal combustion engine, wherein the fault identification result contains: a fault state, a normal state, and a fault type in the fault state, such as insufficient oil supply, advanced injection angle, delayed injection angle, etc. Here, the fault identification model contains a denoising encoder, an attention mechanism module and a bidirectional gated recurrent unit.

[0043] S102: Extract first deep features of each parameter group through the denoising encoder.

[0044] After obtaining the plurality of parameter groups, first, the first deep feature of each parameter group can be extracted by the denoising encoder, so as to obtain the deep feature representation of each parameter group. By extracting the first deep feature, the purposes of dimension reduction and denoising of data can be achieved, so as to improve the speed and accuracy of fault identification.

[0045] S103: The first deep feature is processed by the attention mechanism module to obtain attention data; wherein the attention data is used to indicate the contribution degree of the first deep feature to fault identification.

[0046] After obtaining the first deep feature, the first deep feature can be processed by the attention mechanism module, so as to obtain the attention data used to indicate the contribution degree of the first deep feature to fault identification.

[0047] S104: The attention data is processed by the bidirectional gated recurrent unit to obtain the hidden state parameter of the current time step, and the internal combustion engine is identified according to the hidden state parameter; wherein the hidden state parameter is used to indicate the running state of the internal combustion engine.

[0048] In the embodiment of the present disclosure, by dividing the plurality of state parameters into a plurality of parameter groups, the complexity of the model can be reduced; by extracting the first deep feature of each parameter group through the denoising encoder, processing the first deep feature through the attention mechanism module to obtain the attention data, and processing the attention data through the bidirectional gated recurrent unit to obtain the hidden state parameter of the current time step, and identifying the internal combustion engine according to the hidden state parameter, the way can realize automatic diagnosis and evaluation of the health state of the internal combustion engine. Compared with the diagnosis way by threshold comparison in the prior art, the way provided in the embodiment of the present disclosure can improve the diagnosis accuracy of fault diagnosis, and has important theoretical value and engineering application prospect for improving the operation and maintenance level of the internal combustion engine.

[0049] In the embodiment of the present disclosure, before the plurality of parameter groups are processed by the denoising encoder, the attention mechanism module and the bidirectional gated recurrent unit in the fault identification model, the fault identification model also needs to be built and trained. The building process and the training process of the fault identification model will be introduced in detail below, wherein the building process and the training process include the following steps:

[0050] Step S1: Collect training samples.

[0051] Here, we collect data from internal combustion engines under different operating conditions and organize it to create a dataset. We then preprocess the dataset. We apply one-hot encoding to assign unique labels to the dataset to generate training samples. These samples are then divided into a training set and a test set.

[0052] Here, the collected data set can be expressed as Formula 1: D = {F1, F2, ..., F i ,...,F N}; where D is the collected data set, F i represents the data set under the i-th operating condition, and N is the number of operating conditions.

[0053] In formula 1, F i It can be expressed as formula 2: in, represents the jth sample data under the i-th operating condition data set, and n represents the number of samples under the i-th operating condition data set.

[0054] In formula 2, It can be expressed as formula 3: in, represents the mth state parameter of the jth sample data under the i-th operating condition, and M represents the number of state parameters of the j-th sample data under the i-th operating condition.

[0055] After obtaining the dataset described above, the dataset may be preprocessed to obtain a preprocessing result.

[0056] Here, we can use formula 4 Normalize the sample data, where X new represents the sample data after normalization, X represents the sample data that needs to be normalized, and X max Indicates the maximum value of the state parameter in the sample data that needs to be normalized, X min Indicates the minimum value of the sample parameter in the sample data that needs to be normalized.

[0057] After normalizing the sample data, the dataset of each state parameter under all operating conditions can be determined based on the normalized sample data, thereby obtaining training samples. For example, the normalized sample data can be transposed to obtain training samples.

[0058] Here, the above training samples can be expressed by formula 5: Y = {y1, y2, ..., y m ,...,y M}, Y is the state parameter set under all operating conditions (i.e., training samples), ym represents a data set (i.e., a parameter set) of all operating conditions under the mth state parameter, and M is the number of types of state parameters.

[0059] In the embodiments of the present disclosure, simulation software (for example, GT-POWER) can be used to obtain the required simulation data, so as to determine the training samples based on the simulation data.

[0060] Firstly, a simulation model of a four-cylinder diesel engine is established in the GT-POWER software; secondly, based on the simulation model, the operating conditions to be simulated are set, and the simulation of the diesel engine fault is realized by adjusting the parameter settings in the simulation model. Specifically, the corresponding parameters of the fault mode can be changed by performing corresponding adjustment operations on the simulation model of the diesel engine in the GT-POWER software, while keeping other parameters unchanged. In the embodiments of the present disclosure, three common faults of the diesel engine, namely, insufficient fuel supply, advanced injection angle, and delayed injection angle, and the normal operating condition of the diesel engine are set and simulated; and based on the SAE J1939 protocol, 35 state parameters representing the operating state of the diesel engine are selected from the simulation results.

[0061] After obtaining the training samples, the training samples can be grouped to obtain a plurality of training parameter groups. Here, the plurality of training parameter groups can be determined by the following method, specifically including the following steps:

[0062] Step S11: determining the mutual information values between the plurality of parameter sets;

[0063] Step S12: performing parameter classification processing on the plurality of parameter sets based on the mutual information values to obtain a classification result; wherein the classification result includes a plurality of classification groups, and the mutual information values between the state parameters in the same classification group meet the grouping requirements;

[0064] Step S13: determining the plurality of training parameter groups based on the plurality of classification groups in the classification result; wherein each classification group corresponds to a training parameter group.

[0065] In the embodiments of the present disclosure, the independent variable grouping analysis method can be applied, and the size of the mutual information value can be used as the grouping basis. The state parameters with large mutual information values are divided into the same group, and the state parameters with small mutual information values are divided into different groups, so as to realize automatic grouping processing of the training samples, thereby obtaining a plurality of training parameter groups.

[0066] In specific implementation, firstly, the mutual information values between the plurality of parameter sets can be determined. Here, the plurality of parameter sets can be understood as the data set y mAfter determining the mutual information values ​​between multiple parameter sets, the multiple parameter sets can be classified based on the mutual information values ​​to obtain multiple classification groups, where the same classification group can include at least one parameter set. If the same classification group includes multiple parameter sets, the mutual information values ​​between the multiple parameter sets meet the grouping requirements, for example, the mutual information values ​​are greater than a preset threshold.

[0067] Here, the number of training parameter groups can be pre-set, and then parameter classification processing can be performed on multiple parameter sets based on this number. For example, assume that the multiple parameter sets include set A, set B, and set C, and the number of training parameter groups is 2. In this case, the mutual information value M1 between set A and set B can be calculated, the mutual information value M2 between set A and set C can be calculated, and the mutual information value M4 between set C and set B can be calculated. If the mutual information value M1 is greater than a preset threshold, set A and set B are classified into the same classification group, and set C is classified into another classification group, thereby obtaining two classification groups. Afterwards, these two classification groups can be determined as training parameter groups.

[0068] For example, Figure 2 As shown, assuming that there are 35 types of state parameters, numbered 0-34, then you can follow Figure 2 The grouping method shown in is used to group multiple parameter sets. For example, Figure 2 As shown, assuming that the number of training parameter groups is 4, the parameter sets numbered 31, 0, 6, 1, 7, 4, 11, 8 and 9 can be divided into the same training parameter group, and the parameter sets numbered 10, 17, 22, 29, 5, 12, 13 and 14 can be divided into the same training parameter group, which will not be described one by one here.

[0069] In an optional implementation, determining the mutual information value between the multiple parameter sets specifically includes the following steps:

[0070] Step S111: determining the target information entropy of each parameter set;

[0071] Step S112: determining the joint information entropy between any two parameter sets;

[0072] Step S113: Determine mutual information values ​​between the multiple parameter sets based on the target information entropy and the joint information entropy.

[0073] In the embodiment of the present disclosure, the target information entropy of each parameter set can be calculated using Formula 6:

[0074] Where, represents the parameter set y m A state parameter in The probability value of occurrence, N x n represents the data quantity of a state parameter.

[0075] Next, the joint information entropy between any two parameter sets can be determined by Formula Seven: Where y m , y n respectively represent the parameter set of all working conditions under the m, n state parameters, represents the probability value of occurrence in the parameter set y m . represents the probability value of occurrence in the parameter set y m . and the parameter set y n .

[0076] After the target information entropy and the joint information entropy are determined, the mutual information value between multiple parameter sets can be determined by Formula Eight:

[0077] Where H(y i ) can be calculated by Formula Six, and H((y1,...y m ),(y m ,...y M )) can be calculated by Formula Seven.

[0078] In the above embodiment, the multiple parameter sets are classified and processed by the mutual information value to obtain multiple training parameter groups. In the case where the types of state parameters are more, the state parameters can be grouped according to the association between the state parameters, so as to reduce the number of training parameter groups, and further reduce the model size of the fault recognition model and improve the running speed of the fault recognition model. At the same time, by classifying and processing the multiple parameter sets, it can be ensured that the number of state parameters in each training parameter group meets the training requirements of the fault recognition model, and further ensures the training accuracy of the fault recognition model.

[0079] Step S2: model training.

[0080] In the embodiment of the present disclosure, after the training samples are grouped in the manner described above to obtain the training parameter groups, the fault recognition model can be obtained based on the building of the noise reduction encoder, the attention mechanism module and the bidirectional gated recurrent unit.

[0081] In an optional embodiment, the noise reduction encoder can be built and trained in the following manner, specifically including the following steps:

[0082] ​​Step S21: extracting second deep features corresponding to the training parameter group in the training sample through the initial encoder; wherein the training sample includes multiple training parameter groups, each training parameter group includes multiple parameter sets, and the same parameter set includes state parameters of the same type under various working conditions;

[0083] Step S22: decoding the second deep features extracted by the initial encoder through the initial decoder to obtain a decoding result;

[0084] Step S23: performing self-training on the initial decoder and the initial encoder based on the difference between the decoding result and the corresponding training parameter group, and obtaining the noise reduction encoder after training.

[0085] In the disclosed embodiment, after automatically grouping multiple training parameter groups, a separate initial denoising autoencoder can be pre-built for each training parameter group. The initial denoising autoencoder includes an initial decoder and an initial encoder. Appropriate hyperparameters can then be selected for each initial denoising autoencoder to reduce the model's loss.

[0086] In specific implementation, a training parameter group X={x1,x2,...,x n}, where n is the dimension of the training parameter group. Then, as Figure 3 As shown, noise is added to the training parameter group (i.e., input data) to obtain Afterwards, if Figure 3 As shown, the second deep features of the training parameter group can be extracted by the initial encoder. The second deep features can be understood as deep feature representations of the training parameter group, which can describe the characteristics of the training parameter group as a whole. The feature dimension of the second deep features is smaller than the data dimension of the training parameter group. In other words, the second deep features are deep feature representations that can describe the training parameter group as a whole after dimensionality reduction and denoising of the training parameter group.

[0087] Here, we can use formula 9 Implement the encoding process of the initial encoder. Next, Figure 3 As shown, the second deep features extracted by the initial encoder can be decoded by the initial decoder to obtain a decoding result. Implement the decoding process of the initial decoder.

[0088] The encoding process of the initial encoder and the decoding process of the initial decoder are constructed into a complete denoising autoencoder. By adjusting the hyperparameters such as the learning rate, number of cycles, loss function, activation function, etc. in the denoising autoencoder and selecting appropriate network values, the loss value is reduced so that the reconstructed data X output by Formula 10 is as close as possible to the input X, and the denoising encoder is obtained. Figure 4 As shown, the hyperparameters of the Dense1 network to the Dense3 network can be adjusted. At this point, the denoising encoder can extract features from the training parameter group using Formula 9 to obtain a deep feature representation of the input data.

[0089] Here, in the above formula 9 and formula 10, Z={z1,z2,...,z m} is the deep representation feature of the hidden layer output (i.e., the second deep feature), g1, W1, and b1 are the node activation function, weight matrix, and node bias from the input layer to the hidden layer during the encoding process, respectively. is the reconstructed data output by the output layer, g2, W2, and b2 are the node activation function, weight matrix, and node bias from the hidden layer to the output layer during the decoding process, respectively.

[0090] During the construction phase of the fault recognition model, after the noise reduction encoder is trained in the manner described above, an attention mechanism module can be constructed to process the second deep features through the constructed attention mechanism module to obtain attention data during training.

[0091] Here, the attention mechanism module includes a pooling layer, a first fully connected layer, and a second fully connected layer; wherein, the structures of the first fully connected layer and the second fully connected layer can be the same. The first fully connected layer and the second fully connected layer each include two fully connected operations, and the two fully connected operations of the first fully connected layer and the two fully connected operations of the second fully connected layer have the same operating process. Here, the number of neurons in the first fully connected operation of the fully connected layer (the first fully connected layer and the second fully connected layer) is relatively small, and the number of neurons in the second fully connected operation is the same as the number of channels. Here, the pooling layer includes an average pooling layer and a maximum pooling layer.

[0092] When implementing it specifically, Figure 5 As shown, after obtaining the feature matrix (i.e., the feature matrix of the second deep features), the second deep features of each channel can be processed respectively by the average pooling layer and the maximum pooling layer to obtain the average pooling processing result and the maximum pooling processing result.

[0093] Here, the average pooling operation and the maximum pooling operation can be performed on the second deep layer features through Formula 11 W1=f([MLP(AvgPool(X)),MLP(MaxPool(X))]).

[0094] Then, ifFigure 5 As shown, the average pooling processing result can be subjected to twice full connection operations by a first full connection layer to obtain a full connection operation result A, and the maximum pooling processing result can be subjected to twice full connection operations by a second full connection layer to obtain a full connection operation result B. Next, as shown, the full connection operation result A and the full connection operation result B can be combined by an Add layer, and then subjected to processing by a Sigmoid function to obtain a feature weight W1. Figure 5

[0095] In the above formula eleven, f is a function expression of Sigmoid, f(x) = 1 / (1+e- x The range of the weight W1 is between 0 and 1.

[0096] After the feature weight and the feature matrix are subjected to multiplication processing, an updated feature matrix, i.e., feature data processed by the channel attention mechanism module (i.e., attention data in training) can be obtained.

[0097] In the embodiments of the present disclosure, by applying the channel attention mechanism, a corresponding feature weight can be calculated for each feature in the deep feature, and then the feature weight is used to represent the contribution degree of different features in the deep feature to fault identification, so as to highlight the feature part that is helpful to the model training result and ignore the feature part that is not helpful to the model training result, thereby improving the recognition accuracy of the fault identification model.

[0098] After the attention mechanism module is built in the above-described manner, a bidirectional gate recurrent unit can be built. By constructing the bidirectional gate recurrent unit, the dependency relationship on a sequence with a long time step distance in the vibration signal can be better captured.

[0099] In an optional embodiment, the bidirectional gate recurrent unit can be built and trained in the following manner, specifically including the following steps:

[0100] Step S24: obtaining training attention data obtained by processing the second deep feature by the attention mechanism module;

[0101] Step S25: determining a forward transmission hidden state parameter of a current training time step based on the training attention data and a hidden state parameter of a previous training time step;

[0102] Step S26: determining a backward transmission hidden state parameter of the current training time step based on the training attention data and a hidden state parameter of a next training time step;

[0103] ​Step S27: Based on the forward transmission hidden state parameters and the backward transmission hidden state parameters, determine the bidirectional transmission hidden state parameters of the current training time step, and train the bidirectional gated recurrent unit based on the bidirectional transmission hidden state parameters.

[0104] In the embodiment of the present disclosure, firstly, R (T) =σ(X (T) W xr +H (T-1) W hr +b r ) Construct a reset gate for the input data. By formula 13 Z (T) =σ(X (T) W xz +H (T-1) W hz +b z ) constructs an update gate for the input data.

[0105] In Formula 12 and Formula 13, for a given time step T, R (T) To reset the gate, Z (T) is the update gate, X (T) is the input data, W xr 、W xz and W hr 、W hz is the weight matrix, b r 、b z is the node bias.

[0106] Here, we can also use formula 14 Construct candidate hidden states for the input data.

[0107] In formula 14, for a given time step T, W xh 、W hh is the weight matrix, b h is the node bias, the symbol ⊙ is the Hadamard product operator, H (T-1) is the hidden state at time step T-1.

[0108] Here, we can also use formula 15 Update the hidden state. Combined with the update gate Z (T) The result further determines the new hidden state H (T) and Then, if Figure 6 As shown, it can be obtained by formula 16 and seventeen Determine the forward transmission hidden layer and the backward transmission hidden layer; then, you can use formula 18 The forward transmission hidden layer and the backward transmission hidden layer are added on the basis of the gated recurrent unit to construct a bidirectional gated recurrent unit.

[0109] As shown in formula sixteen, formula seventeen and formula eighteen, Figure 6 is a forward transmission hidden layer representation of the T time step, is a backward transmission hidden layer representation of the T time step, H (T) is an implicit state of the T time step output, X (T) is the input data of the T time step. In formula eighteen, W1 T is a weight matrix of the forward transmission hidden layer of the T time step, is a weight matrix of the backward transmission hidden layer of the T time step, b (T) is a node bias of the T time step.

[0110] In the training phase of the bidirectional gated recurrent unit, the hidden state parameters of each training time step T can be determined according to formula twelve to formula fifteen; then, the attention data in the training and the hidden state parameters of the next training time step are processed through formula sixteen to obtain the forward transmission hidden state parameters of each training time step; and the attention data in the training and the hidden state parameters of the next training time step are processed through seventeen to obtain the backward transmission hidden state parameters of each training time step; finally, the forward transmission hidden state parameters and the backward transmission hidden state parameters are processed through formula eighteen to obtain the bidirectional transmission hidden state parameters of each training time step. Then, the bidirectional gated recurrent unit can be trained by the bidirectional transmission hidden state parameters and the sample labels of the training samples.

[0111] After obtaining the denoising encoder, the attention mechanism module and the bidirectional gated recurrent unit, the denoising encoder, the attention mechanism module and the bidirectional gated recurrent unit can be combined together to establish an internal combustion engine fault early warning and diagnosis model (i.e., a fault recognition model) based on the attention mechanism and the bidirectional gated recurrent unit, so as to realize the fault early warning and diagnosis of the internal combustion engine.

[0112] After the denoising encoder, the attention mechanism module and the bidirectional gated recurrent unit are combined together, a binary classification and a multi-classification model can be established by using formula nineteen and formula twenty respectively, so as to train the model after combination by using formula nineteen and formula twenty, thereby obtaining a complete internal combustion engine fault early warning and diagnosis network model. As shown in formula nineteen, Figure 7 As shown in the curve diagram of the model loss function of the fault recognition model, Figure 8 As shown in the curve diagram of the recognition accuracy of the fault recognition model.

[0113] Formula nineteen:

[0114] ​

[0115] Equation twenty:

[0116]

[0117] In Equation nineteen and Equation twenty, L bp represents a binary cross-entropy function, L mp represents a multi-class cross-entropy function, is the probability that the model predicts the sample to be a positive example, y is the sample label, and takes the value 1 if the sample belongs to a positive example, otherwise takes the value 0.

[0118] In the embodiments of the present disclosure, after the fault identification model is trained in the manner described above, the steps S101 to S104 described above can be performed.

[0119] In step S101, the plurality of state parameters can be grouped according to the grouping manner of the plurality of parameter sets, thereby obtaining a plurality of parameter groups. Then, the first deep features of each of the parameter groups can be extracted by the denoising encoder.

[0120] In an optional implementation, in the case where the denoising encoder includes a plurality of encoders, the step S102 of extracting the first deep features of each of the parameter groups by the denoising encoder specifically includes the following steps:

[0121] Step S1021: determining the encoder corresponding to each of the parameter groups;

[0122] Step S1022: extracting the first deep features of the corresponding parameter groups by the encoder to obtain the first deep features of each of the parameter groups, and fusing the first deep features of the parameter groups to process the fused first deep features by the attention mechanism module.

[0123] In the embodiments of the present disclosure, the encoder corresponding to each of the parameter groups can be determined in the plurality of encoders; then, the first deep features of the corresponding parameter groups are extracted by the encoders in the manner described in Equation nine to obtain the first deep features of each of the parameter groups.

[0124] After obtaining the first deep features of each of the parameter groups, the first deep features of the parameter groups can be fused to obtain the fused first deep features. Then, the fused first deep features can be processed by the attention mechanism module.

[0125] It can be known from the above description that the attention mechanism module includes the pooling layer, the first full connection layer and the second full connection layer. On this basis, the step S103 processes the first deep feature through the attention mechanism module to obtain attention data, specifically including the following steps:

[0126] Step S1031: performing average pooling processing on the first deep feature through the pooling layer to obtain a first pooling result, and performing maximum pooling processing on the first deep feature to obtain a second pooling result;

[0127] Step S1032: performing full connection processing on the first pooling result through the first full connection layer to obtain a first full connection result, and performing full connection processing on the second pooling result through the second full connection layer to obtain a second full connection result;

[0128] Step S1033: merging the first full connection result and the second full connection result to obtain a merged full connection result;

[0129] Step S1034: processing the merged full connection result through a Sigmoid function to obtain a feature weight of the first deep feature, and determining the attention data based on the feature weight.

[0130] In the embodiments of the present disclosure, as shown in Figure 5 After the first deep feature is extracted, the first deep feature of each channel can be processed through an average pooling layer and a maximum pooling layer respectively to obtain a first pooling result and a second pooling result.

[0131] Here, the average pooling operation and the maximum pooling operation on the first deep feature can be performed through formula eleven W1=f([MLP(AvgPool(X)),MLP(MaxPool(X))]).

[0132] Then, as shown in Figure 5 , the first full connection layer can be used to perform twice full connection operation on the first pooling result to obtain a first full connection result, and the second full connection layer can be used to perform twice full connection operation on the second pooling result to obtain a second full connection result. Next, as shown in Figure 5 , the first full connection result and the second full connection result can be merged through an Add layer to obtain a merged full connection result, and then the merged full connection result is processed through a Sigmoid function to obtain a feature weight.

[0133] In formula eleven, f is a function expression of Sigmoid, f(x)=1 / (1+e -x ), and the range of the feature weight is between 0 and 1.

[0134] After the characteristic weight and the first deep layer feature are multiplied, the attention data processed by the channel attention mechanism module can be obtained.

[0135] In the embodiments of the present disclosure, by applying the channel attention mechanism, the corresponding characteristic weight of each feature in the deep layer feature can be calculated, and then the contribution degree of different features in the deep layer feature to the fault identification is represented by the characteristic weight, so as to highlight the feature part that is helpful to the model training result and ignore the feature part that is not helpful to the model training result, thereby improving the recognition accuracy of the fault identification model.

[0136] In an optional implementation, the step S104 processes the attention data by a bidirectional gated recurrent unit to obtain the hidden state parameter of the current time step, and specifically includes the following steps:

[0137] Step S1041: determining reset gate data of the current time step based on the attention data, and determining update gate data of the current time step based on the attention data;

[0138] Step S1042: determining candidate hidden state parameter of the current time step based on the reset gate data;

[0139] Step S1043: updating the candidate hidden state parameter based on the update gate data, and obtaining the hidden state parameter after the updating.

[0140] In the embodiments of the present disclosure, the attention data can be processed by formula twelve R (T) =σ(X (T) W xr +H (T-1) W hr +b r ) to obtain the reset gate data. Then, the attention data can be processed by formula thirteen Z (T) =σ(X (T )W xz +H (T-1) W hz +b z ) to obtain the update gate data.

[0141] Next, the attention data can be processed by formula fourteen to obtain the candidate hidden state parameter. Then, the candidate hidden state parameter can be updated by formula fifteen to obtain the hidden state parameter after the updating. The hidden state parameter can be used to identify the fault of the internal combustion engine, for example, the internal combustion engine is in a fault operating state, or the internal combustion engine is in a normal operating state.

[0142] The internal combustion engine fault identification method will be described below Figure 9 The internal combustion engine fault identification method will be described below Figure 9 The internal combustion engine fault identification method will be described below

[0143] S901: Collect data of the internal combustion engine under different operating conditions to obtain training samples.

[0144] S902: Group multiple state parameters by the independent variable grouping method to obtain multiple parameter groups, for example, training parameter group 1 to training parameter group n.

[0145] S903: Extract deep features of each training parameter group by the denoising encoder to obtain deep feature 1 to deep feature n.

[0146] S904: Fuse the deep features of each training parameter group to obtain fused deep features.

[0147] S905: Process the fused deep features by the attention mechanism module to obtain attention data.

[0148] S906: Process the attention data by the bidirectional gated recurrent unit to obtain hidden state parameters of the current time step.

[0149] S907: Perform full connection processing on the hidden state parameters by the full connection layer to obtain a full connection processing result.

[0150] S908: Process the full connection processing result and the classification label of the training sample by the softmax classification layer, and then train the fault identification model according to the processing result to obtain a trained fault identification model.

[0151] S909: Obtain multiple state parameters of the internal combustion engine, group the multiple state parameters to obtain multiple parameter groups, and process the multiple parameter groups by the trained fault identification model to obtain a fault identification result of the internal combustion engine.

[0152] In the embodiments of the present disclosure, by dividing the multiple state parameters into multiple parameter groups, the complexity of the model can be reduced; by extracting first deep features of each parameter group by the denoising encoder, processing the first deep features by the attention mechanism module to obtain attention data, processing the attention data by the bidirectional gated recurrent unit to obtain hidden state parameters of the current time step, and identifying the fault of the internal combustion engine according to the hidden state parameters, the health status of the internal combustion engine can be automatically diagnosed and evaluated, and compared with the diagnosis by threshold comparison in the prior art, the diagnosis accuracy of the fault diagnosis can be improved, which has important theoretical value and engineering application prospect for improving the operation and maintenance level of the internal combustion engine

[0153] Those skilled in the art can understand that, in the above method of the specific embodiment, the writing order of each step does not mean a strict execution order and does not constitute any limitation on the implementation process. The specific execution order of each step should be determined by its function and possible internal logic.

[0154] Based on the same inventive concept, the disclosure embodiments also provide an internal combustion engine fault identification device corresponding to the internal combustion engine fault identification method. Since the principle of solving problems in the device of the disclosure embodiments is similar to the above-mentioned internal combustion engine fault identification method of the disclosure embodiments, the implementation of the device can be referred to the implementation of the method, and the repeated parts will not be described here.

[0155] Embodiment two

[0156] Referring to Figure 10 Fig. 1 shows an architecture schematic diagram of an internal combustion engine fault identification device provided by the disclosure embodiments, the device comprises a grouping unit 10, an extraction unit 20, a feature processing unit 30 and an identification unit 40; wherein,

[0157] The grouping unit 10 is configured to obtain a plurality of state parameters of the internal combustion engine in a current time step, and group the plurality of state parameters to obtain a plurality of parameter groups.

[0158] The extraction unit 20 is configured to extract first deep features of each of the parameter groups through a denoising encoder.

[0159] The feature processing unit 30 is configured to process the first deep features through an attention mechanism module to obtain attention data; wherein the attention data is used to indicate the contribution of the first deep features to fault identification.

[0160] The identification unit 40 is configured to process the attention data through a bidirectional gated recurrent unit to obtain a hidden state parameter of the current time step, and perform fault identification on the internal combustion engine according to the hidden state parameter; wherein the hidden state parameter is used to indicate the running state of the internal combustion engine.

[0161] In a possible implementation, in the case where the denoising encoder comprises a plurality of encoders, the extraction unit 20 is further configured to: determine an encoder corresponding to each of the parameter groups; extract the first deep features of the corresponding parameter groups through the encoders to obtain the first deep features of each of the parameter groups, and fuse the first deep features of the parameter groups to process the fused first deep features through the attention mechanism module.

[0162] In a possible implementation, in the case that the attention mechanism module comprises a pooling layer, a first full connection layer and a second full connection layer; the feature processing unit 30 is further configured to: perform average pooling processing on the first deep layer feature through the pooling layer to obtain a first pooling result, and perform maximum pooling processing on the first deep layer feature to obtain a second pooling result; perform full connection processing on the first pooling result through the first full connection layer to obtain a first full connection result, and perform full connection processing on the second pooling result through the second full connection layer to obtain a second full connection result; combine the first full connection result and the second full connection result to obtain a combined full connection result; perform processing on the combined full connection result through a Sigmoid function to obtain a feature weight of the first deep layer feature, and determine the attention data based on the feature weight.

[0163] In a possible implementation, the identifying unit 40 is further configured to: determine reset gate data of the current time step based on the attention data, and determine update gate data of the current time step based on the attention data; determine a candidate hidden state parameter of the current time step based on the reset gate data; update the candidate hidden state parameter based on the update gate data, and obtain the hidden state parameter after the update.

[0164] In a possible implementation, the apparatus is further configured to train the denoising encoder in the following manner: extract, through an initial encoder, a second deep layer feature of a corresponding training parameter group in a training sample; wherein the training sample comprises a plurality of training parameter groups, each training parameter group comprising a plurality of parameter sets, and a same parameter set comprising state parameters of parameters of a same type in various working conditions; perform decoding processing on the second deep layer feature extracted by the initial encoder through an initial decoder to obtain a decoding result; perform self-training on the initial decoder and the initial encoder based on a difference between the decoding result and the corresponding training parameter group, and obtain the denoising encoder after the training.

[0165] In a possible implementation, the apparatus is further configured to train the bidirectional gated recurrent unit in the following manner: obtain training attention data obtained after the attention mechanism module processes the second deep layer feature; determine a forward transmission hidden state parameter of a current training time step based on the training attention data and a hidden state parameter of a previous training time step; determine a backward transmission hidden state parameter of the current training time step based on the training attention data and a hidden state parameter of a next training time step; determine a bidirectional transmission hidden state parameter of the current training time step based on the forward transmission hidden state parameter and the backward transmission hidden state parameter, and train the bidirectional gated recurrent unit based on the bidirectional transmission hidden state parameter.

[0166] In a possible implementation, the apparatus is further configured to determine the plurality of training parameter groups by: determining mutual information values between the plurality of parameter sets; performing parameter classification processing on the plurality of parameter sets based on the mutual information values to obtain a classification result; wherein the classification result includes a plurality of classification groups, and mutual information values between state parameters in a same classification group satisfy a grouping requirement; and determining the plurality of training parameter groups based on the plurality of classification groups in the classification result; wherein each classification group corresponds to a training parameter group.

[0167] In a possible implementation, the apparatus is further configured to: determine a target information entropy of each parameter set; determine a joint information entropy between any two parameter sets; and determine mutual information values between the plurality of parameter sets based on the target information entropy and the joint information entropy.

[0168] The description of the processing procedure of each module in the apparatus and the interaction procedure between the modules can refer to the related description in the method embodiments, and will not be repeated here.

[0169] Corresponding to the internal combustion engine fault identification method in Figure 1 The embodiments of the present disclosure also provide a computer device 1100, as shown in Figure 11 The computer device 1100 provided by the embodiments of the present disclosure is shown in a structural schematic diagram, which includes:

[0170] a processor 111, a memory 112, and a bus 113; the memory 112 is configured to store execution instructions, including an internal memory 1121 and an external memory 1122; the internal memory 1121 is also called an internal storage, and is configured to temporarily store operation data in the processor 111 and exchange data with the external memory 1122 such as a hard disk; the processor 111 exchanges data with the external memory 1122 through the internal memory 1121; when the computer device 1100 is running, the processor 111 and the memory 112 communicate through the bus 113, so that the processor 111 executes the following instructions:

[0171] obtain a plurality of state parameters of an internal combustion engine at a current time step, and group the plurality of state parameters to obtain a plurality of parameter groups;

[0172] extract first deep features of each parameter group through a denoising encoder;

[0173] process the first deep features through an attention mechanism module to obtain attention data; wherein the attention data is configured to indicate a contribution degree of the first deep features to fault identification;

[0174] The attention data is processed by a bidirectional gated recurrent unit to obtain a hidden state parameter of the current time step, and the internal combustion engine is identified according to the hidden state parameter.

[0175] The disclosure also provides a computer readable storage medium, which stores a computer program. When the computer program is run by a processor, the steps of the internal combustion engine fault identification method described in the above method embodiments are executed. The storage medium can be a volatile or non-volatile computer readable storage medium.

[0176] The disclosure also provides a computer program product, which carries a program code. The instructions included in the program code can be used to execute the steps of the internal combustion engine fault identification method described in the above method embodiments. For details, refer to the above method embodiments, which will not be described here.

[0177] The disclosure also provides a computer program, which, when executed by a processor, implements any one of the methods of the preceding embodiments. The computer program product can be specifically implemented by hardware, software or a combination thereof. In an optional embodiment, the computer program product is specifically embodied as a computer storage medium, and in another optional embodiment, the computer program product is specifically embodied as a software product, such as a software development kit (Software Development Kit, SDK) and the like.

[0178] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-described system and device can refer to the corresponding process in the above method embodiments, which will not be described here. In several embodiments provided by the disclosure, it should be understood that the disclosed system, device and method can be implemented by other ways. The above-described device embodiments are only schematic, for example, the division of the units is only a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some communication interface, device or unit, which can be electrical, mechanical or other forms.

[0179] The units described as separate components may or may not be physically separate, and the components displayed as units may or may not be physical units, i.e. may be located in one place, or may be distributed on multiple network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiment scheme.

[0180] In addition, each functional unit in various embodiments of the present disclosure can be integrated into one processing unit, or each unit can exist physically, or two or more units can be integrated into one unit.

[0181] If the functions are realized in the form of software functional units and sold or used as independent products, they can be stored in a non-volatile computer readable storage medium executable by a processor. Based on this understanding, the technical solutions of the present disclosure essentially or the part of the prior art or the part of the technical solutions can be embodied in the form of a software product, which is stored in a storage medium and includes a number of instructions for making a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the method described in various embodiments of the present disclosure. The foregoing storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), magnetic disk or optical disk, and various program code storage media.

[0182] Finally, it should be noted that: the above-described embodiments are only specific embodiments of the present disclosure, used to illustrate the technical solutions of the present disclosure, and not to limit them, the protection scope of the present disclosure is not limited thereto, although the present disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand: any person skilled in the art in the technical range disclosed by the present disclosure, it can still modify or easily think of changes to the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to part of the technical features; and these modifications, changes or replacements do not make the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present disclosure, and all should be covered in the protection scope of the present disclosure. Therefore, the protection scope of the present disclosure should be subject to the protection scope of the claims.

Claims

1. An internal combustion engine failure recognition method characterized by comprising: The method comprises: obtaining a plurality of state parameters of the internal combustion engine at a current time step, and grouping the plurality of state parameters to obtain a plurality of parameter groups; extracting first deep features of each of the parameter groups through a denoising encoder; processing the first deep features through an attention mechanism module to obtain attention data, wherein the attention data is used to indicate the contribution of the first deep features to fault identification; processing the attention data through a bidirectional gated recurrent unit to obtain a hidden state parameter of the current time step, and identifying a fault of the internal combustion engine according to the hidden state parameter, wherein the hidden state parameter is used to indicate the operating state of the internal combustion engine; the denoising encoder comprises a plurality of encoders; the first deep features of each of the parameter groups are extracted through the denoising encoder, comprising: determining an encoder corresponding to each of the parameter groups; extracting the first deep features of the corresponding parameter groups through the encoder to obtain the first deep features of each of the parameter groups, and fusing the first deep features of the parameter groups to process the fused first deep features through the attention mechanism module; the plurality of training parameter groups are determined in the following manner: determining mutual information values between the plurality of parameter sets; performing parameter classification processing on the plurality of parameter sets based on the mutual information values to obtain a classification result, wherein the classification result comprises a plurality of classification groups, and the mutual information values between the state parameters in the same classification group satisfy the grouping requirement; determining the plurality of training parameter groups based on the plurality of classification groups in the classification result, wherein each of the classification groups corresponds to one of the training parameter groups; the mutual information values between the plurality of parameter sets are determined in the following manner: determining a target information entropy of each of the parameter sets; determining a joint information entropy between any two of the parameter sets; determining the mutual information values between the plurality of parameter sets based on the target information entropy and the joint information entropy.

2. The method of claim 1, wherein, The attention mechanism module comprises a pooling layer, a first fully connected layer and a second fully connected layer; the first deep features are processed through the attention mechanism module to obtain the attention data, comprising: performing average pooling processing on the first deep features through the pooling layer to obtain a first pooling result, and performing maximum pooling processing on the first deep features to obtain a second pooling result; performing fully connected processing on the first pooling result through the first fully connected layer to obtain a first fully connected result, and performing fully connected processing on the second pooling result through the second fully connected layer to obtain a second fully connected result; merging the first fully connected result and the second fully connected result to obtain a merged fully connected result; processing the merged fully connected result through a Sigmoid function to obtain a feature weight of the first deep features, and determining the attention data based on the feature weight.

3. The method of claim 1, wherein, the attention data is processed through the bidirectional gated recurrent unit to obtain the hidden state parameter of the current time step, comprising: determine reset gate data of the current time step based on the attention data, and determine update gate data of the current time step based on the attention data; determine candidate hidden state parameters of the current time step based on the reset gate data; update the candidate hidden state parameters based on the update gate data, to obtain the hidden state parameters.

4. The method of claim 1, wherein, The method further comprises training the denoising encoder by: extracting second deep features of a corresponding training parameter group in a training sample through an initial encoder; wherein the training sample comprises a plurality of training parameter groups, each training parameter group comprising a plurality of parameter sets, and a same parameter set comprising state parameters of parameters of a same type under various working conditions; decoding the second deep features extracted by the initial encoder through an initial decoder to obtain a decoding result; training the initial decoder and the initial encoder based on differences between the decoding result and the corresponding training parameter group, to obtain the denoising encoder.

5. The method of claim 4, wherein, The method further comprises training the bidirectional gated recurrent unit by: obtaining training attention data obtained by processing the second deep features through an attention mechanism module; determining forward transmission hidden state parameters of a current training time step based on the training attention data and hidden state parameters of a previous training time step; determining backward transmission hidden state parameters of the current training time step based on the training attention data and hidden state parameters of a next training time step; determining bidirectional transmission hidden state parameters of the current training time step based on the forward transmission hidden state parameters and the backward transmission hidden state parameters, and training the bidirectional gated recurrent unit based on the bidirectional transmission hidden state parameters.

6. An internal combustion engine failure recognition device characterized by comprising: comprise: a grouping unit configured to obtain a plurality of state parameters of an internal combustion engine in a current time step, and group the plurality of state parameters to obtain a plurality of parameter groups; an extraction unit configured to extract first deep features of each of the parameter groups through a denoising encoder; a feature processing unit configured to process the first deep features through an attention mechanism module to obtain attention data; wherein the attention data is used to indicate a contribution degree of the first deep features to fault identification; an identification unit configured to process the attention data through a bidirectional gated recurrent unit to obtain hidden state parameters of the current time step, and identify a fault of the internal combustion engine according to the hidden state parameters; wherein the hidden state parameters are used to indicate an operating state of the internal combustion engine. The denoising encoder comprises a plurality of encoders; and the extraction unit is specifically configured to: determine an encoder corresponding to each of the parameter groups; extract the first deep features of the corresponding parameter groups through the encoders to obtain the first deep features of each of the parameter groups, and fuse the first deep features of the parameter groups to process the fused first deep features through the attention mechanism module; determine the plurality of training parameter groups by: determining mutual information values between the plurality of parameter sets; The mutual information values between the plurality of parameter sets are determined, including: The target information entropy of each parameter set is determined; The joint information entropy between any two parameter sets is determined; The mutual information values between the plurality of parameter sets are determined based on the target information entropy and the joint information entropy. The mutual information values between the plurality of parameter sets are determined, including: The target information entropy of each parameter set is determined; 7. A computer device, comprising: The joint information entropy between any two parameter sets is determined; The mutual information values between the plurality of parameter sets are determined based on the target information entropy and the joint information entropy. The mutual information values between the plurality of parameter sets are determined, including: The target information entropy of each parameter set is determined; The joint information entropy between any two parameter sets is determined; The mutual information values between the plurality of parameter sets are determined based on the target information entropy and the joint information entropy. The mutual information values between the plurality of parameter sets are determined, including: The target information entropy of each parameter set is determined; The joint information entropy between any two parameter sets is determined; The mutual information values between the plurality of parameter sets are determined based on the target information entropy and the joint information entropy. The mutual information values between the plurality of parameter sets are determined, including: The target information entropy of each parameter set is determined; The joint information entropy between any two parameter sets is determined; The mutual information values between the plurality of parameter sets are determined based on the target information entropy and the joint information entropy. The mutual information values between the plurality of parameter sets are determined, including: The target information entropy of each parameter set is determined; The joint information entropy between any two parameter sets is determined; The mutual information values between the plurality of parameter sets are determined based on the target information entropy and the joint information entropy. The mutual information values between the plurality of parameter sets are determined, including: The target information entropy of each parameter set is determined; The joint information entropy between any two parameter sets is determined; The mutual information values between the plurality of parameter sets are determined based on the target information entropy and the joint information entropy. The mutual information values between the plurality of parameter sets are determined, including: The target information entropy of each parameter set is determined; The joint information entropy between any two parameter sets is determined; The mutual information values between the plurality of parameter sets are determined based on the target information entropy and the joint information entropy. The mutual information values between the plurality of parameter sets are determined, including: The target information entropy of each parameter set is determined; The joint

Images