Fault diagnosis method and device for fuel injector and vehicle

Abstract

The invention relates to the technical field of fuel injector fault diagnosis, in particular to a fault diagnosis method and device for a fuel injector and a vehicle. The method comprises the steps that vibration data of an oil injector are obtained, to-be-recognized data are obtained, and the to-be-recognized data are the vibration data different from the working condition with label data; adaptively training a to-be-trained fault recognition model based on improved joint distribution by using the to-be-recognized data and the labeled data to obtain a trained fault recognition model; and inputting the to-be-identified data into the fault identification model to obtain a fault identification result. The to-be-trained fault recognition model is trained through the labeled data, the to-be-recognized data different from the labeled data in working condition and the improved joint distribution adaptation, and the to-be-trained fault recognition model can be trained according to different working conditions, so that the accuracy of fault recognition is improved.

Description

Technical Field

[0001] The present application relates to the technical field of fuel injector fault diagnosis, and in particular to a fuel injector fault diagnosis method, device and vehicle. Background Art

[0002] Traditional fault diagnosis methods mainly rely on expert experience and simple signal processing technology, and have the disadvantages of low diagnostic accuracy, poor real-time performance, and weak adaptability to changes in operating conditions. In practical applications, high-pressure common rail injectors often work under variable operating conditions, including pressure changes, load changes, speed fluctuations, fuel quality differences, and ambient temperature changes. These variable operating conditions have a significant impact on the working state and fault characteristics of high-pressure common rail injectors, increasing the complexity and difficulty of fault diagnosis. Therefore, research on fault diagnosis technology for high-pressure common rail injectors under variable operating conditions is particularly important.

[0003] In summary, there is a need to provide a fault diagnosis method, device and vehicle for a fuel injector that can improve the accuracy of fault diagnosis under different operating conditions. Summary of the invention

[0004] To solve the above problems, the present application proposes a fault diagnosis method, device and vehicle for a fuel injector.

[0005] On the one hand, the present application proposes a fault diagnosis method for a fuel injector, comprising the following steps:

[0006] Acquire vibration data of the injector to obtain data to be identified, wherein the data to be identified is the vibration data of a working condition different from that of the labeled data;

[0007] Adaptively training the fault recognition model to be trained using the data to be recognized and the labeled data based on the improved joint distribution to obtain the trained fault recognition model;

[0008] The data to be identified is input into the fault identification model to obtain a fault identification result.

[0009] Further, in the above-mentioned method for fault diagnosis of a fuel injector, the method of using the data to be identified and the labeled data to adaptively train the fault identification model to be trained based on the improved joint distribution to obtain the trained fault identification model includes:

[0010] Inputting the data to be identified and the labeled data into a feature extraction unit in the fault identification model to be trained, and acquiring features of the data to be identified and the labeled data respectively based on an attention mechanism to obtain a first feature and a second feature, wherein the first feature is a feature of the data to be identified, and the second feature is a feature of the labeled data;

[0011] Inputting the first feature and the second feature into a fault prediction unit to obtain a first prediction result and a second prediction result, wherein the first prediction result is a prediction result obtained according to the first feature, and the second prediction result is a prediction result obtained according to the second feature;

[0012] Inputting the first feature, the second feature, a label and a pseudo label into a domain adaptation unit, and determining a domain adaptation loss based on the improved joint distribution adaptation, wherein the label is a label of the labeled data, and the pseudo label is the first prediction result obtained by the fault prediction unit in a previous prediction of the first feature;

[0013] The model to be trained is updated according to the domain adaptation loss and the classification loss to obtain the trained fault recognition model.

[0014] Furthermore, the above-mentioned fault diagnosis method for a fuel injector, before updating the model to be trained according to the domain adaptation loss and the classification loss to obtain the trained fault identification model, further comprises:

[0015] The classification loss is determined according to the second prediction result and the label.

[0016] Further, in the above-mentioned fault diagnosis method for a fuel injector, the step of inputting the first feature, the second feature, the label and the pseudo label into a domain adaptation unit, and determining the domain adaptation loss based on the improved joint distribution adaptation comprises:

[0017] The domain adaptation loss is determined according to the first feature, the second feature, the label, the pseudo label and the improved joint distribution adaptation, and the calculation method is as follows:

[0018]

[0019] Wherein, IJDA represents the domain adaptation loss, s represents the labeled data, t represents the data to be identified, and Y s represents a labeled data set including the labels of the labeled data, Y t A pseudo-label dataset including the pseudo-labels of the data to be identified, Indicates the characteristics of the calculated data set. Represents the first feature, represents the second feature, E(·) represents the expectation of the data, represents the expectation of calculating the second characteristic marginal probability distribution, represents the expectation of calculating the marginal probability distribution of the first feature, It means calculating the expected probability of the second feature occurring when the category is c. It means to calculate the expected probability of the first feature occurring when the category is c, P(X s ) represents the marginal probability distribution of the data set with labeled data, P(Y s =c) represents the category prior distribution of the label, P(Y t =c) represents the category prior distribution of the pseudo label, c represents the data category, X s represents the first data set including the data to be identified, X t Represents a second data set including the labeled data.

[0020] Furthermore, in the above-mentioned fault diagnosis method for a fuel injector, the updating of the model to be trained according to the domain adaptation loss and the classification loss to obtain the trained fault identification model includes:

[0021] updating the feature extraction unit, the fault prediction unit, and the domain adaptation unit according to the domain adaptation loss and the classification loss;

[0022] Continue to execute the step of inputting the data to be identified and the labeled data into the feature extraction unit in the fault identification model to be trained, and respectively acquiring the features of the data to be identified and the labeled data based on the attention mechanism to obtain the first feature and the second feature, until a preset training threshold is reached, or the domain adaptation loss and / or the classification loss is less than or equal to a preset loss threshold, to obtain the trained fault identification model including the feature extraction unit and the fault prediction unit.

[0023] Furthermore, in the above-mentioned method for fault diagnosis of a fuel injector, the features of the data to be identified and the labeled data are respectively acquired based on the attention mechanism to obtain the first feature and the second feature, including:

[0024] Based on the convolution block attention module, a plurality of frequency features of the data to be identified and the labeled data are respectively extracted to obtain a plurality of first frequency features corresponding to the data to be identified and a plurality of second frequency features corresponding to the labeled data;

[0025] The first frequency features are combined and then features are extracted again to obtain a first feature, and the second frequency features are combined and then features are extracted again to obtain a second feature.

[0026] Furthermore, in the above-mentioned fault diagnosis method for a fuel injector, the step of obtaining vibration data of the fuel injector to obtain data to be identified includes:

[0027] Acquiring the vibration data of the injector;

[0028] Perform working condition judgment on the vibration data to determine the data to be identified.

[0029] In a second aspect, the present application proposes a fault diagnosis device for a fuel injector, comprising: a data acquisition module and a fault identification module;

[0030] The data acquisition module is used to acquire vibration data of the injector to obtain data to be identified;

[0031] The fault identification module is used to use the data to be identified and the labeled data to adaptively train the fault identification model to be trained based on the improved joint distribution to obtain the trained fault identification model; and input the data to be identified into the fault identification model to obtain a fault identification result.

[0032] Further, in the above-mentioned fault diagnosis device for a fuel injector, the fault identification module comprises: a feature extraction unit, a fault prediction unit, a domain adaptation unit and a training unit;

[0033] The feature extraction unit is used to input the data to be identified and the labeled data into the feature extraction unit of the fault identification model to be trained, and obtain the features of the data to be identified and the labeled data respectively based on the attention mechanism to obtain a first feature and a second feature, wherein the first feature is the feature of the data to be identified, and the second feature is the feature of the labeled data;

[0034] The fault prediction unit is used to input the first feature and the second feature into the fault prediction unit to obtain a first prediction result and a second prediction result, wherein the first prediction result is a prediction result obtained according to the first feature, and the second prediction result is a prediction result obtained according to the second feature;

[0035] The domain adaptation unit is configured to input the first feature, the second feature, a label and a pseudo label into the domain adaptation unit, and determine a domain adaptation loss based on the improved joint distribution adaptation, wherein the label is a label of the labeled data, and the pseudo label is the first prediction result obtained by the fault prediction unit in predicting the first feature in a previous time;

[0036] The training unit is used to update the model to be trained according to the domain adaptation loss and the classification loss to obtain the trained fault recognition model.

[0037] In a third aspect, the present application proposes a vehicle, comprising an on-board terminal, wherein the on-board terminal is used to execute a fault diagnosis method for a fuel injector as described in any one of the first aspects.

[0038] The advantages of the present application are that by training the fault recognition model to be trained with labeled data, data to be identified with working conditions different from the labeled data, and improved joint distribution adaptation, the fault recognition model to be trained can be trained for different working conditions, thereby improving the accuracy of fault recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] By reading the detailed description of the preferred embodiment below, various other advantages and benefits will become clear to those of ordinary skill in the art. The accompanying drawings are only used for the purpose of illustrating the preferred embodiment and are not to be considered as limiting the present application. Moreover, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:

[0040] Figure 1 is a schematic diagram of a fault diagnosis method for a fuel injector provided by the present application;

[0041] Figure 2 It is a schematic diagram of a process of a fault diagnosis method for a fuel injector provided by the present application;

[0042] Figure 3 is a schematic diagram of a feature extraction unit for a fault diagnosis method for a fuel injector provided by the present application;

[0043] Figure 4 is a schematic diagram of another fault diagnosis method for a fuel injector provided by the present application;

[0044] Figure 5 is a schematic diagram of a fault diagnosis device for a fuel injector provided by the present application;

[0045] Figure 6 It is a schematic diagram of another fault diagnosis device for a fuel injector provided in the present application. DETAILED DESCRIPTION

[0046] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0047] Existing fault identification methods still face challenges such as insufficient feature extraction and insufficient model generalization when dealing with complex operating conditions and diverse fault modes. Especially under variable operating conditions, traditional models often have difficulty adapting to the changes in fault characteristics under different operating conditions, resulting in a significant reduction in diagnostic accuracy and robustness.

[0048] To solve the above problems, the embodiments of the present application propose a fault diagnosis method, device and vehicle for an injector. The fault recognition model to be trained is trained by adapting labeled data, data to be identified under working conditions different from those of the labeled data, and an improved joint distribution. The fault recognition model to be trained can be trained for different working conditions, thereby improving the accuracy and robustness of fault recognition.

[0049] In the first aspect, according to the implementation mode of the present application, a fault diagnosis method for a fuel injector is proposed, such as Figure 1 As shown, the following steps are included:

[0050] S101, acquiring vibration data of the injector to obtain data to be identified, wherein the data to be identified is vibration data of a working condition different from that of the labeled data;

[0051] S102, using the data to be identified and the labeled data to adaptively train the fault identification model to be trained based on the improved joint distribution to obtain a trained fault identification model;

[0052] S103, inputting the data to be identified into a fault identification model to obtain a fault identification result.

[0053] like Figure 2 As shown, the vibration data of the fuel injector is obtained to obtain the data to be identified, including: obtaining the vibration data of the fuel injector; and performing working condition judgment on the vibration data to determine the data to be identified.

[0054] Among them, the working conditions include the pressure during operation (working pressure). The vibration data can be classified according to the working pressure of the vibration data. When the working pressure of the vibration data does not belong to the working pressure of the labeled data, the vibration data is used as data to be identified. That is, assuming that the working pressure of the labeled data includes 70 MPa and 90 MPa, if the working pressure of the vibration data is 70 MPa, then the vibration data is not used as data to be identified, and if the working pressure of the vibration data is 80 MPa, then the vibration data is used as data to be identified.

[0055] like Figure 2 As shown, data is collected by installing high-precision vibration sensors at key locations of the high-pressure common rail injector equipment. The vibration sensors are used to capture vibration signals of the injector equipment under different working conditions in real time and convert the vibration signals into electrical signals for output.

[0056] like Figure 2As shown, the electrical signal of the vibration sensor can be received through the data acquisition board. The board has a multi-channel high-speed acquisition function and can perform high-precision digital processing on the vibration signal. After real-time filtering and combined with adaptive noise reduction technology, it can effectively remove external interference noise and low-frequency interference or other noise sources that may be introduced during the acquisition process, ensuring high-quality output of the acquired signal and improving the clarity and accuracy of the signal. The acquisition board can also dynamically adjust the noise reduction strategy, optimize the noise reduction effect according to the characteristics of the vibration signal under different working conditions, and improve the reliability and stability of the data. The board supports high-resolution digital processing to ensure that the details of the vibration signal are retained with high fidelity. In this process, the board automatically adjusts the filter parameters at the current moment through the filter parameters obtained at the previous moment, and then the filter at the current moment filters and reduces the noise of the acquired digital signal to achieve real-time noise reduction, providing more accurate signal input for subsequent data analysis and fault diagnosis. The digitized vibration signal is transmitted to the data division module, such as the host computer, as vibration data through the data transmission interface.

[0057] like Figure 2 As shown, the data division module includes a data processing unit and a storage unit. The data processing unit performs real-time analysis on the vibration signal (vibration data) according to the working conditions of the high-pressure common rail injector, divides the vibration data into source domain data (labeled data) containing fault categories and target domain data (data to be identified) not containing fault categories according to the working conditions, and saves the divided data to the storage unit for further analysis and fault identification (fault diagnosis).

[0058] Before updating the model to be trained according to the domain adaptation loss and the classification loss to obtain the trained fault recognition model, the method further includes: determining the classification loss according to the second prediction result and the label.

[0059] Wherein, the classification loss includes a cross entropy loss. The classification loss can be determined by a cross entropy loss function according to the second prediction result and the label.

[0060] like Figure 2As shown, the to-be-recognized data and the labeled data are used to train the to-be-recognized fault recognition model based on the improved joint distribution adaptation to obtain a trained fault recognition model, including: inputting the to-be-recognized data and the labeled data into the feature extraction unit in the to-be-recognized fault recognition model, and acquiring the features of the to-be-recognized data and the labeled data based on the attention mechanism respectively, to obtain a first feature and a second feature, wherein the first feature is the feature of the to-be-recognized data, and the second feature is the feature of the labeled data; inputting the first feature and the second feature into the fault prediction unit to obtain a first prediction result and a second prediction result, wherein the first prediction result is a prediction result obtained according to the first feature, and the second prediction result is a prediction result obtained according to the second feature; inputting the first feature, the second feature, the label and the pseudo-label into the domain adaptation unit, and determining the domain adaptation loss based on the improved joint distribution adaptation, wherein the label is the label of the labeled data, and the pseudo-label is the first prediction result obtained by the fault prediction unit in the previous prediction of the first feature; updating the to-be-recognized model according to the domain adaptation loss and the classification loss to obtain a trained fault recognition model.

[0061] Among them, when the first feature and the second feature are predicted for the first time (first time), since the previous first prediction result before the first time is not obtained, that is, there is no previous pseudo label corresponding to the first prediction result obtained for the first time, therefore, when the first feature and the second feature are predicted for the first time, after obtaining the first prediction result and the second prediction result, the step of obtaining the vibration data of the injector and obtaining the data to be identified is continued. That is, when the first feature and the second feature are predicted for the first time, the classification loss is not determined according to the second prediction result and the label, and the model to be trained is not updated according to the domain adaptation loss and the classification loss.

[0062] Updating the model to be trained according to the domain adaptation loss and the classification loss: including updating the feature extraction unit, the fault prediction unit and the domain adaptation unit according to the domain adaptation loss and the classification loss.

[0063] Based on the attention mechanism, the features of the data to be identified and the labeled data are respectively obtained to obtain the first feature and the second feature, including: based on the Convolutional Block Attention Module (CBAM), multiple frequency features of the data to be identified and the labeled data are respectively extracted to obtain multiple first frequency features corresponding to the data to be identified and multiple second frequency features corresponding to the labeled data; after merging the multiple first frequency features, the features are extracted again to obtain the first feature, and after merging the multiple second frequency features, the features are extracted again to obtain the second feature.

[0064] The implementation method of the present application constructs a multi-attention convolutional network as a feature extraction unit to perform feature extraction, uses multi-channel convolution to expand the receptive field, and constructs a one-dimensional multi-scale convolutional neural network model for feature extraction. Among them, since a wider convolution kernel (such as any convolution kernel with a width of 1×64 to 1×128) can capture a wider range of features, it can be used to extract low-frequency features in vibration signals, while a smaller (narrower) convolution kernel (such as 1×3) can be used to capture local details and high-frequency features in vibration signals. The feature extraction unit of the implementation method of the present application includes at least 3 channels, each of which includes 2 convolution blocks, and 1 convolution block is also included after all channels and before pooling.

[0065] like Figure 3 As shown, a feature extraction unit including 3 channels and 7 convolution blocks is taken as an example. Among them, the first channel Ch1 includes the first convolution block 11 and the first second convolution block 12; the second channel Ch2 includes the second first convolution block 21 and the second second convolution block 22; the third channel Ch3 includes the third first convolution block 31 and the third second convolution block 32; after the three channels, a fourth convolution block 40 is also included. Among them, each convolution block includes: a one-dimensional convolution layer 51, a batch normalization layer 52, a convolution attention layer 53 and a nonlinear activation layer 54 in sequence. Among them, the convolution attention layer 53 includes a convolution block attention module for dynamically focusing on the key parts of the features to enhance the network's ability to express the features of vibration signals (vibration data). The attention mechanism is a way to imitate the human attention allocation process, which is usually used in neural networks and machine learning models. In deep learning, the attention mechanism can enable the model to focus on specific parts of the input data in order to better process information. By adding a convolutional attention layer 53 with an attention mechanism after the one-dimensional convolutional layer 51, the channel attention and spatial attention can be jointly applied to further improve the feature extraction capability.

[0066] like Figure 3 As shown, in the implementation manner of the present application, among the convolution kernels of the first convolution block 11, the second convolution block 21 and the third convolution block 31, at least one is a smaller (narrow) convolution kernel (such as 1×3), and the other two are different two wider convolution kernels (such as any convolution kernel with a width of 1×64 to 1×128). The features of different frequencies in the data to be identified or the labeled data are extracted respectively through three channels, and then the features of different frequencies are spliced, and then more sufficient features are extracted through the fourth convolution block. After the fourth convolution block 40, an adaptive average pooling layer (such as a one-dimensional adaptive average pooling layer) 55 and a fully connected layer 56 are also included, wherein the fully connected layer 56 is used to flatten the received features to obtain the first feature or the second feature.

[0067] The implementation method of the present application proposes a high-pressure common rail injector feature extraction method based on multi-attention convolution under variable operating conditions to reduce the confusion between different faults in the target domain data (data to be identified) and extract more comprehensive features under the complex conditions of variable operating conditions.

[0068] By simultaneously inputting the obtained labeled data and the data to be identified into the feature extraction unit with a multi-attention convolutional network for feature extraction, features with more characteristics can be obtained. After that, the joint distribution adaptation is improved by Bayes' theorem, and pseudo labels are used to calculate the conditional distribution to make up for the deficiencies in the joint distribution adaptation, thereby improving the accuracy of fault identification.

[0069] like Figure 4 As shown, the first feature, the second feature, the label, and the pseudo label are input into the domain adaptation unit, and the domain adaptation loss is determined based on the improved joint distribution adaptation, including: determining the domain adaptation loss according to the first feature, the second feature, the label, the pseudo label, and the improved joint distribution adaptation, and the calculation method is as follows:

[0070]

[0071] Among them, IJDA represents domain adaptation loss, s represents labeled data, t represents data to be identified, and Y s Y represents a labeled dataset containing labeled data. t represents a pseudo-label dataset including pseudo-labels of the data to be identified, Indicates the characteristics of the calculated data set. Represents the first feature, represents the second feature, E(·) represents the expectation of the data, represents the expectation of calculating the second characteristic marginal probability distribution, represents the expectation of calculating the marginal probability distribution of the first feature, P(X|Y=c) represents the calculation of the class conditional probability distribution, It means calculating the expected probability of the second feature occurring when the category is c. It means to calculate the expected probability of the first feature occurring when the category is c, P(X s ) represents the marginal probability distribution of the data set with labeled data, P(Y s =c) represents the category prior distribution of the label, P(Y t =c) represents the category prior distribution of pseudo labels, c represents the data category, X s represents the first data set including the data to be identified, X t represents a second data set including labeled data.

[0072] like Figure 4As shown, the model to be trained is updated according to the domain adaptation loss and the classification loss to obtain a trained fault recognition model, including: updating the feature extraction unit, the fault prediction unit and the domain adaptation unit according to the domain adaptation loss and the classification loss; continuing to execute the step of inputting the data to be identified and the labeled data into the feature extraction unit in the fault recognition model to be trained, and respectively acquiring the features of the data to be identified and the labeled data based on the attention mechanism to obtain the first feature and the second feature, until a preset training threshold is reached, or the domain adaptation loss and / or the classification loss is less than or equal to the preset loss threshold, to obtain a trained fault recognition model including the feature extraction unit and the fault prediction unit.

[0073] Among them, when the training round requirement (preset training threshold) is reached or the domain adaptation loss and / or classification loss is less than or equal to the preset loss threshold, the trained multi-attention convolutional network is output as the fault recognition model, otherwise the training continues.

[0074] Afterwards, the obtained data to be identified is input into the trained fault identification model, and the fault diagnosis results under variable working conditions are output.

[0075] Below, the improved joint distribution adaptation determination domain adaptation loss in the embodiment of the present application is further explained.

[0076] First, suppose X s represents the source domain dataset, i.e., the labeled dataset including labeled data, Y s represents the source domain data label set, i.e., the label data set including the labels of the labeled data, X t represents the target domain dataset, i.e., the dataset to be identified including the data to be identified, Y t represents the target domain data label set, that is, the pseudo-label data set including the pseudo-labels of the data to be identified, s represents the data distribution of the source domain, that is, the labeled data, and t represents the data distribution of the target domain, that is, the data to be identified. The objective function of Joint Distribution Adaptation (JDA) is defined as:

[0077]

[0078] in, Indicates the characteristics of the calculated data set. Represents the source domain data characteristics, represents the target domain data feature, c represents the data category, P(X) represents the marginal probability distribution of the calculated data set X, and P(X s ) represents the marginal probability distribution of the source domain data set, E(·) is the expectation of the data, represents the expectation of calculating the marginal probability distribution of source domain data features, represents the expected marginal probability distribution of the target domain data feature, P(X|Y=c) represents the conditional probability distribution of the class, It means calculating the expected probability of occurrence of source domain data features when category c is used. It means calculating the expected probability of occurrence of target domain data features when category is c.

[0079] Secondly, in order to overcome the poor final migration effect caused by the calculation of conditional probability distribution adaptation (CDA) in formula (1) using class conditional probability distribution to approximate the conditional probability distribution, an improved conditional probability distribution adaptation is proposed to adjust the conditional probability distribution in the two domains (source domain and target domain). Using Bayes' theorem, the calculation of conditional probability distribution is expressed by class conditional probability distribution, and its expression is:

[0080]

[0081] Among them, P(Y=c) represents the category prior distribution, which can be calculated by the following formula:

[0082]

[0083] Among them, c represents the category label value, Indicates the size of the c-th category in the source domain (labeled data), It represents the size of the c-th category in the target domain (data to be identified), so the following formula exists.

[0084]

[0085] Among them, Bayes' theorem is an important theorem in probability theory, which describes the probability of another related event occurring under the condition that a certain event is known to occur. This theorem uses the relationship between prior probability and posterior probability to calculate the posterior probability by updating the prior probability. In machine learning and statistical inference, Bayes' theorem is widely used in parameter estimation, classification problems and model prediction.

[0086] In the implementation of the present application, the working conditions corresponding to the data (labeled data) that already has a label (real label) are all used as the source domain, and the working conditions corresponding to the vibration data without a real label are used as the target domain. Therefore, when the vibration data acquired in real time has the same working condition as the source domain data (i.e., labeled data with labels) in the existing (original) database, this data can be used, and only the vibration data obtained under different working conditions is used as the target domain data (data to be identified).

[0087] Afterwards, the improved conditional probability distribution adaptation using equation (2) can be expressed as follows:

[0088]

[0089] Finally, by improving the conditional probability distribution adaptation, the final improved joint distribution adaptation (Improve JointDistribution Adaptation, IJDA) can be defined as:

[0090]

[0091] Among them, IJDA represents domain adaptation loss, s represents labeled data, t represents data to be identified, and Y s Y represents a labeled dataset containing labeled data. t represents a pseudo-label dataset including pseudo-labels of the data to be identified, Indicates the characteristics of the calculated data set. Represents the first feature, represents the second feature, E(·) represents the expectation of the data, represents the expectation of calculating the second characteristic marginal probability distribution, represents the expectation of calculating the marginal probability distribution of the first feature, P(X|Y=c) represents the calculation of the class conditional probability distribution, It means calculating the expected probability of the second feature occurring when the category is c. It means to calculate the expected probability of the first feature occurring when the category is c, P(X s ) represents the marginal probability distribution of the data set with labeled data, P(Y s =c) represents the category prior distribution of the label, P(Y t =c) represents the category prior distribution of pseudo labels, c represents the data category, X s represents the first data set including the data to be identified, X t represents a second data set including labeled data.

[0092] The implementation method of the present application uses pseudo labels instead of real labels to calculate the conditional probability distribution based on Bayes' theorem, thereby compensating for the poor migration effect caused by the use of approximate calculations in the JDA principle, and proposes an improved joint distribution adaptation method for fault diagnosis of high-pressure common rail injectors under variable operating conditions.

[0093] In a second aspect, according to an embodiment of the present application, a fault diagnosis device for a fuel injector is also provided, such as Figure 5 As shown, it includes: a data acquisition module 100 and a fault identification module 200;

[0094] The data acquisition module 100 is used to acquire the vibration data of the fuel injector to obtain the data to be identified;

[0095] The fault identification module 200 is used to use the data to be identified and the labeled data to adaptively train the fault identification model to be trained based on the improved joint distribution to obtain a trained fault identification model; and input the data to be identified into the fault identification model to obtain a fault identification result.

[0096] The fault identification module includes: a feature extraction unit, a fault prediction unit, a domain adaptation unit and a training unit;

[0097] A feature extraction unit, used for inputting the data to be identified and the labeled data into the feature extraction unit of the fault identification model to be trained, and respectively acquiring the features of the data to be identified and the labeled data based on the attention mechanism to obtain a first feature and a second feature, wherein the first feature is the feature of the data to be identified, and the second feature is the feature of the labeled data;

[0098] A fault prediction unit, used to input the first feature and the second feature into the fault prediction unit to obtain a first prediction result and a second prediction result, wherein the first prediction result is a prediction result obtained according to the first feature, and the second prediction result is a prediction result obtained according to the second feature;

[0099] A domain adaptation unit, configured to input the first feature, the second feature, the label, and the pseudo label into the domain adaptation unit, and determine the domain adaptation loss based on the improved joint distribution adaptation, wherein the label is a label of the labeled data, and the pseudo label is a first prediction result obtained by the fault prediction unit in a previous prediction of the first feature;

[0100] The training unit is used to update the model to be trained according to the domain adaptation loss and the classification loss to obtain a trained fault recognition model.

[0101] like Figure 6 As shown, the data acquisition module includes a vibration sensing unit and a data acquisition and transmission unit. The embodiment of the present application also includes a data partitioning module, which is used to perform working condition judgment on the vibration data and determine the data to be identified. The data partitioning module may include a host computer system. Specifically, the vibration sensing unit adopts a high-precision vibration sensor, which is fixedly installed at a preset position of the high-pressure common rail injector to collect the vibration data of the injector and convert the vibration signal into an electrical signal. Among them, the preset position can be multiple, which can be a key position or a position that is easy to install. The data acquisition and transmission unit includes a multi-channel high-speed acquisition board for receiving the electrical signal of the vibration sensor and transmitting the vibration data to the data partitioning module through digital processing. The acquisition board has a high sampling rate and anti-interference ability, which can ensure the accuracy and real-time nature of the data.

[0102] The data partitioning module, such as the host computer system, includes an industrial computer equipped with a data partitioning algorithm. The industrial computer divides the vibration data into labeled data (source domain data) containing fault categories and data to be identified (target domain data) that does not contain fault categories according to different working conditions (such as different working pressures).

[0103] The labeled data includes a stored labeled data set. The labeled data input to the fault identification module may only use the labeled data of the stored labeled data set.

[0104] The fault identification module includes a feature extraction unit of a multi-attention convolutional network, which inputs the labeled data and the data to be identified into the constructed multi-attention convolutional network to extract the vibration data features as the first feature and the second feature. The domain adaptation unit improves the joint distribution adaptation based on the Bayesian principle to perform feature alignment and determine the domain adaptation loss; the fault prediction unit uses the cross entropy function to determine the classification loss; the training unit updates the iterative fault identification module according to the domain adaptation loss and the classification loss until the training round requirement is met. After the training is completed, only the feature extraction unit and the fault prediction unit can be retained in the fault identification module. The fault prediction unit in the fault identification module performs fault identification and classification and outputs the fault diagnosis result (fault identification result), thereby realizing the fault diagnosis of the high-pressure common rail injector under variable working conditions. Among them, in addition to being able to identify normal and faulty conditions, the implementation method of the present application can also identify specific fault types, such as wear, blockage, etc.

[0105] In a third aspect, according to an embodiment of the present application, a vehicle is further provided, comprising an on-board terminal, wherein the on-board terminal is used to execute a fault diagnosis method for a fuel injector as described in any one of the first aspects.

[0106] In the method of the present application, by training the fault recognition model to be trained with labeled data, data to be identified with working conditions different from those of the labeled data, and improved joint distribution adaptation, the fault recognition model to be trained can be trained for different working conditions, thereby improving the accuracy and robustness of fault identification of high-pressure common rail injectors with variable working conditions.

[0107] The above is only a preferred specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application shall be based on the protection scope of the claims.

Claims

1. A fault diagnosis method for a fuel injector, characterized in that: include: Acquire vibration data of the injector to obtain data to be identified, wherein the data to be identified is the vibration data of a working condition different from that of the labeled data; Adaptively training the fault recognition model to be trained using the data to be recognized and the labeled data based on the improved joint distribution to obtain the trained fault recognition model; The data to be identified is input into the fault identification model to obtain a fault identification result.

2. The fault diagnosis method according to claim 1, characterized in that: The method of using the data to be identified and the labeled data to adaptively train the fault identification model to be trained based on the improved joint distribution to obtain the trained fault identification model includes: Inputting the data to be identified and the labeled data into a feature extraction unit in the fault identification model to be trained, and acquiring features of the data to be identified and the labeled data respectively based on an attention mechanism to obtain a first feature and a second feature, wherein the first feature is a feature of the data to be identified, and the second feature is a feature of the labeled data; Inputting the first feature and the second feature into a fault prediction unit to obtain a first prediction result and a second prediction result, wherein the first prediction result is a prediction result obtained according to the first feature, and the second prediction result is a prediction result obtained according to the second feature; Inputting the first feature, the second feature, a label and a pseudo label into a domain adaptation unit, and determining a domain adaptation loss based on the improved joint distribution adaptation, wherein the label is a label of the labeled data, and the pseudo label is the first prediction result obtained by the fault prediction unit in a previous prediction of the first feature; The model to be trained is updated according to the domain adaptation loss and the classification loss to obtain the trained fault recognition model.

3. The fault diagnosis method according to claim 2, characterized in that: Before updating the model to be trained according to the domain adaptation loss and the classification loss to obtain the trained fault recognition model, the method further includes: The classification loss is determined according to the second prediction result and the label.

4. The fault diagnosis method according to claim 2, characterized in that: The step of inputting the first feature, the second feature, the label, and the pseudo label into a domain adaptation unit, and determining a domain adaptation loss based on the improved joint distribution adaptation, comprises: The domain adaptation loss is determined according to the first feature, the second feature, the label, the pseudo label and the improved joint distribution adaptation, and the calculation method is as follows: Wherein, IJDA represents the domain adaptation loss, s represents the labeled data, t represents the data to be identified, and Y s represents a labeled data set including the labels of the labeled data, Y t A pseudo-label dataset including the pseudo-labels of the data to be identified, Indicates the characteristics of the calculated data set. Represents the first feature, represents the second feature, E(·) represents the expectation of the data, represents the expectation of calculating the second characteristic marginal probability distribution, represents the expectation of calculating the marginal probability distribution of the first feature, It means calculating the expected probability of the second feature occurring when the category is c. It means to calculate the expected probability of the first feature occurring when the category is c, P(X s ) represents the marginal probability distribution of the data set with labeled data, P(Y s =c) represents the category prior distribution of the label, P(Y t =c) represents the category prior distribution of the pseudo label, c represents the data category, X s represents the first data set including the data to be identified, X t Represents a second data set including the labeled data.

5. The fault diagnosis method according to claim 2, characterized in that: The updating of the model to be trained according to the domain adaptation loss and the classification loss to obtain the trained fault recognition model includes: updating the feature extraction unit, the fault prediction unit, and the domain adaptation unit according to the domain adaptation loss and the classification loss; Continue to execute the step of inputting the data to be identified and the labeled data into the feature extraction unit in the fault identification model to be trained, and respectively acquiring the features of the data to be identified and the labeled data based on the attention mechanism to obtain the first feature and the second feature, until a preset training threshold is reached, or the domain adaptation loss and / or the classification loss is less than or equal to a preset loss threshold, to obtain the trained fault identification model including the feature extraction unit and the fault prediction unit.

6. The fault diagnosis method according to claim 2, characterized in that: The method of acquiring the features of the to-be-recognized data and the labeled data based on the attention mechanism to obtain the first feature and the second feature includes: Based on the convolution block attention module, a plurality of frequency features of the data to be identified and the labeled data are respectively extracted to obtain a plurality of first frequency features corresponding to the data to be identified and a plurality of second frequency features corresponding to the labeled data; The first frequency features are combined and then features are extracted again to obtain a first feature, and the second frequency features are combined and then features are extracted again to obtain a second feature.

7. The fault diagnosis method according to claim 1, characterized in that: The step of obtaining the vibration data of the fuel injector to obtain the data to be identified includes: Acquiring the vibration data of the injector; Perform working condition judgment on the vibration data to determine the data to be identified.

8. A fault diagnosis device for a fuel injector, characterized in that: include: Data acquisition module and fault identification module; The data acquisition module is used to acquire vibration data of the injector to obtain data to be identified; The fault identification module is used to use the data to be identified and the labeled data to adaptively train the fault identification model to be trained based on the improved joint distribution to obtain the trained fault identification model; and input the data to be identified into the fault identification model to obtain a fault identification result.

9. The fault diagnosis device according to claim 8, characterized in that: The fault identification module includes: a feature extraction unit, a fault prediction unit, a domain adaptation unit and a training unit; The feature extraction unit is used to input the data to be identified and the labeled data into the feature extraction unit of the fault identification model to be trained, and obtain the features of the data to be identified and the labeled data respectively based on the attention mechanism to obtain a first feature and a second feature, wherein the first feature is the feature of the data to be identified, and the second feature is the feature of the labeled data; The fault prediction unit is used to input the first feature and the second feature into the fault prediction unit to obtain a first prediction result and a second prediction result, wherein the first prediction result is a prediction result obtained according to the first feature, and the second prediction result is a prediction result obtained according to the second feature; The domain adaptation unit is configured to input the first feature, the second feature, a label and a pseudo label into the domain adaptation unit, and determine a domain adaptation loss based on the improved joint distribution adaptation, wherein the label is a label of the labeled data, and the pseudo label is the first prediction result obtained by the fault prediction unit in predicting the first feature in a previous time; The training unit is used to update the model to be trained according to the domain adaptation loss and the classification loss to obtain the trained fault recognition model.

10. A vehicle, characterized in that: It comprises a vehicle-mounted terminal, and the vehicle-mounted terminal is used to execute a fault diagnosis method for a fuel injector as claimed in any one of claims 1 to 7.

Images