Escalator main driving shaft bearing fault diagnosis method based on DAE-BiLSTM

Abstract

The invention provides an escalator main driving shaft bearing fault diagnosis method based on DAE-BiLSTM, and the method comprises the following steps: collecting a vibration signal of an escalator main driving shaft bearing, carrying out the normalization processing of the vibration signal, and dividing the vibration signal into a training set, a verification set, and a test set; constructing a DAE depth auto-encoder network model; inputting the training set into a DAE model for training; after training is completed, a DAE network encoder is reserved, and dimension reconstruction is carried out on a training set; a BiLSTM bidirectional long and short time memory neural network model is constructed; and inputting reconstruction data of the DAE network encoder into the BiLSTM model for training, evaluating a fault diagnosis effect by using the test set after training is completed, obtaining optimal parameters of the BiLSTM model by adopting a grid search mode, and finally obtaining an escalator main driving shaft bearing fault diagnosis model. The method has the advantages of being high in fault detection accuracy and high in diagnosis speed, the potential fault state of the bearing can be effectively pre-warned, and safe operation of the escalator is ensured.

Description

Technical Field

[0001] The invention relates to the technical field of mechanical fault diagnosis, and in particular to a method for diagnosing a main drive shaft bearing fault of an escalator based on DAE-BiLSTM. Background Art

[0002] The main drive shaft is the most important load-bearing component of the escalator. Its function is to receive power from the main drive unit, drive the steps and handrails to move, and bear large torque and torsional forces during operation. The bearing seats are installed at both ends of the main shaft and bear the load of the main shaft. Under long-term high-load operating pressure, the bearings are more susceptible to damage than other components. Normally, the life of the selected bearings should be at least greater than the overhaul cycle of the escalator. However, due to the influence of various internal and external factors, the designed service life is often not achieved. Therefore, it is necessary to diagnose and warn the operating status of the bearings. The structure of the main drive shaft is simpler than that of the main drive unit. When collecting vibration signals, the acceleration sensor is directly installed on the bearing seat. The shorter transmission path makes the signal less interfered by the outside world, and there are not many types of faults in the main drive shaft.

[0003] Currently, deep learning methods have been widely used in the diagnosis of mechanical parts, but traditional methods are difficult to guarantee the accuracy of data features. With the continuous progress and development of deep learning, DAE deep autoencoders extract nonlinear features of vibration signals through unsupervised learning, which can effectively reconstruct the dimensions of features and retain useful fault feature information; while BiLSTM bidirectional long short-term memory neural network can use forward and backward time series information to enhance the learning and memory of potential fault features in vibration signals, thereby improving the accuracy and robustness of fault diagnosis. Therefore, combining DAE with BiLSTM can more comprehensively extract vibration signal features and effectively realize fault diagnosis and early warning of main drive shaft bearings. Summary of the invention

[0004] The purpose of the present invention is to overcome the deficiencies of the prior art and propose a fault diagnosis method for the main drive shaft bearing of an escalator based on DAE-BiLSTM. The steps of the technical solution are as follows:

[0005] Step 1, collecting the vibration signal of the main drive shaft bearing of the escalator as the initial data set, normalizing it and then dividing the data set into a training set, a validation set and a test set;

[0006] Step 2: Build a DAE deep autoencoder network model and determine the hidden layer encoding dimension;

[0007] Step 3: Input the training set into the DAE model and train it through the DAE network encoder and DAE network decoder. The effect of DAE model training is verified by comparing the changes in the reconstruction error curve. After the training is completed, the DAE network encoder is retained and the dimension of the training set is reconstructed.

[0008] Step 4: Build a BiLSTM bidirectional long short-term memory neural network model and set various parameters of the model, including the number of training samples, batch size, and the number of hidden layer neurons;

[0009] Step 5: Input the reconstructed data of the DAE network encoder into the BiLSTM model for training, and use the grid search method to obtain the optimal parameters of the BiLSTM model;

[0010] Step 6: After training, use the test set to evaluate the fault diagnosis effect, and finally obtain the escalator main drive shaft bearing fault diagnosis model.

[0011] Furthermore, step 1 specifically includes the following sub-steps:

[0012] Step 1-1: Use a vibration sensor to collect vibration signals of the main drive shaft bearing of the escalator as an initial data set.

[0013] Step 1-2: Normalize the initial data set. Assume that X i is the i-th value in the initial data set, then the normalization formula can be expressed as follows:

[0014]

[0015] In the formula, X norm is the result of normalizing the initial data set, X min is the minimum value in the initial data set, X max is the maximum value in the initial data set.

[0016] Step 1-3: Divide the normalized data set into three parts: training set, validation set, and test set. The training set is used for DAE model fitting. During the training process, the training error is gradient-dropped to adjust the weights and bias vectors of the DAE model. The validation set is used to adjust the hyperparameters of the BiLSTM model to prevent overfitting. The test set is used to evaluate the fault diagnosis effect.

[0017] Furthermore, step 3 specifically includes the following sub-steps:

[0018] Step 3-1: Input the training set into the DAE model, set the number of training times and the number of hidden layer units of the model, and initialize the weights and bias vectors;

[0019] Step 3-2, perform forward propagation to obtain the output value;

[0020] Step 3-3, update the weight and bias vector of the DAE model through error back propagation;

[0021] Step 3-4: After training the DAE model, retain the DAE network encoder and reconstruct the dimension of the training set.

[0022] Furthermore, step 5 specifically includes the following sub-steps:

[0023] Step 5-1: Input the hidden layer of the DAE model into the BiLSTM network, perform forward and reverse learning on the reconstructed data of the DAE network encoder, use the forward and backward long short-term memory neural networks to extract past and future implicit information, learn the rules of different fault vibration signals, and finally calculate the forward and backward output values, and add them as the output of the hidden layer of the BiLSTM model.

[0024] Step 5-2, connect the Softmax layer, which is an output layer for classification tasks. The Softmax function converts the output real value into a probability distribution so that the sum of all output probabilities is 1. The predicted probability distribution is obtained based on the output of the hidden layer in the previous step, and the fault vibration signal samples are classified.

[0025] Step 5-3: Calculate the cross entropy loss function between the actual output and theoretical output of the BiLSTM model, and optimize the weights and bias vectors by backpropagation based on the gradient descent algorithm. During the training process, use grid search to traverse all possible hyperparameter combinations and evaluate the performance of each combination on the validation set. After each evaluation, record the loss and accuracy of the validation set, and finally select the BiLSTM model parameters that minimize the loss of the validation set.

[0026] The present invention proposes a fault diagnosis method for the main drive shaft bearing of an escalator based on DAE-BiLSTM. The method collects the vibration signal of the main drive shaft bearing through a vibration sensor, reconstructs the dimension of the vibration signal by using the DAE deep autoencoder network model, extracts the most representative features from it, and then inputs these features into the BiLSTM bidirectional long short-term memory neural network model for forward and backward learning. Finally, the fault type is classified and predicted through the Softmax layer, which effectively solves the deficiency of the traditional fault diagnosis method in manually extracting fault features, and can realize fault diagnosis and early warning conveniently and effectively. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0028] Figure 1 It is a flow chart of a method for diagnosing faults of main drive shaft bearings of an escalator based on DAE-BiLSTM proposed by the present invention.

[0029] Figure 2 It is a structural diagram of an escalator main drive shaft bearing fault diagnosis model based on DAE-BiLSTM proposed in the present invention.

[0030] Figure 3 Schematic diagram of data enhancement in an embodiment of the present invention.

[0031] Figure 4 It is the loss function of the DAE model training set and validation set in the embodiment of the present invention.

[0032] FIG. 5( a ) is a curve diagram of the accuracy of the fault diagnosis model in an embodiment of the present invention.

[0033] FIG5( b ) is a graph showing the loss rate of the fault diagnosis model in an embodiment of the present invention.

[0034] Figure 6 It is a fault diagnosis confusion matrix diagram based on the DAE-BiLSTM method in an embodiment of the present invention.

[0035] FIG. 7( a ) is a graph showing the accuracy of training sets of different models in an embodiment of the present invention.

[0036] FIG. 7( b ) is a graph showing the training set loss rate of different models in an embodiment of the present invention. DETAILED DESCRIPTION

[0037] In order to make the objectives, technical solutions and advantages of the present invention more clear, the implementation modes of the present invention will be further described in detail below in conjunction with the accompanying drawings and actual embodiments.

[0038] Reference Figure 1 and Figure 2 ,The present paper proposes a fault diagnosis method for the main driving shaft bearing of escalator based on DAE-BiLSTM.

[0039] Example

[0040] This embodiment takes the motor bearing fault as an example to explain the motor bearing fault diagnosis process in detail, and the steps are as follows:

[0041] Step 1, collecting the vibration signal of the main drive shaft bearing of the escalator, normalizing and dividing it into a training set, a verification set and a test set;

[0042] Step 1-1, this embodiment uses the bearing vibration signal of the Case Western Reserve University Bearing Data Center as the initial data set to verify whether the fault diagnosis model can accurately identify the fault type of the bearing. The bearing faults used in the bearing data set are relatively large in damage, and the fault characteristics are more obvious, making it easy to diagnose. At the same time, because bearings have similar vibration characteristics, the long short-term memory neural network can fully learn their fault characteristics.

[0043] In order to enhance the robustness and generalization ability of the training fault diagnosis model, the present invention uses a sliding window method to perform data enhancement on the vibration signal. The method for generating sample data is as follows: Figure 3 As shown, the vibration signal is divided by setting a sliding window to obtain a sample signal, and the window is moved according to a specified step length for sampling. When the sliding step length of the window is less than the length of a single sample, overlap occurs between samples. In this embodiment, the sliding window size is set to 1024, the fixed moving step size is 28, and the sampled sample signal is used as the initial data set.

[0044] Step 1-2: normalize the initial data set;

[0045] Step 1-3: Divide the normalized initial data set into three parts: training set, validation set, and test set. The training set is used to fit the DAE deep autoencoder network model. During the training process, the training error is gradient-dropped to adjust the weights and bias vectors of the DAE model. The validation set is used to adjust the hyperparameters of the BiLSTM bidirectional long short-term memory neural network model to prevent overfitting. The test set is used to evaluate the fault diagnosis effect.

[0046] In this embodiment, the normalized sample data is divided into 70% as a training set, 20% as a validation set, and 10% as a test set. The number of samples and labels of each fault mode are shown in Table 1.

[0047] Table 1 Composition of the number of samples of different faults

[0048]

[0049] Step 2, build a DAE deep autoencoder network model and set various parameters of the model. The following is an explanation of the role of each DAE model parameter based on this embodiment: ① Learning rate: The size of the learning rate will affect the convergence speed of the model. If the learning rate is set too small, the training time will be too long and overfitting will occur; if it is too large, the training effect will be poor. ② Batch processing: By dividing the training set into data batches of equal size, the consumption of resources during the training process can be reduced and the training speed can be accelerated. ③ Activation function: The DAE model selects the ReLU activation function, which can reduce the amount of calculation when backpropagating the error gradient; on the other hand, it can reduce the dependency between the connection weights, which reduces the occurrence of overfitting to a certain extent. ④ Output layer function: Select the Sigmoid function, and the normalized data can be restored during the decoding process. ⑤ Loss function: Select MSE, which has a good effect in evaluating the deviation between input and output. ⑥Optimizer: The optimizer will affect the training efficiency. During the training process, the stochastic gradient descent algorithm keeps a single learning rate unchanged when updating the weights. The Adam optimizer is stable and faster in training, and can adaptively adjust different learning rates. Therefore, the Adam algorithm is selected for optimizer training. The specific parameters of the DAE model are shown in Table 2.

[0050] Table 2DAE model parameters

[0051]

[0052] Step 3: Input the training set into the DAE model and train it through the DAE network encoder and DAE network decoder. The effect of the model training is verified by comparing the changes in the reconstruction error curve. After the training is completed, the DAE network encoder is retained and the dimension of the training set is reconstructed.

[0053] In this embodiment, the loss rate curves of the DAE model training set and the validation set are as follows: Figure 4 As shown in Figure 3, after 300 trainings, the loss function tends to be stable, and the model can be considered to have converged. After the encoding process of the DAE network encoder, the training set is reconstructed into 400-dimensional data, and the reconstructed feature data is shown in Table 3.

[0054] Table 3DAE model reconstruction data

[0055]

[0056] Step 4, construct a BiLSTM bidirectional long short-term memory neural network model, and set the parameters of the model, including the number of BiLSTM units, optimizer and learning rate. The number of hidden layer units represents the number of nodes used to memorize and store states, and the number of units will affect the diagnosis speed and accuracy. The BiLSTM model parameters set in this embodiment are shown in Table 4, setting the number of units of the BiLSTM model to 64, the learning rate to 0.001, the activation function to tanh, the number of training times to 200, and the dropout neuron inactivation probability to 0.2.

[0057] Table 4 BiLSTM model parameters

[0058]

[0059] Step 5: Input the reconstructed data of the DAE network encoder into the BiLSTM model for learning and training, then input the processing results of the BiLSTM model into the Softmax layer, and match the output layer dimension with the final number of classifications. Adopt the Adam optimizer to continuously update the BiLSTM model parameters according to the loss function.

[0060] Step 6: After training, use the validation set to evaluate the diagnostic effect of the BiLSTM model, and finally obtain the escalator main drive shaft bearing fault diagnosis model. The fault diagnosis model structure is as follows: Figure 2 shown.

[0061] In this embodiment, the changes in the accuracy and loss rate curves of the training set and the validation set of the bearing fault diagnosis model are shown in Figure 5. From the accuracy curve, it can be seen that in the initial stage, due to the random setting of the BiLSTM model weights, the training accuracy of the fault diagnosis model is low, and can only reach 72%. However, as the number of training times increases, the recognition accuracy quickly reaches more than 99%. After 150 trainings, the recognition accuracy is close to 100% and tends to be stable. From the loss rate curve of the training set, the loss rate fluctuates greatly in the first 25 trainings, but as the number of trainings increases, the loss curve gradually shows a smaller fluctuation, and finally the loss rate of the fault diagnosis model gradually stabilizes and approaches 0 after 150 trainings. It can be seen that as the number of trainings increases, accurate recognition of fault state diagnosis can be achieved. Figure 6 This is the confusion matrix of the diagnosis results of the bearing test set. The results show that except for a very small number of samples of the inner and outer rings that are incorrectly identified, the model can effectively identify the normal state, inner ring fault, outer ring fault, and rolling element fault of the main drive shaft bearing.

[0062] In order to verify the superiority of the method proposed in the present invention, the same batch size and learning rate were selected to compare the DAE-BiLSTM method with the method using only the BiLSTM model. The parameters of the BiLSTM model were set as follows: the number of input layer units was 1024, the number of units in the forward and backward hidden layers was 64, the activation function selected tanh, and finally the softmax layer was used for classification.

[0063] The accuracy curves of the training set after 30 trainings of the fault diagnosis model based on the DAE-BiLSTM method and the BiLSTM model are shown in Figure 7. Table 5 compares the training time of the two models. The fault diagnosis model based on the DAE-BiLSTM method has a large change in the first 12 iterations, and the number of training times is more when converging; however, the training speed is fast, and it quickly reaches a stable state, and the accuracy reaches more than 99%. As the number of training times increases, the accuracy curves of the two models overlap, and the training time of the fault diagnosis model based on the DAE-BiLSTM method is shorter.

[0064] Table 5 Comparison of diagnosis of different models

[0065]

[0066] From the above results, it can be seen that the escalator main drive shaft bearing fault diagnosis model based on DAE-BiLSTM proposed in the present invention has a similar accuracy to the BiLSTM model, but the former reconstructs the dimension through the deep autoencoder and compresses the 1024-dimensional input vibration signal to 400 dimensions. Under the premise of retaining the necessary fault features, the calculation cost is greatly reduced and the training time of the model is greatly shortened.

Claims

1. A fault diagnosis method for the main drive shaft bearing of an escalator based on DAE-BiLSTM, characterized in that: The following steps are involved: Step 1, collecting the vibration signal of the main drive shaft bearing of the escalator as the initial data set, normalizing it and then dividing the data set into a training set, a validation set and a test set; Step 2: Build a DAE deep autoencoder network model and determine the hidden layer encoding dimension; Step 3: Input the training set into the DAE model and train it through the DAE network encoder and DAE network decoder. The effect of DAE model training is verified by comparing the changes in the reconstruction error curve. After the training is completed, the DAE network encoder is retained and the dimension of the training set is reconstructed. Step 4: Build a BiLSTM bidirectional long short-term memory neural network model and set various parameters of the model, including the number of training samples, batch size, and the number of hidden layer neurons; Step 5: Input the reconstructed data of the DAE network encoder into the BiLSTM model for training, and use the grid search method to obtain the optimal parameters of the BiLSTM model; Step 6: After training, the test set is used to evaluate the diagnostic effect of the BiLSTM model, and finally the escalator main drive shaft bearing fault diagnosis model is obtained.

2. The escalator main drive shaft bearing fault diagnosis method based on DAE-BiLSTM according to claim 1 is characterized in that: The step 1 specifically includes the following sub-steps: Step 1-1, using a vibration sensor to collect vibration signals of the main drive shaft bearing of the escalator as an initial data set; Step 1-2: normalize the initial data set; Step 1-3: Divide the normalized data set into three parts: training set, validation set and test set; the training set is used for DAE model fitting, and the training error is gradient-decreased during the training process to adjust the weight and bias vector; the validation set is used to adjust the hyperparameters of the BiLSTM model to prevent overfitting; The test set is used to evaluate the fault diagnosis effect.

3. The escalator main drive shaft bearing fault diagnosis method based on DAE-BiLSTM according to claim 1 is characterized in that: The step 3 specifically includes the following sub-steps: Step 3-1: Input the training set into the DAE model, set the number of training times and the number of hidden layer units of the model, and initialize the weight and bias vector; Step 3-2, perform forward propagation to obtain the output value; Step 3-3, update the weight and bias vector of the DAE model through error back propagation; Step 3-4: After training the DAE model, retain the DAE network encoder and reconstruct the dimension of the training set.

4. The escalator main drive shaft bearing fault diagnosis method based on DAE-BiLSTM according to claim 1 is characterized in that: The step 5 specifically includes the following sub-steps: Step 5-1: Input the hidden layer of the DAE model into the BiLSTM bidirectional long short-term memory neural network model, perform forward and reverse learning on the reconstructed data of the DAE network encoder, respectively use the forward and backward LSTM long short-term memory neural networks to extract past and future implicit information, learn the rules of different fault vibration signals, and finally calculate the forward and backward output values, and add the two as the output of the hidden layer of the BiLSTM model; Step 5-2, connect the Softmax layer, obtain the prediction probability matrix according to the hidden layer output, and classify the fault vibration signal samples; Step 5-3: Calculate the cross entropy loss function between the actual output and theoretical output of the BiLSTM model, optimize the weights and bias vectors by backpropagation according to the gradient descent algorithm, and continuously optimize the BiLSTM model through training until the optimal parameters are obtained.

Images