Method for constructing internal combustion engine fault diagnosis hybrid model aiming at few samples

Abstract

The invention aims to provide a method for constructing an internal combustion engine fault diagnosis hybrid model aiming at few samples, which belongs to the field of fault diagnosis, and comprises a model discriminator, a sample discriminator and a parameter optimizer: the model discriminator constructs an internal combustion engine fault diagnosis model and analyzes the accuracy of the internal combustion engine fault diagnosis model; a sample discriminator analyzes the classification result of the internal combustion engine fault diagnosis model; and the parameter optimizer finds hyper-parameters of the internal combustion engine fault diagnosis model and optimal values of sample weight values by using an optimization algorithm, and performs two times of parameter optimization in a model construction process, namely parameter optimization for the fault diagnosis model and weight optimization for misclassification samples. According to the method, the fault sample processing capacity of the model under the small sample condition can be improved, the classification precision is effectively improved, the model construction framework relates to application of various algorithms, adjustment can be carried out according to the actual condition, and the universality is high.

Description

Technical Field

[0001] The present invention relates to a fault diagnosis method, in particular to a fault diagnosis method for an internal combustion engine. Background Art

[0002] Internal combustion engines are required to operate for extended periods in harsh environments, including those characterized by corrosion and high impact. This makes them highly susceptible to failures due to degradation, aging, and wear, leading to major accidents and serious threats to life and property. Fault diagnosis plays a crucial role in predicting and managing the health of internal combustion engines. Fault diagnosis is a process that understands and controls the operating state of a machine, determining overall and local normal and abnormal conditions. It is a technology that can detect faults and their causes at an early stage and predict their development trends.

[0003] As mechanical diagnostic and prognostic technologies incorporating the next generation of artificial intelligence mature, they are enabling the construction of more complex and precise mechanical damage models, thereby improving fault diagnosis and prediction performance. Fault diagnosis models constructed using machine learning methods require the use of large amounts of data samples. Because internal combustion engines cannot sustain faults for long periods of time, the collection of fault data is limited. This data scarcity hinders the development of accurate fault diagnosis models, especially in the context of small sample sizes and highly unbalanced fault modeling, which presents a significant and challenging problem.

[0004] Chinese invention patent CN116933656A discloses a deep learning-based method for diagnosing marine power system faults. This invention divides the power system into the power system, power system modules, power system equipment, and power system parameters based on the type of the marine power system's main engine. Aiming to address the problem of fault diagnosis in low- and zero-carbon marine power systems, it proposes a data mining and deep learning-based fault diagnosis algorithm applicable to multiple power systems. However, this invention has limitations in its application to marine power system fault diagnosis. Furthermore, while deep learning offers advantages in data processing, this invention's combination of data mining and deep learning cannot address the low accuracy and precision of fault diagnosis models with a small number of samples. Summary of the Invention

[0005] The object of the present invention is to provide a method for constructing a hybrid model for internal combustion engine fault diagnosis with a small number of samples, which can improve the fault sample processing capability under small sample conditions.

[0006] The object of the present invention is achieved like this:

[0007] The present invention provides a method for constructing a hybrid model for internal combustion engine fault diagnosis with a small number of samples, which is characterized by comprising a model discriminator, a sample discriminator, and a parameter optimizer: the model discriminator constructs the internal combustion engine fault diagnosis model, analyzes the accuracy of the internal combustion engine fault diagnosis model, and screens high-accuracy fault diagnosis sub-models as the construction basis of the hybrid model; the sample discriminator analyzes the classification results of the internal combustion engine fault diagnosis model, and according to the accuracy of the fault samples of the internal combustion engine fault diagnosis model, performs weighted combination of the samples of the classification results, marks the erroneous samples, and updates the weights of the erroneous samples; the parameter optimizer uses an optimization algorithm to find the optimal values ​​of the hyperparameters and sample weights of the internal combustion engine fault diagnosis model, and performs two parameter optimizations in the model construction process, namely, parameter optimization for the fault diagnosis model and weight optimization for the erroneously classified samples;

[0008] Step 1: Process the raw data obtained by the internal combustion engine sensor to extract the fault sample characteristic parameter F, or simulate the internal combustion engine fault state through a simulation model to obtain the thermodynamic parameters of different internal combustion engine fault types C as the fault sample characteristic parameter F; divide the internal combustion engine fault samples into a training set Trr and a test set Te;

[0009] Step 2: Import the training set divided in step 1 into the machine learning algorithm to build the internal combustion engine fault diagnosis model M, where the input variables are the extracted feature parameters and the outputs are different fault types of the internal combustion engine. Import the test set into the internal combustion engine fault diagnosis model and calculate the accuracy of the classification results.

[0010] Step 3: Import the internal combustion engine fault diagnosis model M constructed in step 2 into the model discriminator MD, use the classification accuracy as the screening criterion, set a threshold, and retain the internal combustion engine fault diagnosis models that exceed the threshold as sub-models of the internal combustion engine fault diagnosis hybrid model. The number of sub-models is greater than or equal to 2;

[0011] Step 4: Use the optimization algorithm in the parameter optimizer OP to optimize the hyperparameters of the sub-model selected in step 3, use the accuracy of the classification results as the optimization indicator, and find the hyperparameters of the internal combustion engine fault diagnosis model with the highest classification accuracy;

[0012] Step 5: Analyze the classification results of the sub-model optimized in step 4, import the classification results into the sample discriminator SD, perform weighted combination on all fault samples in the classification results, initialize the weight value, and mark the misclassified fault samples;

[0013] Step 6: Import the error samples marked in step 5 into the parameter optimizer OP and use the optimization algorithm to optimize the error samples;

[0014] Step 7: Mix the weights in step 6 with the optimized sub-model to establish a hybrid model for internal combustion engine fault diagnosis.

[0015] The present invention may also include:

[0016] 1. The processing of the raw data obtained by the internal combustion engine sensor in step 1 includes denoising the data using the principal component analysis method and the modal decomposition method, extracting the characteristic parameters F of the fault sample, including converting the time domain signal into the frequency domain signal using Fourier transform and wavelet Fourier transform, extracting the time domain and frequency domain characteristic parameters, and extracting the fault sample characteristic parameters using a convolutional neural network.

[0017] 2. In step 1, the internal combustion engine fault samples are divided into training set and test set. Randomly select equal number of samples of each fault type of the internal combustion engine to combine and construct the training set Tr. The remaining samples are used as the test set Te. The training set should be a set containing different fault samples {Tr1, Tr2, Tr3, ..., Tr i}、Test set {Te1,Te2,Te3,…,Te i ,}; For example, the internal combustion engine fault type C contains 5 fault types C = {c1, c2, c3, c4, c5}, each fault type contains 5 characteristic parameters F = {f1, f2, f3, f4, f5}, 100 samples are collected for each fault type, and a total of 500 fault samples, then the training set Tr1 is a fault sample randomly selected from the fault samples of the 5 fault types C. The training set Tr1 contains 5 fault samples, and the remaining 490 fault samples are used as the test set Te1, and so on.

[0018] 3. The model discriminator uses machine learning algorithms to build a variety of internal combustion engine fault diagnosis models, including neural networks, support vector machines, decision trees, and random forests.

[0019] 4. In step 3, the internal combustion engine fault diagnosis model M is imported into the model discriminator MD, and the sub-models are selected based on the accuracy of the results, as shown below:

[0020]

[0021] Where M is the internal combustion engine fault diagnosis model, R is the model classification accuracy, and value is the set accuracy threshold.

[0022] 5. In step 5, the classification results are imported into the sample discriminator SD, the fault samples are weighted and combined, and the weights of the samples are initialized. The weights of the correctly classified samples are retained, and the weights of the incorrectly classified samples are updated by the parameter optimizer:

[0023]

[0024] Where w is the sample weight, i is the sample, N is the number of sub-models screened in step 4, and w' is the weight after the optimization algorithm updates;

[0025] Mark the error samples, import the weights of the error samples into the parameter optimizer for weight update, and reorganize the weights w of the correct samples of each sub-model with the updated weights w' of the error samples. The sum of the weights w of the correct samples of each sub-model is 1, and the sum of the weights w' of the updated error samples is 1:

[0026]

[0027] Where Sample is the weighted combination of samples, R is the number of correctly classified samples, E is the number of incorrectly classified samples, Sr is the number of correctly classified samples, Se is the number of incorrectly classified samples, and Ms is the high-accuracy sub-model selected by the model discriminator.

[0028] 6. The process of the sample discriminator in step 6 is as follows:

[0029] (1) Weight the samples of the N high-accuracy sub-models after screening by the model discriminator, with the weight value being w;

[0030] (2) Initialize the weight w to 1 / N, the sum of the weights to 1, and mark the error samples;

[0031] (3) Use the parameter optimizer to update the weight values ​​of the error samples and find the optimal weight combination w'(N);

[0032] (4) Perform weighted combination of different sub-models according to the optimal weight w'(N);

[0033] (5) Obtain a high-accuracy hybrid model for internal combustion engine fault diagnosis.

[0034] 7. The parameter optimizer uses an optimization algorithm to find the optimal values ​​of the hyperparameters and sample weight values ​​of the internal combustion engine fault diagnosis model. The optimization algorithms include particle swarm optimization algorithm, genetic algorithm, gray wolf optimization algorithm, and snake optimization algorithm. The parameter optimizer performs two optimizations, adopting different optimization algorithms according to different parameters.

[0035] The advantages of the present invention are that: the present invention can improve the model's ability to process fault samples in the case of small samples, and its classification accuracy is effectively improved. The model construction framework of the present invention involves the application of multiple algorithms and can be adjusted according to actual conditions, and its versatility is high. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 is a flow chart of the present invention;

[0037] Figure 2 This is the flow chart of the sample discriminator. DETAILED DESCRIPTION

[0038] The present invention will be described in more detail below with reference to the accompanying drawings:

[0039] Combine Figure 1-2 , the present invention includes a model discriminator, a sample discriminator, and a parameter optimizer:

[0040] The model discriminator is used to build an internal combustion engine fault diagnosis model using multiple machine learning algorithms, analyze the accuracy of multiple internal combustion engine fault diagnosis models, and select high-accuracy fault diagnosis sub-models as the basis for building a hybrid model;

[0041] The sample discriminator is used to analyze the classification results of the internal combustion engine fault diagnosis model. According to the accuracy of the fault samples of the internal combustion engine fault diagnosis model, the sample of the classification result is weighted and combined, the error samples are marked, and the weights of the error samples are updated to improve the sample processing capability of the internal combustion engine fault diagnosis model.

[0042] The parameter optimizer is used to use the optimization algorithm to find the optimal values ​​of the hyperparameters and sample weights of the internal combustion engine fault diagnosis model, and perform two parameter optimizations in the model construction process, namely, parameter optimization for the fault diagnosis model and weight optimization for the misclassified samples.

[0043] The present invention comprises the steps of:

[0044] Step 1: Process the raw data obtained by the internal combustion engine sensor to extract the fault sample characteristic parameter F, or simulate the internal combustion engine fault state through a simulation model to obtain the thermodynamic parameters of different internal combustion engine fault types C as the fault sample characteristic parameter F; divide the internal combustion engine fault samples into a training set Tr and a test set Te;

[0045] Step 2: Import the training set divided in step 1 into multiple machine learning algorithms to build multiple internal combustion engine fault diagnosis models M, where the input variables are the extracted feature parameters and the outputs are different fault types of the internal combustion engine. Import the test set into the multiple internal combustion engine fault diagnosis models and calculate the accuracy of the classification results.

[0046] Step 3: Import the multiple internal combustion engine fault diagnosis models M constructed in step 2 into the model discriminator MD, use the classification accuracy as the screening criterion, set a threshold, and retain the internal combustion engine fault diagnosis models that exceed the threshold as sub-models of the internal combustion engine fault diagnosis hybrid model. The number of sub-models should be greater than or equal to 2;

[0047] Step 4: Use the optimization algorithm in the parameter optimizer OP to optimize the hyperparameters of the sub-model selected in step 3, use the accuracy of the classification results as the optimization indicator, and find the hyperparameters of the internal combustion engine fault diagnosis model with the highest classification accuracy, so as to further improve the accuracy of the internal combustion engine fault diagnosis model;

[0048] Step 5: Analyze the classification results of the sub-model optimized in step 4, import the classification results into the sample discriminator SD, perform weighted combination on all fault samples in the classification results, initialize the weight value, and mark the misclassified fault samples;

[0049] Step 6: Import the error samples marked in step 5 into the parameter optimizer OP, and use the optimization algorithm to optimize the error samples to further improve the accuracy of the optimized sub-model;

[0050] Step 7: Mix the weights in step 6 with the optimized sub-model to establish a hybrid model for internal combustion engine fault diagnosis, export the model, and save it.

[0051] In step 1, the raw data obtained by the internal combustion engine sensor is processed, including noise reduction processing of the data using principal component analysis method and modal decomposition method, extraction of fault sample characteristic parameters F, including conversion of time domain signals into frequency domain signals using Fourier transform, wavelet Fourier transform, etc., extraction of time domain and frequency domain characteristic parameters using formulas, and extraction of fault sample characteristic parameters using convolutional neural networks.

[0052] In step 1, the internal combustion engine fault samples are divided into training set and test set. In order to ensure the average distribution of the training set and the test set, an equal number of samples of each fault type of the internal combustion engine are randomly selected and combined to construct the training set Tr, and the remaining samples are used as the test set Te. In order to study the influence of different sample numbers on the internal combustion engine fault diagnosis model, the training set should be designed as a set containing different fault samples {Tr1, Tr2, Tr3, ..., Tr i}, test set {Te1,Te2,Te3,…,Te i For example, the internal combustion engine fault type C includes five fault types C = {c1, c2, c3, c4, c5}, each fault type includes five characteristic parameters F = {f1, f2, f3, f4, f5}, and 100 samples are collected for each fault type, for a total of 500 fault samples. Then the training set Tr1 is a fault sample randomly selected from the fault samples of the five fault types C. The training set Tr1 contains five fault samples, and the remaining 490 fault samples are used as the test set Te1, and so on.

[0053] The model discriminator uses a variety of machine learning algorithms to build a variety of internal combustion engine fault diagnosis models, where the various machine learning algorithms include the application of a variety of classification algorithms, including neural networks, support vector machines, decision trees, random forests, etc.

[0054] In step 3, multiple internal combustion engine fault diagnosis models M are imported into the model discriminator MD, and sub-models are selected based on the accuracy of the results, as shown below:

[0055]

[0056] Where M is the fault diagnosis model constructed based on different machine learning methods, R is the model classification accuracy, and value is the set accuracy threshold, which is set to 95%.

[0057] In step 5, the classification results are imported into the sample discriminator SD, the fault samples are weighted and combined, and the sample weights are initialized. The weights of correctly classified samples are retained, and the weights of incorrectly classified samples are updated by the parameter optimizer:

[0058]

[0059] Where w is the sample weight, i is the sample, N is the number of sub-models screened in step 4, and w' is the weight after the optimization algorithm updates;

[0060] Mark the error samples, import the weights of the error samples into the parameter optimizer for weight update, and reorganize the weights w of the correct samples of each sub-model with the updated weights w' of the error samples. The sum of the weights w of the correct samples of each sub-model is 1, and the sum of the weights w' of the updated error samples is 1:

[0061]

[0062] Where Sample is the weighted combination of samples, R is the number of correctly classified samples, E is the number of incorrectly classified samples, Sr is the number of correctly classified samples, Se is the number of incorrectly classified samples, and Ms is the high-accuracy sub-model selected by the model discriminator.

[0063] The process of the sample discriminator in step 6 is as follows:

[0064] Step 1: Weight the samples of the N high-accuracy sub-models that have passed the model discriminator, with a weight of w;

[0065] Step 2: Initialize the weight w to 1 / N, the sum of the weights to 1, and mark the error samples;

[0066] Step 3: Use the parameter optimizer to update the weight value of the error sample and find the optimal weight combination w'(N) to improve the accuracy of the error sample;

[0067] Step 4: Perform weighted combination of different sub-models according to the optimal weight w'(N);

[0068] Step 5: Obtain a high-accuracy internal combustion engine fault diagnosis hybrid model to further improve the accuracy of the model.

[0069] The parameter optimizer uses an optimization algorithm to find the optimal values ​​of the hyperparameters and sample weights of the internal combustion engine fault diagnosis model. There are many types of optimization algorithms. Traditional optimization algorithms include particle swarm optimization algorithms and genetic algorithms. New optimization algorithms include gray wolf optimization algorithms and snake optimization algorithms. The parameter optimizer performs two optimizations. Due to the differences in the effects of the optimization algorithms, different optimization algorithms need to be adopted according to different parameters.

Claims

1. A method for constructing a hybrid model for internal combustion engine fault diagnosis based on a small number of samples, characterized by: Include Model Discriminator, sample discriminator, parameter optimizer: The model discriminator constructs an internal combustion engine fault diagnosis model, analyzes the accuracy of the internal combustion engine fault diagnosis model, and selects high-accuracy fault diagnosis sub-models as the basis for constructing the hybrid model; the sample discriminator analyzes the classification results of the internal combustion engine fault diagnosis model, performs weighted combination of the classification result samples based on the accuracy of the fault samples of the internal combustion engine fault diagnosis model, marks the erroneous samples, and updates the weights of the erroneous samples; The parameter optimizer uses an optimization algorithm to find the optimal values ​​of the hyperparameters and sample weights of the internal combustion engine fault diagnosis model, and performs two parameter optimizations during the model construction process: parameter optimization for the fault diagnosis model and weight optimization for the misclassified samples; Step 1: Process the raw data obtained by the internal combustion engine sensor to extract the fault sample characteristic parameter F, or simulate the internal combustion engine fault state through a simulation model to obtain the thermodynamic parameters of different internal combustion engine fault types C as the fault sample characteristic parameter F; The internal combustion engine fault samples are divided into training set Trr and test set Te; Step 2: Import the training set divided in step 1 into the machine learning algorithm to build the internal combustion engine fault diagnosis model M, where the input variables are the extracted feature parameters and the outputs are different fault types of the internal combustion engine. Import the test set into the internal combustion engine fault diagnosis model and calculate the accuracy of the classification results. Step 3: Import the internal combustion engine fault diagnosis model M constructed in step 2 into the model discriminator MD, use the classification accuracy as the screening criterion, set a threshold, and retain the internal combustion engine fault diagnosis models that exceed the threshold as sub-models of the internal combustion engine fault diagnosis hybrid model. The number of sub-models is greater than or equal to 2; Step 4: Use the optimization algorithm in the parameter optimizer OP to optimize the hyperparameters of the sub-model selected in step 3, use the accuracy of the classification results as the optimization indicator, and find the hyperparameters of the internal combustion engine fault diagnosis model with the highest classification accuracy; Step 5: Analyze the classification results of the sub-model optimized in step 4, import the classification results into the sample discriminator SD, perform weighted combination on all fault samples in the classification results, initialize the weight value, and mark the misclassified fault samples; Step 6: Import the error samples marked in step 5 into the parameter optimizer OP and use the optimization algorithm to optimize the error samples; Step 7: Mix the weights in step 6 with the optimized sub-model to establish a hybrid model for internal combustion engine fault diagnosis.

2. The method for constructing a hybrid model for internal combustion engine fault diagnosis based on a small number of samples according to claim 1, wherein: In step 1, the raw data obtained by the internal combustion engine sensor is processed, including noise reduction processing of the data using the principal component analysis method and the modal decomposition method, and the extraction of the fault sample characteristic parameter F, including the use of Fourier transform and wavelet Fourier transform to convert the time domain signal into the frequency domain signal, extract the time domain and frequency domain characteristic parameters, and use the convolutional neural network to extract the fault sample characteristic parameters.

3. The method for constructing a hybrid model for internal combustion engine fault diagnosis based on a small number of samples according to claim 1, wherein: In step 1, the internal combustion engine fault samples are divided into training set and test set. Equal samples of each fault type of the internal combustion engine are randomly selected and combined to construct the training set Tr. The remaining samples are used as the test set Te. The training set should be a set containing different fault samples {Tr1, Tr2, Tr3, ..., Tr i }、Test set {Te1,Te2,Te3,…,Te i ,}; For example, the internal combustion engine fault type C contains 5 fault types C = {c1, c2, c3, c4, c5}, each fault type contains 5 characteristic parameters F = {f1, f2, f3, f4, f5}, 100 samples are collected for each fault type, and a total of 500 fault samples, then the training set Tr1 is a fault sample randomly selected from the fault samples of the 5 fault types C. The training set Tr1 contains 5 fault samples, and the remaining 490 fault samples are used as the test set Te1, and so on.

4. The method for constructing a hybrid model for internal combustion engine fault diagnosis with a small number of samples according to claim 1 is characterized by: The model discriminator uses machine learning algorithms to build a variety of internal combustion engine fault diagnosis models, where the machine learning algorithms include neural networks, support vector machines, decision trees, and random forests.

5. The method for constructing a hybrid model for internal combustion engine fault diagnosis with a small number of samples according to claim 1 is characterized by: In step 3, the internal combustion engine fault diagnosis model M is imported into the model discriminator MD, and the sub-models are selected based on the result accuracy, as shown below: Where M is the internal combustion engine fault diagnosis model, R is the model classification accuracy, and value is the set accuracy threshold.

6. The method for constructing a hybrid model for internal combustion engine fault diagnosis with a small number of samples according to claim 1, characterized in that: In step 5, the classification results are imported into the sample discriminator SD, the fault samples are weighted and combined, the sample weights are initialized, the weights of the correctly classified samples are retained, and the weights of the incorrectly classified samples are updated by the parameter optimizer: Among them, w is the sample weight, i is the sample, N is the number of sub-models screened in step 4, w ' The updated weights for the optimization algorithm; Mark the error samples, import the weights of the error samples into the parameter optimizer for weight update, and compare the weights w of the correct samples of each sub-model with the updated weights w of the error samples. ' After reorganization, the sum of the weights w of the correct samples of each sub-model is 1, and the updated weights w of the wrong samples are ' The sum of is 1: Where Sample is the weighted combination of samples, R is the number of correctly classified samples, E is the number of incorrectly classified samples, Sr is the number of correctly classified samples, Se is the number of incorrectly classified samples, and Ms is the high-accuracy sub-model selected by the model discriminator.

7. The method for constructing a hybrid model for internal combustion engine fault diagnosis with a small number of samples according to claim 1, characterized in that: The process of the sample discriminator in step 6 is as follows: (1) Weight the samples of the N high-accuracy sub-models after screening by the model discriminator, with the weight value being w; (2) Initialize the weight w to 1 / N, the sum of the weights to 1, and mark the error samples; (3) Use the parameter optimizer to update the weight value of the error sample and find the optimal weight combination w ' (N); (4) According to the optimal weight w ' (N), weighted combination of different sub-models; (5) Obtain a high-accuracy hybrid model for internal combustion engine fault diagnosis.

8. The method for constructing a hybrid model for internal combustion engine fault diagnosis with a small number of samples according to claim 1, characterized in that: The parameter optimizer uses an optimization algorithm to find the optimal values ​​of hyperparameters and sample weight values ​​of the internal combustion engine fault diagnosis model. The optimization algorithms include particle swarm optimization algorithm, genetic algorithm, gray wolf optimization algorithm, and snake optimization algorithm. The parameter optimizer performs two optimizations, adopting different optimization algorithms according to different parameters.

Images