Excavator engine fault diagnosis method based on RFTL-1DNet network

Abstract

The invention discloses an excavator engine fault diagnosis method based on an RFTL-1DNet network. The method comprises the following steps: (1) arranging a sensor on an engine of the excavator for collecting original fault data; step (2), establishing a random forest model to perform dimension screening on an original fault, and performing data preprocessing; step (3), the preprocessed data set and the network pre-training optimal parameters are imported into the RFTL-1DNet, and an optimized RFTL-1DNet network model is established; and step (4), loading the data set after dimension screening, copying original data, adding random noise, expanding the data set, performing normalization processing on the data set, and outputting an excavator fault diagnosis result through an RFTL-1DNet network model after optimization. The method has the advantages of being high in model training speed, few in needed samples and high in robustness, and therefore accurate fault diagnosis of the excavator engine is achieved.

Description

Technical Field

[0001] The present invention relates to the field of mining system engineering and mine optimization technology, and in particular to an excavator engine fault diagnosis method based on an RFTL-1DNet network. Background Art

[0002] In open-pit mining, traditional monitoring and surveillance systems mainly rely on the visualization of sensor data and simple threshold analysis. Although this method can monitor the operation process of the equipment, it often cannot identify early failures in key parts of the equipment. As a key equipment in mining operations, the early identification of excavator failures has a significant impact on the overall interests of mining companies.

[0003] In summary, faced with the increasing complexity of faults, traditional equipment fault diagnosis technologies such as expert systems and fault tree analysis have gradually failed to meet actual needs in terms of diagnostic efficiency and identification accuracy. Summary of the invention

[0004] In order to overcome the above technical problems, the purpose of the present invention is to provide an excavator engine fault diagnosis method based on the RFTL-1DNet network. The method has the characteristics of fast model training speed, few required samples and strong robustness, thereby realizing accurate diagnosis of excavator engine faults.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is:

[0006] A method for diagnosing an excavator engine fault based on an RFTL-1DNet network comprises the following steps:

[0007] Step (1): arranging sensors on the excavator engine to collect original fault data;

[0008] Step (2): Establish a random forest model to filter the dimensions of the original fault data, and perform data preprocessing on the filtered data;

[0009] Step (3): Import the preprocessed data set and the optimal network pre-training parameters into RFTL-1DNet to establish the optimized RFTL-1DNet network model;

[0010] Step (4): Load the dataset after filtering the dimensions, copy the original fault data and normalize it by adding random noise and expanding the dataset, and output the excavator fault diagnosis results through the optimized RFTL-1DNet network model.

[0011] In the step (1), the common fault types and fault characteristic parameters of the excavator engine are determined, and the fault type set y = [y1; y2; y3; ...; y n], the corresponding fault characteristic parameter set is X = [X1; X2; X3; ...; X n ],y n Indicates the nth type of fault, X n Represents y n Fault characteristic parameters.

[0012] The fault characteristic parameters include engine speed, injection quantity setting value, current injection quantity, accelerator pedal opening, accelerator pedal 1 original voltage, accelerator pedal 2 original voltage, coolant temperature, oil pressure, oil temperature, atmospheric pressure, intake pressure, intake temperature, rail pressure interpolation, rail pressure setting value, rail pressure peak, multi-state switch status, engine synchronization signal status, torque load rate, output torque, which are collected through the Zhiduoxing road spectrum monitoring device.

[0013] The step (2) is specifically:

[0014] The collected fault feature parameter data set X is standardized to a standard normal distribution with a mean of 0 and a variance of 1. Referring to formula (1), the split contribution of a certain feature in the decision tree is evaluated by formula (2). The median of the feature importance is used as the threshold for selecting important features. Referring to formula (3), the feature importance is sorted and the screening results are output for testing.

[0015]

[0016] I(f)=∑ t∈T ΔG i,t (2)

[0017] Selected = {f i |importance(f i )≥threshold} (3)

[0018] Where, X std is the standardized data; X is the original data; μ is the feature mean; σ is the standard deviation of the feature; I(f) is the feature importance; T is all decision trees; ΔG i is the contribution of feature f in decision tree t; Selected is the selected feature set; f i is feature i; importance(f i ) is the importance of feature i; threshold is the set threshold.

[0019] In the step (2), the collected original fault data of the excavator engine is used as the network input, and the fault type is used as the network output;

[0020] All the collected factors affecting excavator failures are numbered, where the original variables are:

[0021] Engine speed X1, fuel injection setting value X2, current fuel injection amount X3, accelerator pedal opening X4, accelerator pedal 1 original voltage X5, accelerator pedal 2 original voltage X6, coolant temperature X7, oil pressure X8, oil temperature X9, atmospheric pressure X 10 , intake pressure X 11 , intake air temperature X 12 , rail pressure interpolation X 13 , rail pressure setting value X 14 , rail pressure peak value X 15 , multi-state switch state X 16 , engine synchronization signal status X 17 , torque load rate X 18 , output torque X 19 .

[0022] The random forest classifier constructs multiple decision trees and calculates the feature importance of each tree, taking the average value to obtain the final feature importance score, and sorts and screens the importance of all factors affecting excavator failures based on the score;

[0023] Finally, select engine speed X1, fuel injection setting value X2, current fuel injection amount X3, coolant temperature X7, oil temperature X9, intake pressure X 11 , intake air temperature X 12 , rail pressure setting value X 14 , rail pressure peak value X 15 , torque load rate X 18 These ten influencing factors serve as feature dimensions for model training and fault diagnosis.

[0024] In the step (3), RFTL-1DNet includes ResNet18-1D, the input vector dimension of the ResNet18_1D structure is 1, a one-dimensional convolutional layer, a one-dimensional maximum pooling layer and 4 residual layers are used, and the remaining convolution blocks keep the feature vector size unchanged.

[0025] The fault feature parameters are used as the initial data set to input the ResNet18_1D network for pre-training. The ResNet18_1D structure has an input vector dimension of 1, uses a one-dimensional convolutional layer, a one-dimensional maximum pooling layer, and four residual layers, and the remaining convolution blocks keep the feature vector size unchanged. The forward method is selected for forward propagation. After the input data passes through each residual layer in turn, it is subjected to one-dimensional adaptive average pooling and expanded into a one-dimensional vector. Finally, the 512-dimensional feature vector is mapped to the corresponding fault category through a fully connected layer and the network parameters are retained.

[0026] The network training parameters set in step (3) include ResNet18-1D network training and transfer learning fine-tuning parameters.

[0027] The original fault data collected by various sensors is used as the input of the ResNet18-1D network, and the fault type is determined according to the learning and training status of the model.

[0028] In step (3), the data set is randomly divided into 5 groups, 4 of which are selected for model training, and the remaining 1 group is used for model verification, that is, the network model is trained using five-fold cross validation to avoid underfitting of the model. The training model is saved and output; finally, the RFTL-1DNet network model with the best training effect is called for final testing.

[0029] In the step (4), the random noise generation rule refers to formula (4).

[0030] x=μ+σ*Z (4)

[0031] Where μ is the mean of normal distribution, σ is the standard deviation, and Z is the standard normal distribution random variable.

[0032] In the step (4), the data set after dimension filtering is loaded, the original data is copied and normalized by adding random noise and expanding the data set; the detection data is used as network input, and the fault type is determined based on its output result.

[0033] Beneficial effects of the present invention:

[0034] The present invention establishes an excavator engine fault diagnosis model by using the data-driven and deep learning network characteristics in step (2) and step (3) to learn the fault patterns of excavator engine fault types and fault characteristic parameters. The principle is simple and can effectively realize real-time diagnosis of excavator engine fault status.

[0035] By utilizing the present invention in diagnosing excavator engine faults, rapid and accurate diagnosis can be achieved through RFTL-1DNet, and the diagnostic accuracy rate can reach over 99%, turning the risk of shutdown of mobile equipment in mines into a controllable risk, changing the status quo of post-maintenance in mines, and having good scalability and fault tolerance, providing an effective way for excavator fault diagnosis in mining enterprises. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a flow chart of establishing the fault diagnosis model of the present invention.

[0037] Figure 2 This is a schematic diagram of the confusion matrix for diagnosing excavator engine faults based on the RFTL-1DNet model of the present invention.

[0038] Figure 3 This is a schematic diagram of the ResNet18_1D network.

[0039] Figure 4 Schematic diagram of the network structure of a single residual block. DETAILED DESCRIPTION

[0040] The present invention will be further described in detail below in conjunction with the accompanying drawings.

[0041] like Figure 1 As shown in the figure, the excavator engine fault diagnosis method based on RFTL-1DNet network has the following specific implementation steps:

[0042] Step 1: In the excavator engine fault diagnosis, determine the common fault types and fault characteristic parameters of the excavator engine, the fault type set y = [y1; y2; y3; ...; y n ], the corresponding fault characteristic parameter set is X = [X1; X2; X3; ...; X n ], in the process of data collection, the means of data collection is real-time collection through sensors, and the fault data is recorded in the collected raw data, and the fault type and the fault characteristic parameters correspond to each other; the overall rule for selecting fault characteristic parameters under each type of fault type is to meet the work intensity and time diversity, that is, the fault characteristic data is distributed as much as possible in different mining intensities and times.

[0043] Data selection follows the principle of diversity of mechanical work intensity and time. The goal of data selection is to meet the uniformity and effectiveness of data distribution to ensure the accuracy of the diagnostic model training results;

[0044] There are 8 types of faults corresponding to the original fault data, which are:

[0045] y1 excavator engine normal operation, y2 excavator engine high-pressure oil pump failure, y3 excavator engine intake pipe failure, y4 excavator engine supercharger failure, y5 excavator engine air filter blockage failure, y6 excavator engine fan failure, y7 excavator engine intercooler failure and y8 excavator engine low-pressure oil pump failure.

[0046] Step 2: Standardize the collected fault feature parameter data set X to a standard normal distribution with a mean of 0 and a variance of 1. Refer to formula (1) and use formula (2) to evaluate the split contribution of a certain feature in the decision tree. Let the median of the feature importance be the threshold for selecting important features. Refer to formula (3) to sort the feature importance and output the screening results for testing.

[0047]

[0048] I(f)=∑ t∈T ΔG i,t (2)

[0049] Selected = {f i ||importance(f i )≥threshold} (3)

[0050] Where, X std is the standardized data; X is the original data; μ is the feature mean; σ is the standard deviation of the feature; I(f) is the feature importance; T is all decision trees; ΔG i is the contribution of feature f in decision tree t; Selected is the selected feature set; f i is feature i; importance(f i ) is the importance of feature i; threshold is the set threshold.

[0051] Evaluate the split contribution of each dimension in the decision tree in turn, and use the evaluation results as an important basis for dimension screening. Remove the feature dimensions with low split contribution values ​​to improve the learning efficiency and classification accuracy of the entire model, use the screened data set to pre-train the one-dimensional residual network model, and retain the training results; the fault data in the excavator engine fault data that can be collected is relatively small, which is not enough to fully train the deep model, so it is necessary to add random noise to expand the data set, and normalize the data set to obtain the preprocessed data set, and divide the preprocessed data set into a test set, a training set, and a validation set;

[0052] Step 3: Set the engine speed to X1, the injection amount setting value to X2, the current injection amount to X3, the coolant temperature to X7, the oil temperature to X9, the intake pressure to X 11 , intake air temperature X 12 , rail pressure setting value X 14 , rail pressure peak value X 15 , torque load rate X 18 The initial data set is input into the ResNet18_1D network for pre-training.

[0053] like Figure 3 , Figure 4 As shown in the figure, the input vector dimension of the ResNet18_1D structure is 1, and a one-dimensional convolutional layer, a one-dimensional maximum pooling layer and four residual layers are used. The remaining convolutional blocks keep the size of the feature vector unchanged.

[0054] The fault feature parameters are used as the initial data set to input the ResNet18_1D network for pre-training. The ResNet18_1D structure has an input vector dimension of 1, uses a one-dimensional convolutional layer, a one-dimensional maximum pooling layer, and four residual layers, and the remaining convolution blocks keep the feature vector size unchanged. The forward method is selected for forward propagation. After the input data passes through each residual layer in turn, it is subjected to one-dimensional adaptive average pooling and expanded into a one-dimensional vector. Finally, the 512-dimensional feature vector is mapped to the corresponding fault category through a fully connected layer and the network parameters are retained.

[0055] The set network training parameters include ResNet18-1D network training and transfer learning fine-tuning parameters.

[0056] The original fault data collected by various sensors is used as the input of the ResNet18-1D network, and the fault type is determined according to the learning and training status of the model.

[0057] The data set is randomly divided into 5 groups, 4 of which are selected for model training, and the remaining 1 is used for model validation, that is, the network model is trained using five-fold cross validation to avoid underfitting of the model. The trained model is saved and output; finally, the RFTL-1DNet network model with the best training effect is called for the final test.

[0058] Step 4: Load the dataset after filtering the dimensions, copy the original data and normalize it by adding random noise and expanding the dataset. The random noise generation rule refers to formula (4).

[0059] x=μ+σ*Z (4)

[0060] Where μ is the mean of normal distribution, σ is the standard deviation, and Z is the standard normal distribution random variable.

[0061] Step 5: Train a deep learning model on the source domain, with a training dataset D S , a model θs needs to be trained according to formula (5) to minimize the loss function Ls(θs; Ds) of the source domain; the model θs trained in the source domain is used as the initialization parameter of the target domain model, and the θs model of the present invention is a one-dimensional deep residual network; the model trained in the source domain is used for further training in the target domain, and a target domain dataset Dt is provided. The model parameter θt needs to be optimized in the target domain according to formula (6) to minimize the loss function Lt(θt; Dt) of the target domain. When fine-tuning the model, a lower learning rate is selected to avoid over-adjustment of pre-training related parameters.

[0062]

[0063] Where, L s , L tUsually it is cross entropy loss or mean square error.

[0064] Step 6: The training model uses the cross entropy loss function, and the model is trained through forward propagation, back propagation and optimizer. Finally, the average training loss is returned through the test function. In order to comprehensively evaluate the performance of different models, parameters such as precision, recall and mean (F1-score) are used for comparison. F1 is a comprehensive performance indicator for evaluating category imbalance, which takes into account false positives and false negatives. The value range is between 0 and 1, with 1 indicating the best performance and 0 indicating the worst performance. The higher the F1 score, the more accurate and comprehensive the model is in identifying positive samples.

[0065] Step 7: Select the fault diagnosis model with the best training effect as the diagnosis basis, and select another set of data from the selected fault original data as fault test data, as network input, and test the diagnosis model, which further verifies the feasibility of the excavator engine fault diagnosis model. Figure 2 This is the fault diagnosis result diagram of the excavator engine based on the RFTL-1DNet network, with an average diagnostic accuracy of 99.3%;

[0066] The present invention monitors the operating status of the excavator engine in real time through the Internet of Things sensor, and extracts the key characteristic parameters of the training and test data sets to provide a data basis for the training and testing of the deep learning model. The model optimizes the traditional ResNet18 architecture so that it can process one-dimensional signal data to better adapt to the fault diagnosis task of the excavator engine. In order to enhance the robustness and adaptability of the model, the random forest pre- and data enhancement processing strategy is introduced. The network model solves the data imbalance problem and improves the data quality. The training set is expanded by creating additional synthetic data to ensure that the model maintains high accuracy under diverse working conditions. In addition, transfer learning is used to pre-train the RFTL-1DNet model to accelerate the convergence of the model and improve its adaptability to small-scale data sets. This method effectively avoids the impact of insufficient data volume and uneven distribution on model performance. The excavator engine fault diagnosis model of the present invention shows excellent robustness, wide applicability, high accuracy and fast diagnostic capabilities. The model provides strong technical support for the maintenance and fault handling of excavator engines. The model can significantly improve the operating efficiency and reliability of the excavator, reduce maintenance costs, and extend the service life of the equipment.

[0067] The present invention uses the characteristics of deep residual network and transfer learning to learn the fault patterns of excavator engine fault types and fault characteristic parameters, and establishes an excavator engine fault diagnosis model. It can effectively realize the diagnosis of excavator engine fault status, and can be adjusted according to the individual fault conditions of the excavator during actual application.

[0068] The present invention is described above by way of example in conjunction with the accompanying drawings. It is obvious that the specific implementation of the present invention is not limited to the above-mentioned method. As long as various non-substantial improvements are made using the method concept and technical solution of the present invention, or the concept and technical solution of the present invention are directly applied to other occasions without improvement, they are all within the protection scope of the present invention.

Claims

1. A method for diagnosing engine faults of an excavator based on RFTL-1DNet network, characterized in that: The steps include: Step (1): arranging sensors on the excavator engine to collect original fault data; Step (2): Establish a random forest model to filter the dimensions of the original fault data, and perform data preprocessing on the filtered data; Step (3): Import the preprocessed data set and the optimal network pre-training parameters into RFTL-1DNet to establish the RFTL-1DNet network model; Step (4): Load the dataset after filtering the dimensions, copy the original fault data and normalize it by adding random noise and expanding the dataset, and output the excavator fault diagnosis results through the optimized RFTL-1DNet network model.

2. The excavator engine fault diagnosis method based on RFTL-1DNet network according to claim 1 is characterized in that: In the step (1), the common fault types and fault characteristic parameters of the excavator engine are determined, and the fault type set y = [y1; y2; y3; ...; y n ], the corresponding fault characteristic parameter set is X = [X1; X2; X3; ...; X n ],y n Indicates the nth type of fault, X n Represents y n Fault characteristic parameters.

3. The excavator engine fault diagnosis method based on RFTL-1DNet network according to claim 2 is characterized in that: The step (2) is specifically: The collected fault feature parameter data set X is standardized to a standard normal distribution with a mean of 0 and a variance of 1. Referring to formula (1), the split contribution of a certain feature in the decision tree is evaluated by formula (2). The median of the feature importance is used as the threshold for selecting important features. Referring to formula (3), the feature importance is sorted and the screening results are output for testing. I(f)=∑ t∈T ΔG i，t (2) Selected={f i |importance(f i )≥threshold} (3) In the formula, X std is the standardized data; X is the original data; μ is the feature mean; σ is the standard deviation of the feature; I(f) is the feature importance; T is all decision trees; ΔG i is the contribution of feature f in decision tree t; Selected is the selected feature set; f i is feature i; importance(f i ) is the importance of feature i; threshold is the set threshold.

4. The excavator engine fault diagnosis method based on RFTL-1DNet network according to claim 2 is characterized in that: In the step (2), the collected original fault data of the excavator engine is used as the network input, and the fault type is used as the network output; All the collected factors affecting excavator failures are numbered, where the original variables are: Engine speed X1, fuel injection setting value X2, current fuel injection amount X3, accelerator pedal opening X4, accelerator pedal 1 original voltage X5, accelerator pedal 2 original voltage X6, coolant temperature X7, oil pressure X8, oil temperature X9, atmospheric pressure X 10 , intake pressure X 11 , intake air temperature X 12 , rail pressure interpolation X 13 , rail pressure setting value X 14 , rail pressure peak value X 15 , multi-state switch state X 16 , engine synchronization signal status X 17 , torque load rate X 18 , output torque X 19 .

5. The excavator engine fault diagnosis method based on RFTL-1DNet network according to claim 4 is characterized in that: The random forest classifier constructs multiple decision trees and calculates the feature importance of each tree, taking the average value to obtain the final feature importance score, and sorts and screens the importance of all factors affecting excavator failures based on the score; Finally, select engine speed X1, fuel injection setting value X2, current fuel injection amount X3, coolant temperature X7, oil temperature X9, intake pressure X 11 , intake air temperature X 12 , rail pressure setting value X 14 , rail pressure peak value X 15 , torque load rate X 18 These ten influencing factors serve as feature dimensions for model training and fault diagnosis.

6. The excavator engine fault diagnosis method based on RFTL-1DNet network according to claim 1 is characterized in that: In the step (3), RFTL-1DNet includes ResNet18-1D, the input vector dimension of the ResNet18_1D structure is 1, a one-dimensional convolutional layer, a one-dimensional maximum pooling layer and four residual layers are used, and the remaining convolutional blocks keep the feature vector size unchanged; The fault feature parameters are used as the initial data set and input into the ResNet18_1D network for pre-training. The ResNet18_1D structure has an input vector dimension of 1. It uses a one-dimensional convolutional layer, a one-dimensional maximum pooling layer, and four residual layers. The remaining convolution blocks keep the size of the feature vector unchanged. The forward method is selected for forward propagation. After the input data passes through each residual layer in turn, one-dimensional adaptive average pooling is performed and expanded into a one-dimensional vector. Finally, the 512-dimensional feature vector is mapped to the corresponding fault category through a fully connected layer and the network parameters are retained.

7. The excavator engine fault diagnosis method based on RFTL-1DNet network according to claim 1 is characterized in that: In the step (4), the dataset after dimension screening is loaded, the original data is copied and normalized by adding random noise and expanding the dataset. The random noise generation rule refers to formula (4); x=μ+σ*Z (4) Where μ is the mean of normal distribution, σ is the standard deviation, and Z is the standard normal distribution random variable.

8. The excavator engine fault diagnosis method based on RFTL-1DNet network according to claim 7 is characterized in that: The detection data is used as the network input, and the fault type is determined based on its output results.

Images