Triple feature analysis fault diagnosis method for multi-working-condition faults of internal combustion engine based on transfer learning

Abstract

The invention aims to provide a triple feature analysis fault diagnosis method for multi-working-condition faults of an internal combustion engine based on transfer learning, and belongs to the field of fault diagnosis of internal combustion engines. Comprising the steps of feature conversion, feature adaptation and feature reconstruction, the feature conversion reduces the difference of domain structures of source domain data and target domain data, basic features of fault data are converted into statistical features, source domain data features and target domain data features are converted, the feature adaptation determines the relation between the source domain statistical features and the target domain statistical features, and the feature reconstruction is carried out on the source domain statistical features and the target domain statistical features. Weight values of the source domain data and the target domain data are determined, feature reconstruction is carried out according to the source domain data and the target domain data, features of the source domain data and features of the target domain data are reconstructed, and transfer learning is carried out on the target domain data based on the reconstructed features of the source domain data. The multi-working-condition fault diagnosis method can diagnose multi-working-condition faults under the condition that fault data samples are scarce, and is suitable for feature analysis of various fault data.

Description

TECHNICAL FIELD

[0001] The present application relates to an internal combustion engine fault diagnosis method, in particular to an internal combustion engine fault diagnosis method under a few sample conditions. BACKGROUND

[0002] Since the fault experiment causes serious wear and tear to the equipment, and even irreversible, there are limitations in obtaining fault data. Machine learning methods require a large amount of accurate data for model training. In the case of data sample scarcity, the fault diagnosis model constructed by the machine learning algorithm has poor precision, low accuracy, and is difficult to meet the application requirements.

[0003] In the case of a small number of data samples, there are few sample learning, zero sample learning, and transfer learning. Among them, transfer learning is widely used in various fields (such as mechanical equipment, medical treatment, industrial detection, etc.) due to its high computational efficiency, improved model performance, and reduced cost in the case of data scarcity.

[0004] The core idea of transfer learning is to apply the knowledge and model learned in one field (referred to as the "source field") to another related but different field (referred to as the "target field"). In machine learning, traditional supervised learning assumes that training data and test data are independent and identically distributed. Transfer learning bravely breaks this assumption, and it specifically studies how to effectively transfer knowledge when the data distribution of the source field and the target field is different, but the task is related.

[0005] Transfer learning is used in the case of data scarcity and high labeling cost, limited computing resources, breaking through the performance bottleneck of the model, and directly training the model on a small data set is prone to overfitting, and the performance has an upper limit. By using the general features learned on large data, the generalization ability and performance of the model on small data can be improved.

[0006] From the theoretical and practical point of view, transfer learning mainly has the following implementation paradigms: model-based transfer, directly reusing the model structure and weights trained in the source field. Feature-based transfer, find a feature transformation method to map the data in the source field and the target field to the same feature space, and in this new space, their distribution is as similar as possible. Relationship-based transfer, assuming that the source field and the target field share some logical relationships or dependencies, then transfer these relationship knowledge. Instance-based transfer, select samples with high similarity to target field data from source field data, and give these samples higher weights, then learn with target field data. SUMMARY

[0007] The purpose of this invention is to provide a fault diagnosis method for internal combustion engines based on transfer learning, which utilizes triple feature analysis to diagnose multi-condition faults in internal combustion engines, enabling the diagnosis of faults under various operating conditions even when fault data samples are scarce. The objective of this invention is achieved as follows: This invention discloses a fault diagnosis method for internal combustion engines based on transfer learning and triple feature analysis of multi-condition faults, characterized by the following steps: (1) Data acquisition: The fault conditions are simulated by constructing a bench test or simulation model of the internal combustion engine to obtain fault datasets under different operating conditions. Basic features as input variables Fault labels are output variables Determine the characteristics of source domain data Features of target domain data ; (2) Feature transformation: transform the domain data features in step (1) Basic characteristics and target domain data features Basic characteristics Convert to statistical characteristics Construct a statistical feature dataset Furthermore, machine learning algorithms were imported to construct a fault diagnosis model and verify the accuracy of the statistical feature dataset; (3) Feature adaptation: Based on the statistical feature dataset in step (2) The features are further adapted to determine the relationship between the source domain dataset and the target domain data samples. Based on the fault type, target domain data samples are randomly selected, and the difference between the source domain data samples and the selected target domain data samples is calculated as the weight value. ; (4) Feature reconstruction: weight values ​​in step (3) As a standard for feature reconstruction between source domain data and target domain data, weights are combined with source domain data samples to reconstruct the features of target domain data. ; (5) Verification: Use correlation analysis to analyze the characteristics of the target domain data. Reconstructing features in step (4) The domain structure is used to construct a fault diagnosis model using reconstructed features. The target domain dataset is used as a test sample to verify the accuracy of the reconstructed features.

[0008] The present invention may also include: 1. Step (1) Data acquisition includes acquiring fault data of gasoline engines and diesel engines, as well as the fault types of relevant components of the equipment. This includes data acquisition for turbochargers, air filters, catalytic converters, igniters, fuel injectors, and various operating conditions. Basic characteristics related to fault data Parameters related to internal combustion engine failures are used to determine the operating conditions of source domain data and target domain data.

[0009] 2. The feature transformation in step (2) includes: calculating the features of the source domain data. and target domain data features Statistical characteristics Statistical characteristics include mean, median, standard deviation, variance, mean absolute deviation, skewness, kurtosis, root mean square value, average power, and energy statistical characteristics, based on fault type. Build Feature dataset with statistical characteristics: .

[0010] 3. Feature adaptation in step (3) includes: determining the statistical features of the target domain. and Randomly select target domain samples Calculate the weight values :

[0011] 4. Feature reconstruction in step (4) includes: using weight values and statistical characteristics of source domain data Statistical characteristics of target domain data Refactor : .

[0012] 5. The correlation analysis methods used in step (5) to analyze the characteristics of source domain data and target domain data include maximum mean difference, principal component analysis, correlation alignment, and A-distance method.

[0013] 6. The machine learning algorithms used to construct the fault diagnosis model in steps (2) and (5) include neural networks, support vector machines, decision trees, and random forests.

[0014] The advantages of this invention are: it can diagnose faults under multiple operating conditions when fault data samples are scarce; the analysis process of this invention involves feature transformation, feature adaptation, and feature reconstruction, and is applicable to feature analysis of various fault data. Attached Figure Description

[0015] Fig. 1 Flowchart for feature reconstruction; Fig. 2 This is a flowchart of the present invention. Detailed Implementation

[0016] The invention will now be described in more detail with reference to the accompanying drawings: Combination Figs. 1-2 The present invention provides a fault diagnosis method for multi-condition faults of internal combustion engines based on transfer learning, comprising feature transformation, feature adaptation, and feature reconstruction. Feature transformation reduces the difference in domain structure between source domain data and target domain data. Based on statistical principles, it transforms the basic features of fault data into statistical features and transforms the features of source domain data and target domain data. Feature adaptation determines the relationship between statistical features of the source domain and statistical features of the target domain, determines the weight values ​​of source domain data and target domain data, and performs the process based on source domain data and target domain data; Feature reconstruction involves reconstructing the features of source domain data and target domain data, and then performing transfer learning on the target domain data based on the reconstructed features from the source domain data.

[0017] The specific steps are as follows: Step 1: Data Acquisition. Fault conditions are simulated by constructing a bench test or simulation model of the internal combustion engine to obtain fault datasets under different operating conditions. Basic features as input variables Fault labels are output variables Determine the characteristics of source domain data Features of target domain data ; Step 2: Feature transformation, transforming the basic features of the source domain data from Step 1. and the basic characteristics of the target domain data Convert to statistical characteristics Construct a statistical feature dataset Furthermore, machine learning algorithms were imported to construct a fault diagnosis model and verify the accuracy of the statistical feature dataset; Step 3: Feature adaptation, based on the statistical feature dataset from Step 2. The features are further adapted to determine the relationship between the source domain dataset and the target domain data samples. Based on the fault type, a small number of target domain data samples are randomly selected, and the difference between the source domain data samples and the selected target domain data samples is calculated as the weight value. ; Step 4: Feature reconstruction, weight values ​​from step 3 As a standard for feature reconstruction between source domain data and target domain data, weights are combined with source domain data samples to reconstruct the features of target domain data. ; Step 5: Validation, using correlation analysis to analyze the characteristics of the target domain data. Reconstructing features in step 4 The domain structure is used to construct a fault diagnosis model using reconstructed features. The target domain dataset is used as a test sample to verify the accuracy of the reconstructed features. Step 1 involves acquiring internal combustion engine fault data, including fault data from gasoline engines, diesel engines, and other equipment, as well as the fault types of related components. This includes data acquisition for turbochargers, air filters, catalytic converters, igniters, fuel injectors, and various operating conditions. Basic characteristics related to fault data Parameters related to internal combustion engine failures are used to determine the operating conditions of source domain data and target domain data.

[0018] The feature transformation in step 2 calculates the features of the source domain data. and target domain data features Statistical characteristics The statistical characteristics include mean, median, standard deviation, variance, mean absolute deviation, skewness, kurtosis, root mean square value, average power, energy, etc., depending on the fault type. Build A feature dataset with statistical characteristics.

[0019] Feature fitting in step 3 is for determining the statistical features of the target domain. and Randomly select a small number of target domain samples Calculate the weight values .

[0020] The feature reconstruction in step 4 utilizes weight values. and statistical characteristics of source domain data Statistical characteristics of target domain data Refactor .

[0021] The correlation analysis methods used in step 5 to analyze the characteristics of source domain data and target domain data include maximum mean difference, principal component analysis, correlation alignment, A-distance, and other methods.

[0022] The machine learning algorithms used to build the fault diagnosis model in steps 2 and 5 include the application of various classification algorithms, such as neural networks, support vector machines, decision trees, and random forests.

[0023] Feature transformation, feature adaptation, and feature reconstruction can reduce the difference in domain structure between source and target domain data and reduce dimensionality. Furthermore, in the case of scarce samples, it can accurately reconstruct the features of target domain data, making it suitable for fault diagnosis under multiple operating conditions of internal combustion engines.

Claims

1. A fault diagnosis method for internal combustion engines based on transfer learning using triple feature analysis, characterized by: Includes the following steps: (1) Data acquisition: The fault conditions are simulated by constructing a bench test or simulation model of the internal combustion engine to obtain fault datasets under different operating conditions. Basic features as input variables Fault labels are output variables Determine the characteristics of source domain data Features of target domain data ; (2) Feature transformation: transform the domain data features in step (1) Basic characteristics and target domain data features Basic characteristics Convert to statistical characteristics Construct a statistical feature dataset Furthermore, machine learning algorithms were imported to construct a fault diagnosis model and verify the accuracy of the statistical feature dataset; (3) Feature adaptation: Based on the statistical feature dataset in step (2) The features are further adapted to determine the relationship between the source domain dataset and the target domain data samples. Based on the fault type, target domain data samples are randomly selected, and the difference between the source domain data samples and the selected target domain data samples is calculated as the weight value. ; (4) Feature reconstruction: weight values ​​in step (3) As a standard for feature reconstruction between source domain data and target domain data, weights are combined with source domain data samples to reconstruct the features of target domain data. ; (5) Verification: Use correlation analysis to analyze the characteristics of the target domain data. Reconstructing features in step (4) The domain structure is used to construct a fault diagnosis model using reconstructed features. The target domain dataset is used as a test sample to verify the accuracy of the reconstructed features.

2. The fault diagnosis method for multi-condition faults of internal combustion engines based on transfer learning according to claim 1, characterized in that: Step (1) Data acquisition includes acquiring fault data for gasoline engines and diesel engines, as well as the fault types of relevant components of the equipment. This includes data acquisition for turbochargers, air filters, catalytic converters, igniters, fuel injectors, and various operating conditions. Basic characteristics related to fault data Parameters related to internal combustion engine failures are used to determine the operating conditions of source domain data and target domain data.

3. The fault diagnosis method for multi-condition faults of internal combustion engines based on transfer learning according to claim 1, characterized in that: The feature transformation in step (2) includes: calculating the features of the source domain data. and target domain data features Statistical characteristics Statistical characteristics include mean, median, standard deviation, variance, mean absolute deviation, skewness, kurtosis, root mean square value, average power, and energy statistical characteristics, based on fault type. Build Feature dataset with statistical characteristics: 。 4. The fault diagnosis method for multi-condition faults of internal combustion engines based on transfer learning according to claim 1, characterized in that: Feature adaptation in step (3) includes: determining the statistical features of the target domain. and Randomly select target domain samples Calculate the weight values : 。 5. The fault diagnosis method for multi-condition faults of internal combustion engines based on transfer learning according to claim 1, characterized in that: The feature reconstruction in step (4) includes: using weight values and statistical characteristics of source domain data Statistical characteristics of target domain data Refactor : 。 6. The fault diagnosis method for multi-condition faults of internal combustion engines based on transfer learning according to claim 1, characterized in that: The correlation analysis methods used in step (5) to analyze the characteristics of source domain data and target domain data include maximum mean difference, principal component analysis, correlation alignment, and A-distance method.

7. The fault diagnosis method for multi-condition faults of internal combustion engines based on transfer learning according to claim 1, characterized in that: The machine learning algorithms used to construct the fault diagnosis model in steps (2) and (5) include neural networks, support vector machines, decision trees, and random forests.

Images