Method for diagnosing new fault types of equipment based on improved stack autoencoder
By combining stack autoencoder and tag transfer algorithm, the hidden layer features of mechanical equipment are extracted, and the problem of difficulty in identifying new fault types in the existing technology is solved, and efficient diagnosis of mechanical equipment failure types is achieved, especially new fault types are realized. The correct recognition rate reached 92.5%.
Patent Information
- Application Number
- CN202111470928.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-12-03
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2041-12-03
AI Technical Summary
Existing mechanical equipment fault diagnosis methods are difficult to identify emerging fault types because the training sample cannot contain all fault categories, resulting in poor diagnostic performance for emerging fault categories.
Using a fusion method based on improved stack autoencoder and label transfer algorithm, the hidden layer features of the training samples and test samples are obtained through stack autoencoder training, and these features are input into the label transfer algorithm to diagnose the fault types of mechanical equipment, including Build a traditional stack autoencoder and tag delivery algorithm model, extract hidden layer features and diagnose fault types.
It has achieved a high accuracy rate for mechanical equipment to diagnose existing fault types and identify new fault types, with an average diagnostic accuracy rate of 92.5%, effectively solving the problem of identifying new fault types.
Smart Images

Figure CN114282600B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of mechanical equipment fault diagnosis, and in particular to a method for diagnosing new fault types of equipment based on an improved stack autoencoder. Background Art
[0002] During the life cycle of mechanical equipment, changes in the mechanical environment will cause deviations between the stress of parts and components and the normal state. As the equipment works for a long time, the surface of the parts is prone to wear, cracks and other changes. If the early minor damage of the parts cannot be diagnosed in time, once fatigue worsens and the parts break, the mechanical equipment will be shut down. Therefore, monitoring, diagnosing and analyzing mechanical equipment is an effective measure to avoid continued deterioration of parts after damage.
[0003] The basic framework of fault diagnosis can be composed of three parts: signal preprocessing, feature extraction and fault identification. Due to its easy collection, vibration signals are widely used to characterize the state of mechanical equipment. However, the vibration signals collected usually have more interference components and non-stationary and nonlinear characteristics, which also brings difficulties to fault diagnosis. In recent years, deep learning has received widespread attention in the field of fault diagnosis due to its end-to-end signal processing method. The biggest advantage of this type of method is that it learns hidden discriminative features from input signals through multi-layer neural networks. Although the performance of some deep learning-based diagnostic models is satisfactory, the common problem of these fault diagnosis methods is that they assume that the category of the test sample belongs to the training sample. However, the training sample cannot contain all fault categories, which means that the long-term operation of mechanical equipment may encounter new fault categories. Therefore, this kind of research that conforms to the actual operating state of the machinery and diagnoses new fault types is more meaningful. Summary of the invention
[0004] In response to the above-mentioned problems, the present invention aims to provide a method for diagnosing new fault types of equipment based on an improved stack autoencoder, which integrates the stack autoencoder with the label transfer algorithm to diagnose existing fault types of mechanical equipment as well as newly emerging fault types.
[0005] In order to achieve the above object, the technical solution adopted by the present invention is as follows:
[0006] A method for diagnosing new fault types of equipment based on an improved stacked autoencoder is characterized by comprising the following steps:
[0007] S1: Obtain training sample set and test sample set of equipment fault types;
[0008] S2: Preprocess each training sample and test sample;
[0009] S3: constructing a stacked autoencoder diagnostic model based on label transfer; the stacked autoencoder diagnostic model based on label transfer includes a traditional stacked autoencoder diagnostic model and a label transfer algorithm model;
[0010] S4: Use the training samples and corresponding sample labels as inputs to the traditional stacked autoencoder diagnostic model to complete the training of SAE;
[0011] S5: The training samples and test samples are used as the input of the traditional stacked autoencoder diagnostic model to extract hidden layer features;
[0012] S6: The hidden features of the training samples and the corresponding sample labels, as well as the hidden features of the test samples are used as inputs of the label transfer algorithm model, and the diagnosis results of the fault type are output.
[0013] Furthermore, the specific operation of step S1 includes the following steps:
[0014] S101: Select data of different equipment failure types from the equipment failure database to form a training sample set. Let the training sample set be X, and the training sample set X consists of N samples of C categories, that is, The corresponding sample label is y;
[0015] S102: In order to diagnose the newly added equipment fault type, the label y of the training sample is preprocessed, that is, the matrix of the label y of the training sample is increased from C columns to C+1 columns, then y i represents the label of the i-th training sample, Assume that the label of the first training sample belongs to the first category, then y1 = [1, 0, ..., 0] ∈ R C+1 ;
[0016] S103: The equipment failure data collected in real time by the data collection system is used as test samples. All test samples constitute a test sample set Q. The test sample set Q consists of M samples of C+1 categories.
[0017] Furthermore, the specific operation of step S2 includes constructing 11 features for each sample, then the training sample set X∈R N×11 , test sample set Q∈R M×11 ;
[0018] The characteristic parameters of the 11 features are: mean, standard deviation, root amplitude, root mean square value, maximum value, skewness, kurtosis, peak value, margin, shape factor and pulse index;
[0019] Among them: mean
[0020] Standard Deviation
[0021] Square root amplitude
[0022] RMS value
[0023] Maximum value p5 = max|x i |(i=1, 2, ..., n)
[0024] Skewness
[0025] Kurtosis
[0026] Peak
[0027] Margin
[0028] Form Factor
[0029] Pulse Indicator
[0030] Further, the stacked autoencoder diagnostic model based on label transfer described in step S3 includes a traditional stacked autoencoder diagnostic model and a label transfer algorithm model;
[0031] The traditional stacked autoencoder diagnostic model is a two-layer stacked autoencoder, including an input layer, a hidden layer h1, a hidden layer h2 and a Softmax layer. The input layer is connected to the hidden layer h1, the hidden layer h1 is connected to the hidden layer h2, and the hidden layer h2 is connected to the Softmax layer. The loss function of the Softmax layer is In the formula, s i Represents the output of the Softmax layer of the i-th training sample;
[0032] The label transfer algorithm model is also connected to the hidden layer h2. The features of the hidden layers h2 of the training samples and the test samples, as well as the labels of the training samples, are used as inputs of the label transfer algorithm model to obtain the label information of the test samples.
[0033] Furthermore, the neurons of hidden layer h1 and hidden layer h2 are 600 and 100 respectively, and the learning rate is 0.01.
[0034] The beneficial effects of the present invention are:
[0035] 1. The present invention is based on a method for diagnosing new fault types of equipment based on an improved stacked autoencoder, and integrates the stacked autoencoder with the label transfer algorithm. The hidden features of the training samples and the test samples are first obtained through stacked autoencoder training, and then the hidden features of the training samples and the test samples, as well as the labels corresponding to the training samples are input into the label transfer algorithm. This method can not only diagnose the existing fault types of mechanical equipment, but also diagnose the equipment fault types that do not exist in the training sample set.
[0036] 2. Experimental verification shows that the method for diagnosing new equipment fault types based on improved stacked autoencoders proposed in the present invention can achieve an average diagnostic accuracy of up to 92.5% for existing equipment fault types and non-existent equipment fault types in the training sample set, indicating that the present invention is effective and feasible. BRIEF DESCRIPTION OF THE DRAWINGS
[0037] Figure 1 This is a structural diagram of the stacked autoencoder diagnostic model based on label transfer in the present invention.
[0038] Figure 2 This is the fault diagnosis model of the two-layer stacked autoencoder in the present invention.
[0039] Figure 3 This is a test bench for a certain type of artillery swing mechanism in Example 1 of the present invention.
[0040] Figure 4 This is a diagram of the damage of the pressure plate in Example 1 of the present invention.
[0041] Figure 5 This is a crack diagram of the roller in Example 1 of the present invention.
[0042] Figure 6 This is a time domain diagram of the vibration signal of the swing-spring mechanism in the normal state in the first embodiment of the present invention.
[0043] Figure 7 This is a time domain diagram of the vibration signal of the pressure plate damage of the swing-spring mechanism in Example 1 of the present invention.
[0044] Figure 8 This is a time domain diagram of the vibration signal of the roller crack of the swing-spring mechanism in the first embodiment of the present invention.
[0045] Fig. 9 These are the diagnosis results of each model in Example 1 of the present invention. DETAILED DESCRIPTION
[0046] In order to enable those skilled in the art to better understand the technical solution of the present invention, the technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.
[0047] The method for diagnosing new fault types of equipment based on the improved stack autoencoder includes the following steps:
[0048] S1: Obtain training sample set and test sample set of equipment fault types;
[0049] Specifically, S101: select data of different equipment failure types from the equipment failure database to form a training sample set, and let the training sample set be X, and the training sample set X consists of N samples of C categories, that is, The corresponding sample label is y;
[0050] S102: In order to diagnose the newly added equipment fault type, the label y of the training sample is preprocessed, that is, the matrix of the label y of the training sample is increased from C columns to C+1 columns, then y i represents the label of the i-th training sample, Assume that the label of the first training sample belongs to the first category, then y1 = [1, 0, ..., 0] ∈ R C+1 ;
[0051] S103: The equipment failure data collected in real time by the data collection system is used as test samples. All test samples constitute a test sample set Q. The test sample set Q consists of M samples of C+1 categories.
[0052] Further, S2: preprocess each training sample and test sample;
[0053] Specifically, 11 features are constructed for each sample, and the training sample set X∈R N×11 , test sample set Q∈R M×11 ;
[0054] The characteristic parameters of the 11 features are: mean, standard deviation, root amplitude, root mean square value, maximum value, skewness, kurtosis, peak value, margin, waveform factor and pulse index; the calculation method of the 11 characteristic parameters is shown in Table 1 below.
[0055] Table 1 Characteristic parameters
[0056]
[0057] Where x(i) is the time domain signal sequence, i = 1, 2, ..., N, N is the number of sample points, and the time domain characteristic parameters p1 and p3-p5 describe the amplitude and energy changes of the time domain signal; p2 and p6-p 11 Describes the time series distribution of time domain signals.
[0058] Further, S3: construct a stacked autoencoder diagnostic model (LPSAE model) based on label transfer; the stacked autoencoder diagnostic model based on label transfer includes a traditional stacked autoencoder diagnostic model and a label transfer algorithm model, as shown in the attached Figure 1 shown.
[0059] Specifically, the traditional stacked autoencoder diagnostic model is a two-layer stacked autoencoder (SAE model), as shown in the attached Figure 2 As shown in the figure, the SAE model is trained using training samples and their labels; it includes the input layer, hidden layer h1, hidden layer h2 and Softmax layer. The input layer is connected to the hidden layer h1, the hidden layer h1 is connected to the hidden layer h2, and the hidden layer h2 is connected to the Softmax layer. After two AE unsupervised learning, the SAE training is completed, and the features of the hidden layer h1 and the hidden layer h2 can be obtained. The hidden layer h2 is connected to the Softmax layer, that is, the features of the hidden layer h2 are used as the input of the Softmax layer. Combined with the label information of the input layer data, the Softmax layer is supervised for learning. The loss function of the Softmax layer is: In the formula, s i Represents the Softmax layer output of the i-th training sample; through iterative training, the output of the Softmax layer is continuously compared with the real label information of the input layer data, thereby completing the training of the fault diagnosis model based on the stacked autoencoder.
[0060] The label transfer algorithm model is also connected to the hidden layer h2. After completing the training of SAE, the training samples and test samples are used as the input of SAE to extract the features of the hidden layer h2. Then, the features of the hidden layer h2 of the training samples and test samples, as well as the labels of the training samples, are used as the input of the label transfer algorithm model to obtain the label information of the test samples. The neurons of the hidden layer h1 and the hidden layer h2 in LPSAE are set to 600 and 100 respectively, and the learning rate is 0.01.
[0061] It should be noted here that both the stacked autoencoder and the label transfer algorithm adopt the principles of the existing stacked autoencoder and label transfer algorithm, which will not be described in detail in the present invention.
[0062] Further, S4: the training sample set X∈R N×11 The training samples in and the corresponding sample labels y are used as the input of the SAE diagnosis model to complete the SAE training;
[0063] Further, S5: respectively train the sample set X∈R N×11 The training samples in and the test sample set Q∈R M×11 The test sample in is used as the input of SAE to extract the features of hidden layer h2. The features of training samples in hidden layer h2 are represented by X h2 Indicates that there is X h2 ∈R N×100 , the feature of the test sample in the hidden layer h2 is Q h2 Indicates that there is Q h2 ∈RN×100 ;
[0064] S6: The hidden features X of the training samples h2 ∈R N×100 And the corresponding sample label y, and the hidden layer feature Q of the test sample h2 ∈R N×100 As the input of the label transfer algorithm model, the fault type diagnosis result is output, that is, the diagnosis result of the LPSAE model.
[0065] Embodiment 1:
[0066] The data of Example 1 comes from a bench test of a certain type of artillery swing mechanism. Figure 3 As shown in the figure, sensors are arranged on the test bench to collect the vibration acceleration signals generated by the pressure plate and roller movement during the upward and downward swing of the swing arm, with a sampling frequency of 10kHz. Among them, the pressure plate is opened and closed during the recoil and recoil of the barrel, and its damage is as follows Figure 4 As shown in the figure, the roller rolls in the groove during the recoil process, and the crack condition is as follows Figure 5 shown.
[0067] The experimental data includes three types of states: normal state of the swing spring mechanism, damage to the pressure plate, and cracks in the roller. The vibration signals of these three states are as follows: Figure 6-8 As shown, from Figure 6-8 It can be seen that the impact vibration in each stage of the swing is large. The data composition is shown in Table 2 below. Data set D consists of training data set D1, training data set D2 and test data set D0, which are used to test the model's ability to diagnose new fault types. Among them, D1-D0 means that a total of 80 samples of 2 categories (normal, pressure plate damage) are used to train the model, and 60 samples of 3 categories (new roller crack types) are tested; D2-D0 means that a total of 80 samples of 2 categories (normal, roller crack) are used to train the model, and 60 samples of 3 categories (new pressure plate damage types) are tested.
[0068] Table 2 Composition of experimental dataset
[0069]
[0070] In order to evaluate the fault diagnosis ability of the LPSAE model proposed in this invention, the diagnosis results of support vector machine (SVM), SAE, and convolution neural network (CNN) are compared (CNN network structure comes from the literature: Hu Xiaoyi. Research on bearing fault identification based on deep convolutional neural network of CNN-SVM [J]. Vibration and Shock, 2019, 38 (18): 173-178). Each diagnostic model is run 10 times, and the diagnostic accuracy is averaged. Among them, the kernel function of SVM selects the radial basis kernel function, and the kernel parameter is 0.1; the LPSAE model involves two parameters in the label transfer algorithm: the radial basis kernel function σ is 1, and the number of neighbors k of each sample can be the number of training samples of any type, where k is 40.
[0071] For data set D, the diagnostic results of each model are as follows: Fig. 9 As shown in Table 3. Fig. 9 As can be seen from Table 3, whether it is data set D1-D0 or D2-D0, the diagnosis results of SVM, SAE, and CNN models do not exceed 66.67%, indicating that these three models cannot diagnose new fault types. Among them, the diagnosis results of SAE and CNN models are both 66.67%, indicating that they can correctly diagnose the existing fault types in the test samples. Taking D1-D0 as an example, the test samples include 3 types of samples, 2 types of fault types that already exist in the training samples: normal, pressure plate damage, and a new fault type: roller crack. The SAE and CNN models have a 100% diagnosis accuracy for the existing 2 types of fault types (normal, pressure plate damage), and a 0% diagnosis accuracy for the new fault type (roller crack). In comparison, the LPSAE proposed in the present invention can not only diagnose existing fault types, but also identify newly appeared fault types. For example, for the D1-D0 data set, 80 samples of 2 categories (normal, platen damage) are used to train the LPSAE model, and the average diagnostic accuracy of 60 samples of 3 categories (new roller crack types) is 92.5%. This result shows that the LPSAE model has a 100% diagnostic accuracy for the existing 2 types of fault types (normal, platen damage) and a 87.01% diagnostic accuracy for the newly appeared fault type (roller crack). This also verifies the effectiveness of the improved stacked autoencoder proposed in the present invention in diagnosing newly appeared fault types.
[0072] Table 3 Diagnosis results of swing-spring mechanism damage experiment (%)
[0073]
[0074] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.
Claims
1. A method for diagnosing new fault types of equipment based on improved stack autoencoders, It is characterized in that The following steps are included: S1: Obtain training sample set and test sample set of equipment fault types; S2: Preprocess each training sample and test sample; S3: constructing a stacked autoencoder diagnostic model based on label transfer; the stacked autoencoder diagnostic model based on label transfer includes a traditional stacked autoencoder diagnostic model and a label transfer algorithm model; S4: Use the training samples and corresponding sample labels as inputs to the traditional stacked autoencoder diagnostic model to complete the training of SAE; S5: The training samples and test samples are used as the input of the traditional stacked autoencoder diagnostic model to extract hidden layer features; S6: taking the hidden features of the training samples and the corresponding sample labels, as well as the hidden features of the test samples as inputs to the label transfer algorithm model, and outputting the diagnosis results of the fault type; The stacked autoencoder diagnostic model based on label transfer described in step S3 includes a traditional stacked autoencoder diagnostic model and a label transfer algorithm model; The traditional stacked autoencoder diagnostic model is a two-layer stacked autoencoder, including an input layer, a hidden layer h1, a hidden layer h2 and a Softmax layer. The input layer is connected to the hidden layer h1, the hidden layer h1 is connected to the hidden layer h2, and the hidden layer h2 is connected to the Softmax layer. The loss function of the Softmax layer is In the formula, s i Represents the output of the Softmax layer of the i-th training sample; The label transfer algorithm model is also connected to the hidden layer h2. The features of the hidden layers h2 of the training samples and the test samples, as well as the labels of the training samples, are used as inputs of the label transfer algorithm model to obtain the label information of the test samples.
2. The method for diagnosing new fault types of equipment based on improved stacked autoencoder according to claim 1, It is characterized in that The specific operation of step S1 includes the following steps: S101: Select data of different equipment failure types from the equipment failure database to form a training sample set. Let the training sample set be X, and the training sample set X consists of N samples of C categories, that is, The corresponding sample label is y; S102: In order to diagnose the newly added equipment fault type, the label y of the training sample is preprocessed, that is, the matrix of the label y of the training sample is increased from C columns to C+1 columns, then y i represents the label of the i-th training sample, Assuming that the label of the first training sample belongs to the first category, then y 1 =[1,0,…,0]∈R C+1 ; S103: The equipment failure data collected in real time by the data collection system is used as test samples. All test samples constitute a test sample set Q. The test sample set Q consists of M samples of C+1 categories.
3. The method for diagnosing new fault types of equipment based on improved stacked autoencoder according to claim 2, Features: The specific operation of step S2 includes constructing 11 features for each sample, then the training sample set X∈R N×11 , test sample set Q∈R M×11 ; The characteristic parameters of the 11 features are: mean, standard deviation, root amplitude, root mean square value, maximum value, skewness, kurtosis, peak value, margin, shape factor and pulse index; Among them: mean Standard Deviation Square root amplitude RMS value Maximum value p 5 =max|x i |(i=1, 2, ..., n) Skewness Kurtosis Peak Margin Form Factor Pulse Indicator 4. The method for diagnosing new fault types of equipment based on improved stacked autoencoder according to claim 1, Features: The number of neurons in the hidden layer h1 and the hidden layer h2 is 600 and 100 respectively, and the learning rate is 0.01.
