An elevator traction sheave fault judgment method based on an improved DANN model

Abstract

The present invention discloses a method for judging elevator traction sheave faults based on an improved DANN model, including: collecting target domain data and source domain data and recording domain classification labels; performing decomposition and reconstruction on the target domain and source domain data by using wavelet packet decomposition to obtain reconstructed signals; extracting time-frequency features of the reconstructed signals by using a Markov transition matrix to construct a Markov transition field as input features; establishing an improved DANN model; pre-training the feature extractor and classifier of the improved DANN model by using the input features corresponding to the source domain data; inputting the input features and domain classification labels corresponding to the target domain data and the source domain data into the pre-trained model for adversarial training to obtain a final DANN model; and inputting the signal to be detected into the final DANN model to obtain a fault diagnosis result. This method solves the problem of insufficient samples, fully extracts relevant features of vibration signals, combines transfer learning to quickly and accurately diagnose elevator faults, and is convenient for equipment safety maintenance.

Description

Technical Field

[0001] The present invention belongs to the technical field of elevator fault diagnosis, and particularly relates to a method for judging faults of an elevator traction sheave based on an improved DANN model. Background Art

[0002] With the continuous acceleration of the urban development speed, as a vertical and fast means of transportation, elevators have become an indispensable part of high-rise buildings. However, due to frequent daily operation and the prominent problems such as "aging" of elevators, elevator failures and even entrapment accidents occur frequently. The traditional detection method is to conduct regular inspection and maintenance on each elevator device, but it has the disadvantages of over-maintenance, time-consuming, high cost, and poor real-time performance, and at the same time, it cannot completely avoid the occurrence of equipment failures. Therefore, how to be able to display the current operating state of the elevator in real time, give early warnings before faults occur, and change the passive situation of dealing with faults only after they occur has become an urgent problem to be solved.

[0003] As the motion center of the whole elevator, the traction sheave controls the whole process of the elevator from starting, accelerating upward, decelerating and braking to finally standing still and hovering. During the operation of the elevator, the traction sheave is in a state of high-load operation all the time, which will accelerate the wear of the traction sheave and cause abnormal vibration of the traction sheave. Therefore, the wear degree of the traction sheave directly affects the safety performance of the elevator. At the same time, in the actual application of elevators, the data of different elevators are often collected under different working conditions, and there are also differences in the distribution of elevator data of different users and elevator data of test towers. Even due to manufacturing costs and privacy protection, it is impossible to obtain the elevator traction sheave data of users. Therefore, finding appropriate fault diagnosis and identification methods has an urgent sense of the times for effectively diagnosing faults of elevator traction sheaves.

[0004] At present, with the development of big data and artificial intelligence technologies, fault diagnosis has also been widely studied. However, since the artificial intelligence algorithm model is data-driven, combining relevant algorithms to achieve fault diagnosis and prediction often requires a large amount of data, especially fault data. Due to the particularity of elevator equipment, it is difficult to collect fault data, and thus it is impossible to ensure that there are enough training samples to support model training. Therefore, transfer learning needs to be introduced. However, traditional transfer learning is mainly applied to image processing, has poor adaptability to vibration signals, does not fully utilize the time-domain, frequency-domain and spatial-domain features of vibration signals, and the obtained accuracy is not high. Therefore, it is urgent to develop new elevator fault diagnosis methods to solve the problem of insufficient samples while fully extracting the relevant features of vibration signals and combining transfer learning algorithms to achieve efficient and accurate fault diagnosis. Summary of the Invention

[0005] The object of the present invention is to address the above problems and propose a method for judging the faults of elevator traction wheels based on an improved DANN model. While solving the problem of insufficient samples, it can fully extract the relevant features of vibration signals, and combine transfer learning to achieve rapid and accurate diagnosis of elevator faults, facilitating the safety maintenance of intelligent devices.

[0006] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0007] A method for judging the faults of elevator traction wheels based on an improved DANN model proposed by the present invention includes the following steps:

[0008] S1. Collect the vibration signals during the operation of the elevator as the target domain data and record the domain classification labels. The target domain data includes normal samples and traction wheel fault samples;

[0009] S2. Use the elevator 3D motion model to simulate the normal operation state and the traction wheel fault operation state of the elevator, collect the corresponding sample data, label them as the source domain data, and record the domain classification labels;

[0010] S3. Use wavelet packet decomposition to decompose and reconstruct the target domain data and the source domain data to obtain the reconstructed signals;

[0011] S4. Use the Markov transition matrix to extract the time-frequency features of the reconstructed signals and construct the Markov transition field as the input features;

[0012] S5. Establish an improved DANN model. The improved DANN model includes a feature extractor, a classifier, a domain discriminator, and a gradient reversal layer. The feature extractor is respectively connected to the input ends of the classifier and the gradient reversal layer, and the output end of the gradient reversal layer is connected to the input end of the domain discriminator. Among them, the feature extractor includes a first convolutional module, a second convolutional module, a third convolutional module, a fully connected layer, and a Faltten layer connected in sequence. The first convolutional module, the second convolutional module, and the third convolutional module each include a convolutional layer and a pooling layer connected in sequence. The domain discriminator includes a fourth convolutional module and a fully connected layer connected in sequence. The fourth convolutional module includes at least one convolutional layer;

[0013] S6. Use the input features corresponding to the source domain data to pre-train the feature extractor and the classifier of the improved DANN model;

[0014] S7. Input the input features and domain classification labels corresponding to the target domain data and the source domain data into the pre-trained improved DANN model for adversarial training until the model converges to obtain the final DANN model;

[0015] S8. Input the vibration signal to be detected into the final DANN model to obtain the fault diagnosis result.

[0016] Preferably, the wavelet packet decomposition is a three-layer wavelet packet decomposition, and the structural formula is as follows:

[0017]

[0018] Among them, the high-pass filter coefficient and the low-pass filter coefficient satisfy the orthogonal relationship of formula (2):

[0019] g(k) = (-1) k h(1 - k) (2)

[0020] The decomposed signals obtained at different layers are calculated according to formulas (3) and (4):

[0021]

[0022]

[0023] In the formula, S 0,0 (m) represents the original signal before decomposition at the current time series, m is the discrete time series, k is the time shift factor, S i+1,2j (m) represents the decomposed signal of the current time series signal at the 2j-th node of the (i + 1)-th layer, S i+1,2j+1 (m) represents the decomposed signal of the current time series signal at the (2j + 1)-th node of the (i + 1)-th layer, S i , j (k) represents the decomposed signal corresponding to the j-th node of the i-th layer of the current time series signal, i = 0, 1, 2, j = 0 to 7, and when j is even, it represents the low-frequency component signal decomposed by the low-pass filter coefficient g(k), and when j is odd, it represents the high-frequency component signal decomposed by the high-pass filter coefficient h(k).

[0024] Preferably, the convolutional layer of the first convolutional module uses 3 7×7 convolutional kernels, the convolutional layer of the second convolutional module uses 16 3×3 convolutional kernels, and the convolutional layer of the third convolutional module uses 60 3×3 convolutional kernels.

[0025] Preferably, the input features and domain classification labels corresponding to the target domain data and the source domain data are input into the pre-trained improved DANN model for adversarial training, including:

[0026] S71. Map the input feature x through the feature extractor f = G f (x; θ f ) to convert it into a D-dimensional feature vector, that is And map the feature vector corresponding to the source domain data through the classifier G y (x; θ y ) to output the corresponding classification label prediction result, and map the feature vectors corresponding to the source domain data and the target domain data through the domain discriminator G d (x; θd )Output the corresponding domain classification results, where θ f is the weight parameter of the feature extractor, θ y is the weight parameter of the classifier, θ d is the weight parameter of the domain discriminator;

[0027] S72. Calculate the classification loss and the domain discrimination loss as shown in Equation (5):

[0028]

[0029] In the formula, is the classification loss of the i-th sample, is the domain discrimination loss of the i-th sample, L y is the classification loss, L d is the domain discrimination loss, y i is the classification label corresponding to the i-th sample, d i is the domain classification label corresponding to the i-th sample;

[0030] Then the total loss function is as shown in Equation (6):

[0031]

[0032] In the formula, n is the number of samples in the source domain data, n' is the number of samples in the target domain data, N is the total number of samples, and λ is the parameter of the gradient reversal layer;

[0033] S73. Use the gradient reversal layer to automatically reverse the gradient direction during backpropagation and perform an identity transformation during forward propagation, as shown in Equation (7):

[0034]

[0035] In the formula, is the output feature of the gradient reversal layer, is the input feature of the gradient reversal layer, and I is the current gradient;

[0036] And the update formula of the λ value is as shown in Equation (9):

[0037]

[0038] In Equation (9), λ p is the λ value in the p stage, p represents the relative value of the iteration process, that is, the ratio of the current iteration number to the total iteration number, and γ = 10.

[0039] Preferably, input the input features and domain classification labels corresponding to the target domain data and the source domain data into the pre-trained improved DANN model for adversarial training, including:

[0040] S71. Map the input feature x through the feature extractor f = G f (x; θ f ) to a D-dimensional feature vector, that is And map the feature vector corresponding to the source domain data through the classifier G y (x; θ y ) to output the corresponding classification label prediction result. Map the feature vectors corresponding to the source domain data and the target domain data through the domain discriminator G d (x; θ d ) to output the corresponding domain classification result, where θ f is the weight parameter of the feature extractor, θ y is the weight parameter of the classifier, θ d is the weight parameter of the domain discriminator;

[0041] S72. Calculate the classification loss and the domain discrimination loss as shown in Equation (5):

[0042]

[0043] In the formula, is the classification loss of the i-th sample, is the domain discrimination loss of the i-th sample, L y is the classification loss, L d is the domain discrimination loss, y i is the classification label corresponding to the i-th sample, d i is the domain classification label corresponding to the i-th sample;

[0044] S73. Use the gradient reversal layer to automatically reverse the gradient direction during the backpropagation process and perform an identity transformation during the forward propagation process, as shown in Equation (7):

[0045]

[0046] In the formula, is the output feature of the gradient reversal layer, is the input feature of the gradient reversal layer, and I is the current gradient;

[0047] Then the total loss function is as shown in Equation (8):

[0048]

[0049] And the update formula for the λ value is as shown in Equation (9):

[0050]

[0051] In Equation (9), λ pThe λ value for the p stage, where p represents the relative value of the iteration process, i.e., the ratio of the current iteration number to the total iteration number, and γ = 10.

[0052] Preferably, inputting the input features and domain classification labels corresponding to the target domain data and the source domain data into the pre-trained improved DANN model for adversarial training further includes:

[0053] Updating the learning rate as shown in Equation (10):

[0054]

[0055] In Equation (10), μ0 is the initial learning rate, with a value of 0.01, and μ p is the learning rate for the p stage, where p represents the relative value of the iteration process, i.e., the ratio of the current iteration number to the total iteration number, and α and β are hyperparameters, α = 10, β = 0.75.

[0056] Preferably, both the classification loss and the domain discrimination loss adopt the maximum mean discrepancy loss function.

[0057] Compared with the prior art, the beneficial effects of the present invention are as follows: By acquiring the vibration signals of the elevator traction sheave during the elevator operation process and the elevator 3D motion model simulation process, performing wavelet packet decomposition and Markov transition on the vibration signals to form input features, pre-training the feature extractor and classifier of the improved DANN model with the input features corresponding to the source domain data, and then inputting the input features corresponding to the target domain data and the source domain data into the improved DANN model for adversarial training, optimizing the parameters to enable the model to have a better fault diagnosis effect, so as to achieve fast and accurate diagnosis of elevator faults. While solving the problem of insufficient samples, it can fully extract the relevant features of the vibration signals, and combine transfer learning to achieve efficient and accurate fault diagnosis, which is convenient for the safety maintenance of intelligent devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 is a flowchart of the elevator traction sheave fault judgment method based on the improved DANN model of the present invention;

[0059] Figure 2 is the signal of each node after wavelet packet decomposition and the corresponding spectrogram of the present invention;

[0060] Figure 3 is the structural diagram of the improved DANN model of the present invention;

[0061] Figure 4 is the experimental result diagram of the training accuracy and loss of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0062] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present application without creative efforts shall fall within the protection scope of the present application.

[0063] It should be noted that unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the technical field to which this application belongs. The terms used in the specification of this application herein are only for the purpose of describing specific embodiments and are not intended to limit this application.

[0064] In order to overcome the problem of low accuracy in elevator fault diagnosis in the prior art, a method for judging elevator faults based on big data feature analysis is proposed.

[0065] As Figures 1-4 shown, a method for judging the faults of elevator traction wheels based on an improved DANN model includes the following steps:

[0066] S1. Collect the vibration signals during the operation of the elevator as target domain data and record the domain classification labels. The target domain data includes normal samples and traction wheel fault samples. For example, use a high-precision sensor to collect the vibration data of the traction wheel during the operation of the elevator, which contains a large number of normal samples and a small number of fault data in the worn state of the traction wheel, and label this part of the data set as target domain data.

[0067] S2. Use the elevator 3D motion model to simulate the normal operation state and the traction wheel fault operation state of the elevator, collect the corresponding sample data, label them as source domain data, and record the domain classification labels. The sample data in the traction wheel fault operation state is, for example, the fault data in the worn state of the traction wheel.

[0068] S3. Use wavelet packet decomposition to decompose and reconstruct the target domain data and the source domain data to obtain the reconstructed signals.

[0069] In one embodiment, the wavelet packet decomposition is a three-layer wavelet packet decomposition, and the structural formula is as follows:

[0070]

[0071] Among them, the high-pass filter coefficient and the low-pass filter coefficient satisfy the orthogonal relationship of formula (2):

[0072] g(k) = (-1) k h(1 - k) (2) The decomposed signals obtained in different layers are calculated according to formulas (3) and (4):

[0073]

[0074]

[0075] In the formula, S 0,0 (m) represents the original signal before decomposition at the current time sequence, m is a discrete time series, k is a time shift factor, S i+1,2j (m) represents the decomposed signal of the current time sequence signal at the 2j-th node of the (i + 1)-th layer, S i+1,2j+1 (m) represents the decomposed signal of the current time sequence signal at the (2j + 1)-th node of the (i + 1)-th layer, S i,j (k) represents the decomposed signal corresponding to the j-th node of the i-th layer of the current time sequence signal, i = 0, 1, 2, j = 0 to 7, and when j is an even number, it represents the low-frequency component signal obtained by decomposition with the low-pass filter coefficient g(k), and when j is an odd number, it represents the high-frequency component signal obtained by decomposition with the high-pass filter coefficient h(k).

[0076] The signal is decomposed and reconstructed by three-layer wavelet packet decomposition to obtain 8 characteristic components. As Figure 2 shown, the left figure is the original signal before decomposition, and the decomposed signals of the 0th to 7th nodes of the 3rd layer are shown from top to bottom on the right.

[0077] S4. Extract the time-frequency features of the reconstructed signal using the Markov transition matrix to construct a Markov transition field as the input feature;

[0078] S5. Establish an improved DANN model. The improved DANN model includes a feature extractor, a classifier, a domain discriminator, and a gradient reversal layer. The feature extractor is respectively connected to the input ends of the classifier and the gradient reversal layer, and the output end of the gradient reversal layer is connected to the input end of the domain discriminator. Among them, the feature extractor includes a first convolution module, a second convolution module, a third convolution module, a fully connected layer, and a Flatten layer connected in sequence. The first convolution module, the second convolution module, and the third convolution module all include a convolution layer and a pooling layer connected in sequence. The domain discriminator includes a fourth convolution module and a fully connected layer connected in sequence. The fourth convolution module includes at least one convolution layer.

[0079] In an embodiment, the convolution layer of the first convolution module uses 3 7*7 convolution kernels, the convolution layer of the second convolution module uses 16 3*3 convolution kernels, and the convolution layer of the third convolution module uses 60 3*3 convolution kernels.

[0080] Among them, the improved DANN model improves the feature extractor, which includes a three-layer convolutional module, a fully connected layer, and a Flatten layer. In the first convolutional module, 3 large convolutional kernels of 7×7 are used for feature extraction. In the second convolutional module, 16 small convolutional kernels of 3×3 are used for feature extraction again. In the third convolutional module, 60 small convolutional kernels of 3×3 are used for in-depth feature mining, and then they are passed into the gradient reversal layer through the fully connected layer and the Flatten layer. Larger convolutional kernels can be set to suppress interference. The convolutional layer is used to extract the shallow features of the vibration signal, and then continuous small convolutional kernels are used to expand the network depth to improve the network performance. The domain discriminator consists of a convolutional layer and a fully connected layer, which can effectively solve the problem of eliminating gradient disappearance. For example, the domain discriminator includes at least one convolutional layer (forming a convolutional structure) and a fully connected layer. The convolutional layer is used to better judge the domain to which the data belongs. For example, it is 1 to 3 convolutional layers, and the effect is better when it is 3 layers, but the training time is too long. To balance accuracy and efficiency, one convolutional layer is adopted in this embodiment.

[0081] Introduce the idea of the generative adversarial network into transfer learning, and make the data features in the target domain as close as possible to the source domain data through adversarial learning, so as to be not restricted by whether the data labels in the target domain are available, and it is suitable for both supervised learning and unsupervised learning. During the training process, the role of the feature extractor is to extract the common features of the corresponding input features of the source domain data and the target domain data. The role of the domain discriminator is to correctly judge the domain to which the data belongs. Through the adversarial training between the feature extractor and the domain discriminator, the data distribution in the target domain is continuously close to the data distribution in the source domain, thereby improving the classification accuracy of the model.

[0082] S6. Use the input features corresponding to the source domain data to pre-train the feature extractor and classifier of the improved DANN model.

[0083] S7. Input the input features and domain classification labels corresponding to the target domain data and the source domain data into the pre-trained improved DANN model for adversarial training until the model converges to obtain the final DANN model.

[0084] In one embodiment, inputting the input features and domain classification labels corresponding to the target domain data and the source domain data into the pre-trained improved DANN model for adversarial training includes:

[0085] S71. Map the input feature x through the feature extractor f = G f (x; θ f ) to convert it into a D-dimensional feature vector, that is, And map the feature vector corresponding to the source domain data through the classifier G y (x; θ y) The predicted classification label results corresponding to the mapped output are obtained. The feature vectors corresponding to the source domain data and the target domain data are both passed through the domain discriminator G d (x; θ d ) to output the corresponding domain classification results, where θ f are the weight parameters of the feature extractor, and θ y are the weight parameters of the classifier, and θ d are the weight parameters of the domain discriminator;

[0086] S72. Calculate the classification loss and the domain discrimination loss as shown in Equation (5):

[0087]

[0088] In the formula, is the classification loss of the i-th sample, is the domain discrimination loss of the i-th sample, and L y is the classification loss, and L d is the domain discrimination loss, y i is the classification label corresponding to the i-th sample, and d i is the domain classification label corresponding to the i-th sample;

[0089] Then the total loss function is as shown in Equation (6):

[0090]

[0091] In the formula, n is the number of samples in the source domain data, n' is the number of samples in the target domain data, N is the total number of samples, and λ is the parameter of the gradient reversal layer;

[0092] S73. Use the gradient reversal layer to automatically reverse the gradient direction during the backpropagation process and perform an identity transformation during the forward propagation process, as shown in Equation (7):

[0093]

[0094] In the formula, is the output feature of the gradient reversal layer, is the input feature of the gradient reversal layer, and I is the current gradient;

[0095] And the update formula for the λ value is as shown in Equation (9):

[0096]

[0097] In Equation (9), λ p is the λ value at the p stage, where p represents the relative value of the iteration process, that is, the ratio of the current iteration number to the total iteration number, and γ = 10.

[0098] In one embodiment, inputting the input features and domain classification labels corresponding to the target domain data and the source domain data into the pre-trained improved DANN model for adversarial training includes:

[0099] S71. Mapping the input feature x through the feature extractor f = G f (x; θ f ) to convert it into a D-dimensional feature vector, that is And outputting the corresponding classification label prediction result by mapping the feature vector corresponding to the source domain data through the classifier G y (x; θ y ). Outputting the corresponding domain classification result by mapping the feature vectors corresponding to both the source domain data and the target domain data through the domain discriminator G d (x; θ d ), where θ f is the weight parameter of the feature extractor, θ y is the weight parameter of the classifier, and θ d is the weight parameter of the domain discriminator;

[0100] S72. Calculating the classification loss and the domain discrimination loss as shown in Equation (5):

[0101]

[0102] In the formula, is the classification loss of the i-th sample, is the domain discrimination loss of the i-th sample, L y is the classification loss, L d is the domain discrimination loss, y i is the classification label corresponding to the i-th sample, and d i is the domain classification label corresponding to the i-th sample;

[0103] S73. Using the gradient reversal layer to automatically reverse the gradient direction during the backpropagation process and implement the identity transformation during the forward propagation process, as shown in Equation (7):

[0104]

[0105] In the formula, is the output feature of the gradient reversal layer, is the input feature of the gradient reversal layer, and I is the current gradient;

[0106] Then the total loss function is as shown in Equation (8):

[0107]

[0108] And the update formula for the λ value is as shown in Equation (9):

[0109]

[0110] In Equation (9), λ p is the λ value in the p stage, where p represents the relative value of the iteration process, that is, the ratio of the current iteration number to the total iteration number, and γ = 10.

[0111] In one embodiment, input features and domain classification labels corresponding to the target domain data and the source domain data are input into a pre-trained improved DANN model for adversarial training, and it further includes:

[0112] Updating the learning rate as shown in Equation (10):

[0113]

[0114] In Equation (10), μ0 is the initial learning rate, and its value is 0.01. μ p is the learning rate in the p stage, where p represents the relative value of the iteration process, that is, the ratio of the current iteration number to the total iteration number, and α and β are hyperparameters, α = 10, β = 0.75.

[0115] In one embodiment, both the classification loss and the domain discrimination loss adopt the maximum mean discrepancy loss function. The maximum mean discrepancy (MMD) is a technique well-known to those skilled in the art and will not be elaborated here.

[0116] Specifically, during the adversarial training of the improved DANN model, the input of the model is the source domain data with classification labels, the target domain data without classification labels, and the domain classification labels of the source domain data and the target domain data.

[0117] The input feature x of the improved DANN model will first pass through the feature extractor f = G f (x; θ f ) for mapping transformation into a D-dimensional feature vector, that is, Then the improved DANN model will be divided into two branches, namely the classifier G y (x; θ y ) and the domain discriminator G d (x; θ d ). The feature vector corresponding to the source domain data will pass through the mapping of G y (x; θ y ) to obtain the corresponding classification label prediction result. At the same time, the feature vectors of both the source domain data and the target domain data will pass through G d (x; θ d ) to obtain the corresponding domain classification result.

[0118] During the training phase, the following two tasks need to be completed. First, accurately classify the source domain data to minimize the image classification error. Second, confuse the source domain data and the target domain data to maximize the domain classification error. Then, the total loss function of the improved DANN model can be defined as shown in Equation (6) or (8).

[0119] To avoid the staged training method of separately fixing the parameters of the generator and discriminator like in the GAN model and achieve true end-to-end training, a new Gradient Reversal Layer (GRL) is proposed in the improved DANN model, which automatically reverses the gradient direction during the backpropagation process and performs an identity transformation during the forward propagation process. The relevant mathematical representation is shown in Equation (7) below.

[0120] The gradient reversal layer is mainly between the feature extractor and the domain discriminator. Then, during the backpropagation process, the gradient of the domain classification loss of the domain discriminator will be automatically reversed before backpropagating to the parameters of the feature extractor, thereby achieving the adversarial loss.

[0121] Of course, in the Gradient Reversal Layer (GRL), the parameter λ is not a fixed value but changes dynamically. Its change expression is shown in Equation (9). At the same time, in the improved DANN model, the learning rate also changes with the iteration process, and the change formula is shown in Equation (10).

[0122] S8. Input the vibration signal to be detected into the final DANN model to obtain the fault diagnosis result.

[0123] In this embodiment, the target domain data is divided into a training set and a test set. As Figure 4 shown, the abscissa in the figure is the number of iterations, and the ordinate is the accuracy rate. In the left figure, the solid line is the accuracy rate of the training set, up to 98%, and the dashed line is the accuracy rate of the test set, up to 94%. In the right figure, the solid line is the total loss of the training, the dashed line is the loss of the domain discriminator, and the dotted line is the loss of the classifier, all of which can converge stably to a small value.

[0124] This method obtains the vibration signal of the elevator traction wheel during the elevator operation process and the elevator 3D motion model simulation process, performs wavelet packet decomposition and Markov transfer on the vibration signal to form input features, pre-trains the feature extractor and classifier of the improved DANN model with the input features corresponding to the source domain data, and then inputs the input features corresponding to the target domain data and the source domain data into the improved DANN model for adversarial training. By optimizing the parameters, the model has a good fault diagnosis effect to achieve fast and accurate diagnosis of elevator faults. While solving the problem of insufficient samples, it can fully extract the relevant features of the vibration signal, combine transfer learning to achieve efficient and accurate fault diagnosis, and facilitate the safety maintenance of intelligent devices.

[0125] The technical features of the above-described embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification.

[0126] The above-described embodiments only express the embodiments of the present application that are relatively specific and detailed in description, but should not be construed as a limitation on the scope of the patent application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the appended claims.

Claims

1. A method for judging the faults of elevator traction wheels based on an improved DANN model, characterized in that: The elevator traction sheave fault judgment method based on the improved DANN model includes the following steps: S1. Collect vibration signals during the operation of the elevator as target domain data and record domain classification labels. The target domain data includes normal samples and traction sheave fault samples; S2. Use the elevator 3D motion model to simulate the normal operation state of the elevator and the traction sheave fault operation state, collect the corresponding sample data, label them as source domain data, and record domain classification labels; S3. Use wavelet packet decomposition to decompose and reconstruct the target domain data and source domain data to obtain reconstructed signals; S4. Use the Markov transition matrix to extract the time-frequency features of the reconstructed signals and construct a Markov transition field as input features; S5. Establish an improved DANN model. The improved DANN model includes a feature extractor, a classifier, a domain discriminator, and a gradient reversal layer. The input end of the feature extractor is respectively connected to the classifier and the gradient reversal layer, and the output end of the gradient reversal layer is connected to the input end of the domain discriminator. Among them, the feature extractor includes a first convolution module, a second convolution module, a third convolution module, a fully connected layer, and a Faltten layer connected in sequence. The first convolution module, the second convolution module, and the third convolution module each include a convolution layer and a pooling layer connected in sequence. The domain discriminator includes a fourth convolution module and a fully connected layer connected in sequence. The fourth convolution module includes at least one convolution layer; S6. Use the input features corresponding to the source domain data to pre-train the feature extractor and classifier of the improved DANN model; S7. Input the input features and domain classification labels corresponding to the target domain data and source domain data into the pre-trained improved DANN model for adversarial training until the model converges to obtain the final DANN model; S8. Input the vibration signal to be detected into the final DANN model to obtain a fault diagnosis result.

2. The elevator traction sheave fault judgment method based on the improved DANN model according to claim 1, characterized in that: The wavelet packet decomposition is a three-layer wavelet packet decomposition, and the structural formula is as follows: Among them, the high-pass filter coefficient and the low-pass filter coefficient satisfy the orthogonal relationship of formula (2): g(k) = (-1) k h(1 - k) (2) The decomposed signals obtained in different layers are calculated according to formulas (3) and (4): Where S 0,0 (m) represents the original signal before decomposition at the current time series, m is the discrete time series, k is the time shift factor, and S i+1,2j (m) represents the decomposed signal of the current time series signal at the 2j-th node of the (i + 1)-th layer, and S i+1,2j+1 (m) represents the decomposed signal of the current time series signal at the (2j + 1)-th node of the (i + 1)-th layer, and S i,j (k) represents the decomposed signal corresponding to the j-th node of the i-th layer of the current time series signal, where i = 0, 1, 2, j = 0 to 7, and when j is an even number, it represents the low-frequency component signal obtained by decomposition with the low-pass filter coefficient g(k), and when j is an odd number, it represents the high-frequency component signal obtained by decomposition with the high-pass filter coefficient h(k).

3. The elevator traction sheave fault judgment method based on the improved DANN model according to claim 1, characterized in that: The convolution layer of the first convolution module uses 3 7*7 convolution kernels, the convolution layer of the second convolution module uses 16 3*3 convolution kernels, and the convolution layer of the third convolution module uses 60 3*3 convolution kernels.

4. The elevator traction sheave fault judgment method based on the improved DANN model according to claim 1, characterized in that: The inputting the input features and domain classification labels corresponding to the target domain data and source domain data into the pre-trained improved DANN model for adversarial training includes: S71. Map the input feature x through the feature extractor f = G f (x; θ f ) into a D-dimensional feature vector, that is And map the feature vector corresponding to the source domain data through the classifier G y (x; θ y ) to output the corresponding classification label prediction result. Map the feature vectors corresponding to the source domain data and the target domain data through the domain discriminator G d (x; θ d ) to output the corresponding domain classification result, where θ f is the weight parameter of the feature extractor, θ y is the weight parameter of the classifier, and θ d is the weight parameter of the domain discriminator; S72. Calculate the classification loss and domain discrimination loss as shown in formula (5): Wherein, is the classification loss of the i-th sample, is the domain discrimination loss of the i-th sample, and L y is the classification loss, and L d is the domain discrimination loss, y i is the classification label corresponding to the i-th sample, and d i is the domain classification label corresponding to the i-th sample; Then the total loss function is as shown in formula (6): In the formula, n is the number of samples of the source domain data, n' is the number of samples of the target domain data, N is the total number of samples, and λ is the parameter of the gradient reversal layer; S73. Use the gradient reversal layer to automatically reverse the gradient direction during the backpropagation process and perform an identity transformation during the forward propagation process, as shown in formula (7): In the formula, is the output feature of the gradient reversal layer, is the input feature of the gradient reversal layer, and I is the current gradient; And the λ value update formula is as shown in formula (9): In formula (9), λ p is the value of λ in the p stage, where p represents the relative value of the iteration process, that is, the ratio of the current iteration number to the total iteration number, and γ = 10.

5. The elevator traction sheave fault judgment method based on the improved DANN model according to claim 1, characterized in that: The inputting the input features and domain classification labels corresponding to the target domain data and source domain data into the pre-trained improved DANN model for adversarial training includes: S71. Map the input feature x through the feature extractor f = G f (x; θ f ) into a D-dimensional feature vector, that is And map the feature vector corresponding to the source domain data through the classifier G y (x; θ y ) to output the corresponding classification label prediction result. Map the feature vectors corresponding to both the source domain data and the target domain data through the domain discriminator G d (x; θ d ) to output the corresponding domain classification result, where θ f is the weight parameter of the feature extractor, θ y is the weight parameter of the classifier, and θ d is the weight parameter of the domain discriminator; S72. Calculate the classification loss and the domain discrimination loss as shown in Equation (5): Wherein, is the classification loss of the i-th sample, is the domain discrimination loss of the i-th sample, L y is the classification loss, L d is the domain discrimination loss, y i is the classification label corresponding to the i-th sample, d i is the domain classification label corresponding to the i-th sample; S73. Use the gradient reversal layer to automatically reverse the gradient direction during backpropagation and perform an identity transformation during forward propagation, as shown in Equation (7): In the formula, is the output feature of the gradient reversal layer, is the input feature of the gradient reversal layer, and I is the current gradient; Then the total loss function is as shown in Equation (8): And the update formula for the λ value is as shown in Equation (9): In Equation (9), λ p is the value of λ in the p stage, where p represents the relative value of the iteration process, that is, the ratio of the current iteration number to the total iteration number, and γ = 10.

6. The elevator traction sheave fault judgment method based on the improved DANN model according to any one of claims 4 or 5, characterized in that: The step of inputting the input features and domain classification labels corresponding to the target domain data and the source domain data into the pre-trained improved DANN model for adversarial training further includes: Update the learning rate, as shown in Equation (10): In Equation (10), μ0 is the initial learning rate, and its value is 0.

01. μ p is the learning rate at the p-th stage. p represents the relative value of the iteration process, that is, the ratio of the current iteration number to the total iteration number. α and β are hyperparameters, where α = 10 and β = 0.

75.

7. The elevator traction sheave fault judgment method based on the improved DANN model according to any one of claims 4 or 5, characterized in that: Both the classification loss and the domain discrimination loss adopt the maximum mean discrepancy loss function.

Images