Elevator fault diagnosis method based on one-dimensional convolutional neural network and meta-learning

Abstract

The present invention provides an elevator fault diagnosis method based on a one-dimensional convolutional neural network and meta-learning. The method first obtains the acceleration signal generated by the elevator car during its movement. The acceleration signal is then preprocessed and denoised using wavelet soft thresholding. The contour features of the denoised data are extracted. The data dimension is fixed using resampling technology to generate a complete data set. The complete data set is then partitioned according to the N-way K-shot principle to meet the requirements of the meta-learning strategy. The partitioned data is then fed into a one-dimensional convolutional neural network in batches, and the model is trained and tested using the meta-learning strategy. This method can effectively reduce the difficulty of extracting elevator operation features and obtain an elevator fault diagnosis model with strong generalization ability and high accuracy using only a small number of samples.

Description

Technical Field

[0001] The present invention belongs to the technical field of elevator data analysis, and in particular relates to an elevator fault diagnosis method based on one-dimensional convolutional neural network and meta-learning. Background Art

[0002] With the rapid development of social productivity in recent years, the number of elevators has also increased. People's focus on elevators has gradually shifted from operating speed to comfort and safety. The elevator car is not only the main component of the elevator, but also the part closest to the user. Its safe operation is directly related to the safety of users' lives and property.

[0003] The elevator car, as the main component of an elevator, is closely connected to other elevator components. Failures in these components can significantly cause abnormal car operation. If the elevator car experiences abnormal operation, its vertical acceleration signal will also fluctuate accordingly. Therefore, it is necessary to find a suitable method to diagnose the vertical acceleration signal during elevator car operation.

[0004] When an elevator malfunctions, the vertical acceleration data of the car will show significant fluctuations. These fluctuations contain information about the elevator fault. Extracting fault characteristics from acceleration signals has become an unavoidable topic in the field of elevator fault diagnosis.

[0005] The existing traditional approach is to extract the time domain and frequency domain features of the acceleration signal, manually select the extracted features, and then combine them with a support vector machine to complete fault diagnosis. The number of acceleration samples is small, this method relies on expert experience, and the diagnostic model has poor generalization ability. This makes the use of the vertical acceleration signal generated by the car during operation for elevator fault diagnosis inaccurate, and it is very difficult to determine elevator faults. Summary of the Invention

[0006] Therefore, the technical problem to be solved by the present invention is to provide an elevator fault diagnosis method based on one-dimensional convolutional neural network and meta-learning, which can solve the problems of small number of traditional elevator car acceleration signal samples, difficult feature extraction process, and weak generalization ability of the diagnostic model, resulting in inaccurate elevator fault diagnosis using the vertical acceleration signal generated during car operation, and great difficulty in judging elevator faults.

[0007] To solve the above problems, the present invention provides an elevator fault diagnosis method based on a one-dimensional convolutional neural network and meta-learning, comprising the following steps:

[0008] Step 1: Obtain vertical acceleration data generated by the elevator cars during operation, and perform data preprocessing on the acceleration signals of the elevator cars;

[0009] Step 2: Perform noise reduction on the pre-processed acceleration signal data;

[0010] Step 3: Extract the contour features of the acceleration signal data after noise reduction and resample it to construct a complete data set;

[0011] Step 4: Divide the complete dataset into training set and test set;

[0012] Step 5: According to the requirements of the meta-learning strategy, the existing training set is divided into a training set support set and a training set query set, and the existing test set is divided into a test set support set and a test set query set;

[0013] Step 6: Construct a feature extraction network and a feature classification network based on a one-dimensional convolutional neural network. Use the divided training set support set, training set query set, test set support set, and test set query set as network inputs, and use a meta-learning strategy to train the feature extraction network and feature classification network based on a one-dimensional convolutional neural network.

[0014] Optionally, in step 1, the vertical acceleration data generated by the elevator cars during operation are obtained, and the acceleration signals of the elevator cars are preprocessed using a zero-mean normalization method. The normalization formulas are as follows:

[0015]

[0016]

[0017] Among them, μ is the mean of the sample data; σ is the standard deviation of the sample data; is the amplitude of the sample point; N is the number of sample points, and z is the standardized sample data.

[0018] Optionally, step 2 includes:

[0019] Step 2.1: The pre-processed acceleration signal data is recorded as S(t), and the acceleration signal S(t) is decomposed by wavelet for n times. The acceleration signal S obtained after the j-th decomposition is j (w), S j (w) is composed of detail component CD j and approximate component CA j The calculation formulas of wavelet transform are as follows:

[0020]

[0021] S j (w)=CD j +CA j j=1…n (4)

[0022]

[0023] in, is the wavelet basis function, a is the scaling scale of the wavelet function, b is the translation of the wavelet function, w is the angular frequency, and n is the number of layers of wavelet decomposition;

[0024] Step 2.2: Decompose the acceleration signal S obtained by wavelet decomposition j (w) Perform soft threshold denoising. Wavelet soft threshold denoising is performed on the basis of wavelet decomposition. j The soft threshold noise reduction method is used for noise reduction. The calculation formulas of the soft threshold noise reduction method are as follows:

[0025]

[0026]

[0027]

[0028] in, is the jth wavelet decomposition detail component after denoising, CA j is the approximate component of the jth wavelet decomposition, S′ j (w) is the recombined signal, λ is the noise reduction threshold, N is the signal length, sign(x) is the sign function: when x>0, sign(x)=1, when x=0, sign(x)=0, when x<0, sign(x)=-1, and n is the number of layers of wavelet decomposition;

[0029] Step 2.3: All the detail components after noise reduction With all the untreated approximate components CA j Reconstruction is performed to obtain the acceleration signal S′(t) generated in the vertical direction of the elevator car after noise reduction.

[0030] Optionally, step 3 includes:

[0031] Step 3.1: Collect the status signals generated during the operation of the elevator. The status signals reflect the operating status of the elevator, including: up signal, down signal, and door closing signal. The status signals and acceleration signals are consistent in time. According to the characteristics of acceleration changes, the operation process of the elevator can be divided into four processes: ascending acceleration, ascending deceleration, descending acceleration, and descending deceleration.

[0032] Step 3.2: Extract the index corresponding to the door closing signal and the upward signal, intercept and save the vertical acceleration signal of the elevator car when it moves upward from s′(t) and record it as S′up (t); extract the index corresponding to the door closing signal and the downward signal, intercept and save the vertical acceleration signal of the elevator car when it moves downward from s′(t) and record it as S′ down (t);

[0033] Step 3.3: Read the intercepted acceleration signal S′ up (t) and S′ down (t), respectively obtain S′ up (t) and S′ down (t) The index corresponding to the maximum amplitude point and the minimum amplitude point in the signal, calculated as formula (9):

[0034]

[0035] Among them, Index is the index acquisition function, min(S′ up (t)) is the minimum amplitude of the uplink signal, max(S′ up (t)) is the maximum amplitude of the uplink signal, min(S′ down (t)) is the minimum amplitude of the downlink signal, max(S′ down (t)) is the maximum amplitude value of the downlink signal;

[0036] Step 3.4: From S′ up (t) starts searching from the maximum and minimum amplitude indexes of the index, respectively, to both sides of the index until the amplitude reaches 0, stops searching, records the final index range, intercepts the corresponding acceleration signal and calculates the amplitude sum, where the process where the amplitude sum is greater than 0 is the rising acceleration process and its corresponding acceleration signal is S′ up,speed (t), the process where the amplitude and value are less than 0 is the rising and decelerating process, and the corresponding acceleration signal is S′ up,slow (t), save the corresponding process; from S′ down (t) starts searching from the maximum and minimum amplitude indexes of the index, respectively, and stops searching when the amplitude reaches 0. The final index range is recorded, the corresponding acceleration signal is intercepted, and the amplitude sum is calculated. The process where the amplitude sum is less than 0 is a descending acceleration process, and its corresponding acceleration signal is S′. down,speed (t), the process where the amplitude and value are greater than 0 is the descending deceleration process, and the corresponding acceleration signal is S′ down,slow (t), save the corresponding process;

[0037] Step 3.5: For S′ up,speed (t), S′ up,slow (t), S′ down,speed (t), S′ down,slow(t) The corresponding acceleration signal is resampled to 800 dimensions and a curve is drawn. The data is manually classified to produce a complete data set.

[0038] Optionally, when dividing the training set and the test set in step 4, W categories of data are extracted from the complete data set without replacement to form the training set, and the remaining categories of data constitute the test set, and the data categories of the training set and the test set do not overlap.

[0039] Optionally, step 5 includes:

[0040] Step 5.1: Each time, randomly extract N types of data from the training set data without replacement, and then randomly extract K + M samples from the N types of data. The first K samples form the training set support set, and the remaining M samples are the training set query set. All the data extracted each time constitute a task;

[0041] Step 5.2: Each time, randomly extract N types of data from the test set data without replacement, and then randomly extract K + M samples from the N types of data. The first K samples constitute the support set of the test set, and the remaining M samples constitute the query set of the test set. All the data extracted each time constitute a task.

[0042] Optionally, step 6 includes:

[0043] Step 6.1: The one-dimensional convolutional neural network consists of two parts: a feature extraction network and a feature classification network. The process of constructing the one-dimensional convolutional neural network is to construct the feature extraction network and the feature classification network in sequence.

[0044] Step 6.2: The feature extraction network consists of four convolutional layers and four regularization layers. The parameters of the convolutional layers and regularization layers are the same. The convolutional layer has 64 convolution kernels, the convolution kernel size is 2, the stride is 2, and the activation function is ReLU.

[0045] Step 6.3: The feature classification network uses a fully connected layer and a Softmax function to classify existing features. The features extracted by the convolution layer are stretched and combined into a column vector, which is the input of the feature extraction and classification network. The feature classification network contains two fully connected layers. The first fully connected layer is connected to the last layer of the feature extraction network and the activation function is Relu. The second fully connected layer is connected to the first fully connected layer and its activation function is the Softmax function. The first fully connected layer has 256 neurons, and the second fully connected layer has N neurons.

[0046] Step 6.4: Set the learning rate of the base model to 0.1, divide the data of the training set support set into batches of 10 tasks, feed the data into the feature extraction network in batches, feed the features extracted by the feature extraction network into the feature classification network for classification, use the cross entropy function to obtain the sample loss of the batch training set support set, use the backpropagation algorithm to update the parameters of the feature extraction network and feature classification network, and complete the update of the base model;

[0047] Step 6.5: Set the learning rate of the meta-learning model to 0.001, divide the training set and query set data into batches of 10 tasks, feed the data batches into the updated model for feature extraction and classification, use the cross-entropy function to calculate the average loss of all samples in the training set and query set, and use the backpropagation algorithm to fine-tune the parameters of the feature extraction network and feature classification network to complete the meta-learning model update;

[0048] Step 6.6: After every 10 iterations, the model's generalization ability needs to be tested. The actual classification model learning rate is set to 0.01. The test set and query set data are divided into batches of 10 tasks. The data is fed into the meta-learning model for calculation. The cross-entropy function is used to obtain the loss of the batch samples. The back-propagation algorithm is used to update the parameters of the feature extraction network and the feature classification network to complete the actual classification model training.

[0049] Step 6.7: Feed the data in the test set query set into the actual classification model in batches for calculation to obtain the loss and accuracy of the actual classification model. The loss and accuracy are used as indicators to evaluate the classification and generalization capabilities of the model.

[0050] Step 6.8: Repeat steps 6.1 to 6.7 to complete the training of the elevator fault diagnosis model.

[0051] Beneficial effects

[0052] The elevator fault diagnosis method based on a one-dimensional convolutional neural network and meta-learning provided in an embodiment of the present invention first obtains the acceleration signal generated by the elevator car during movement, then preprocesses the acceleration signal and performs wavelet soft threshold noise reduction; extracts the contour features of the noise-reduced data, uses resampling technology to fix the data dimension, and generates a complete data set; divides the complete data set according to the N-way K-shot principle to meet the requirements of the meta-learning strategy, and feeds the divided data into a one-dimensional convolutional neural network in batches to complete model training and testing using the meta-learning strategy; This method can effectively reduce the difficulty of extracting elevator operation features and obtain an elevator fault diagnosis model with strong generalization ability and high accuracy using a small number of samples. It solves the problem that the existing traditional approach extracts time-domain and frequency-domain features of the acceleration signal, manually selects the extracted features, and then combines them with a support vector machine to complete fault diagnosis. The number of acceleration samples is small, this method relies on expert experience, and the generalization ability of the diagnostic model is not strong. This makes elevator fault diagnosis using the vertical acceleration signal generated by the car operation inaccurate, and it is very difficult to determine elevator faults. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 This is a flow chart of an elevator fault diagnosis method according to an embodiment of the present invention;

[0054] Figure 2 A schematic diagram of three-layer wavelet decomposition according to an embodiment of the present invention;

[0055] Figure 3 This is a comparison diagram of the wavelet soft threshold noise reduction effect of an embodiment of the present invention;

[0056] Figure 4 This is a structural diagram of the feature extraction network and feature classification network according to an embodiment of the present invention. DETAILED DESCRIPTION

[0057] See also Figures 1 to 4 As shown, according to an embodiment of the present invention, an elevator fault diagnosis method based on a one-dimensional convolutional neural network and meta-learning is shown in FIG. Figure 1 , including the following steps:

[0058] Step 1: Obtain vertical acceleration data generated by the elevator cars during operation, and perform data preprocessing on the acceleration signals of the elevator cars;

[0059] The vertical acceleration signal generated during the operation of the elevator car is collected and preprocessed. The data preprocessing method used in the present invention is zero-mean normalization. The normalization formulas are as follows:

[0060]

[0061]

[0062] Among them, μ is the mean of the sample data; σ is the standard deviation of the sample data; is the amplitude of the sample point; N is the number of sample points, and z is the standardized sample data.

[0063] After the acceleration signal is normalized using zero mean, the data unit impact caused by inconsistent sensor models during data acquisition can be eliminated.

[0064] Step 2: Perform noise reduction on the preprocessed acceleration signal data. To reduce the interference of external unknown factors on the acceleration signal and improve the accuracy of fault diagnosis, the preprocessed acceleration signal data needs to be subjected to noise reduction.

[0065] Specifically include:

[0066] Step 2.1: The pre-processed acceleration signal data is recorded as S(t), and the acceleration signal S(t) is decomposed by wavelet for n times. The acceleration signal obtained after the j-th decomposition is S j (w), S j (w) is composed of detail component CD j and approximate component CA j The calculation formulas of wavelet decomposition are as follows:

[0067]

[0068] S j (w)=CD j +CA j j=1…n (4)

[0069]

[0070] in, is the wavelet basis function, a is the scaling scale of the wavelet function, b is the translation of the wavelet function, w is the angular frequency, and n is the number of layers of wavelet decomposition;

[0071] Further, the sym8 wavelet basis function used in wavelet transform can be found in Figure 2 The decomposition level is three. The acceleration signal S(t) is decomposed to obtain the approximate component CA1 and the detail component CD1. CA1 is decomposed again to obtain the second-decomposition approximate component CA2 and the detail component CD2. CA2 is decomposed again to obtain the third-decomposition approximate component CA3 and the detail component CD3.

[0072] Step 2.2: Decompose the acceleration signal S obtained by wavelet decompositionj (w) Perform soft threshold denoising. Wavelet soft threshold denoising is performed on the basis of wavelet transform. j The soft threshold noise reduction method is used for noise reduction. The calculation formulas of the soft threshold noise reduction method are as follows:

[0073]

[0074]

[0075]

[0076] in, is the detail component of the jth wavelet decomposition after denoising, CA j is the approximate component after the jth wavelet decomposition, S′ j (w) is the recombined signal, λ is the noise reduction threshold, N is the signal length, n is the number of wavelet decomposition layers, sign(x) is the sign function: when x>0, sign(x)=1, when x=0, sign(x)=0, when x<0, sign(x)=-1;

[0077] Furthermore, the threshold function used in soft threshold denoising is the maximum and minimum threshold.

[0078] Step 2.3: All the detail components after noise reduction With all untreated approximate components CA j Reconstruct and obtain the acceleration signal S′(t) generated in the vertical direction of the elevator car after noise reduction. Please refer to Figure 3 After noise reduction, the signal fluctuation is reduced, the outline is clear, and the characteristics of the signal are more prominent.

[0079] Step 3: Extract the contour features of the acceleration signal data after noise reduction and resample it to build a complete data set. To eliminate the interference of unknown information and improve data utilization, it is necessary to extract the contour features of the acceleration signal; to avoid the adverse effects of different data dimensions on model training, the contour features need to be resampled.

[0080] Specifically include:

[0081] Step 3.1: Collect the status signals generated during the elevator's operation. Unlike acceleration signals, status signals reflect the elevator's operating status and include: up signals, down signals, and door closing signals. Status signals and acceleration signals are temporally consistent. Based on the characteristics of acceleration changes, the elevator's operation can be divided into four stages: ascending acceleration, ascending deceleration, descending acceleration, and descending deceleration. Extracting the contour features of the operation process can fully utilize the data, reduce the interference of unnecessary information, and improve the accuracy of the model.

[0082] Step 3.2: Extract the index corresponding to the door closing signal and the upward signal, intercept and save the vertical acceleration signal S′ of the elevator car when it moves upward corresponding to the index up (t); Extract the index corresponding to the door closing signal and the downward signal, intercept and save the vertical acceleration signal S' of the elevator car when it moves downward corresponding to the index down (t);

[0083] Step 3.3: Read the intercepted acceleration signal S′ up (t) and S′ down (t), respectively obtain S′ up (t) and S′ down (t) The index corresponding to the maximum amplitude point and the minimum amplitude point in the signal, calculated as formula (9):

[0084]

[0085] Among them, Index is the index acquisition function, min(S′ up (t)) is the minimum amplitude of the uplink signal, max(S′ up (t)) is the maximum amplitude of the uplink signal, min(S′ down (t)) is the minimum amplitude of the downlink signal, max(S′ down (t)) is the maximum amplitude value of the downlink signal;

[0086] Step 3.4: From S′ up Start searching from the maximum and minimum amplitude indexes of (t) to both sides of the indexes respectively, and stop searching when the amplitude reaches 0. Record the final index range, intercept the corresponding acceleration signal and calculate the amplitude sum. The process where the amplitude sum is greater than 0 is the rising acceleration process, and its corresponding acceleration signal is S′ up,speed (t), the process where the amplitude and value are less than 0 is the rising and decelerating process, and the corresponding acceleration signal is S′ up,slow (t), save the corresponding process;

[0087] From S′ down Start searching from the maximum and minimum amplitude indexes of (t) to both sides of the indexes respectively, and stop searching when the amplitude is 0. Record the final index range, intercept the corresponding acceleration signal and calculate the amplitude sum. The process where the amplitude sum is less than 0 is a descending acceleration process, and its corresponding acceleration signal is S′ down,speed (t), the process where the amplitude and value are greater than 0 is the descending deceleration process, and the corresponding acceleration signal is S′ down,slow (t), save the corresponding process;

[0088] Step 3.5: For S′ up,speed(t), S′ up,slow (t), S′ down,speed (t), S′ down,slow (t) The corresponding acceleration signal is resampled to a fixed data dimension and a curve is drawn. The data is manually classified to produce a complete data set.

[0089] Step 4: The training set and test set required by the meta-learning strategy are different from the traditional training set and test set. When using the meta-learning strategy to divide the data set, all data of W categories must be randomly extracted from the complete data set without replacement to form the training set, and all data of the remaining categories must form the training set. The data in the training set is used for model training, and the data in the test set is used to test the actual performance of the model. The data categories of the training set and the test set do not overlap.

[0090] Step 5: Divide the training set into a training set support set and a training set query set, and divide the test set into a test set support set and a test set query set. Each time, randomly extract N types of data from the training set data or the test set data without replacement, and then randomly extract K+M samples from the N categories without replacement, where the number of samples in the support set is K and the number of samples in the query set is M.

[0091] A support set and a query set constitute a task. The dataset ultimately used for model training and testing should contain multiple tasks. The division of the support set and query set needs to refer to the N-way K-shot principle, that is, the dataset is divided based on the principle that there are N types of samples in a task and K samples in each type. The support set and query set are divided based on the existing training set and test set.

[0092] By dividing the training set data and the test set data in the above way, a large task can be divided into many different small tasks, which can greatly improve the efficiency of data utilization.

[0093] Specifically include:

[0094] Step 5.1: Each time, randomly extract N types of data from the training set without replacement. From the N types of data, randomly extract K + M samples without replacement to generate the training set support set and the training set query set. The training set support set has K samples, and the training set query set has M samples. All the data extracted each time constitute a task.

[0095] Step 5.2: Each time, randomly extract N types of data from the test set without replacement. From the N types of data, randomly extract K + M samples without replacement to generate the test set support set and the test set query set. The test set support set has K samples, and the test set query set has M samples. All the data extracted each time constitute a task.

[0096] Step 5.3: Task generation in the training set and test set is performed synchronously. Multiple tasks constitute a batch of tasks, and a batch of tasks is sent to the model for training.

[0097] Please refer to Figure 4 , Step 6: Construct a feature extraction network and a feature classification network based on a one-dimensional convolutional neural network. The feature extraction network includes four convolutional layers and four regularization layers. The parameters of the four convolutional layers are the same, and the parameters of the four regularization layers are the same; the feature classification network includes two fully connected layers. The first fully connected layer is connected to the last regularization layer of the feature extraction network, and the second fully connected layer is connected to the first fully connected layer.

[0098] The contour features of the vertical acceleration during the operation of the elevator car are used as model input, and the model is trained using a meta-learning strategy. The meta-learning strategy is a model-independent meta-learning method. During model training, data needs to be input into the fault diagnosis model according to the N-way K-shot principle. The training set support set is used to update the basic model, the training set query set is used to fine-tune the basic model and complete the update of the meta-learning model, the test set support set is used to update the actual classification model, and the test set query set is used to test the performance of the actual classification model.

[0099] The introduction of a regularization layer in the feature extraction network can effectively avoid overfitting during model training. The models trained using the meta-learning strategy often have strong generalization capabilities. When the model encounters a new task, a small amount of samples can be used to update the model, allowing the model to quickly meet the requirements of the new task.

[0100] Specifically include:

[0101] Step 6.1: The one-dimensional convolutional neural network consists of two parts: a feature extraction network and a feature classification network. The process of constructing the one-dimensional convolutional neural network is to construct the feature extraction network and the feature classification network in sequence.

[0102] Please refer to Figure 4, Step 6.2: The feature extraction network contains four convolutional layers and four regularization layers; the input dimension of the first convolutional layer is 1*800, the output dimension is 64*400, there are 64 convolution kernels, the size of the convolution kernel is 1*2, the convolution kernel step is 2, and the activation function is Relu; the input of the second convolutional layer is 64*400, the output dimension is 64*200, there are 64 convolution kernels, the size of the convolution kernel is 1*2, the convolution kernel step is 2, and the activation function is Relu; the input dimension of the third convolutional layer is 64*100, the output dimension is 64*50, there are 64 convolution kernels, the size of the convolution kernel is 1*2, the convolution kernel step is 2, and the activation function is Relu; the input dimension of the fourth convolutional layer is 65*50, the output dimension is 64*25, there are 64 convolution kernels in total, the convolution kernel size is 1*2, the convolution step is 2, and the activation function is ReLU; the input data dimension and output data dimension of the regularization layer are the same, the input and output dimensions of the first regularization layer are both 64*400, and the number of parameters to be regularized is 64+64*2; the input and output dimensions of the second regularization layer are both 64*200, and the number of parameters to be regularized is 64+64*2; the input and output dimensions of the third regularization layer are both 64*100, and the number of parameters to be regularized is 64+64*2; the input and output dimensions of the fourth regularization layer are both 64*50, and the number of parameters to be regularized is 64+64*2.

[0103] Step 6.3: The feature classification network uses a fully connected layer and a Softmax function to complete the classification of existing features. The features output by all convolution kernels in the fourth convolutional layer are stretched and combined into a column vector, which is the input of the feature extraction and classification network. The feature classification network contains two fully connected layers. The first fully connected layer is connected to the last layer of the feature extraction network and the activation function is Relu. The second fully connected layer is connected to the first fully connected layer and its activation function is the Softmax function. The first fully connected layer has 256 neurons, the input dimension is 3200*1, and the output dimension is 256*1. The second fully connected layer has N neurons, the input dimension is 256*1, and the output dimension is N*1.

[0104] Step 6.4: Set the learning rate of the base model to 0.1, divide the data of the training set support set into batches of 10 tasks, feed the data into the feature extraction network in batches, feed the features extracted by the feature extraction network into the feature classification network for classification, use the cross entropy function to obtain the sample loss of the batch training set support set, use the backpropagation algorithm to update the parameters of the feature extraction network and feature classification network, and complete the update of the base model;

[0105] Step 6.5: Set the learning rate of the meta-learning model to 0.001, divide the training set and query set data into batches of 10 tasks, feed the data batches into the updated model for feature extraction and classification, use the cross-entropy function to calculate the average loss of all samples in the training set and query set, and use the backpropagation algorithm to fine-tune the parameters of the feature extraction network and feature classification network to complete the meta-learning model update;

[0106] Step 6.6: After every 10 iterations, the model's generalization ability needs to be tested. The actual classification model learning rate is set to 0.01. The test set and query set data are divided into batches of 10 tasks. The data is fed into the meta-learning model for calculation. The cross-entropy function is used to obtain the loss of the batch samples. The back-propagation algorithm is used to update the parameters of the feature extraction network and the feature classification network to complete the actual classification model training.

[0107] Step 6.7: Feed the data in the test set query set into the actual classification model in batches for calculation to obtain the loss and accuracy of the actual classification model. The loss and accuracy are used as indicators to evaluate the classification and generalization capabilities of the model.

[0108] Step 6.8: Repeat steps 6.1 to 6.7 to complete the training of the elevator fault diagnosis model.

[0109] The present invention first obtains the vertical acceleration signal generated by the elevator car during movement, and then performs zero-mean normalization on the acceleration signal to eliminate the unit influence caused by different sensor models; uses wavelet soft threshold noise reduction to remove noise information in the acceleration signal, thereby improving the accuracy of diagnosis; extracts the contour features of the noise-reduced data to eliminate the interference of unknown factors and improve the utilization rate of data; uses resampling technology to fix the data dimension to avoid the adverse effects of data dimension on model training, and generates a complete data set; in order to meet the requirements of the meta-learning strategy, the completed data set is divided into a training set and a test set, and the categories in the training set and the test set are different; according to the N-way K-shot principle, the training set and the test set are divided into a training set support set and a training set support set. The data categories of the support set and query set are the same, and the two are generated at the same time to form a task. This division method can greatly improve the efficiency of data utilization; the divided tasks are sent to the feature extraction network and feature classification network based on the one-dimensional convolutional neural network in batches of 10 tasks and trained and tested using the meta-learning strategy. The feature extraction network based on the one-dimensional convolutional neural network can automatically extract features from the time series information, thereby avoiding the influence of human factors on the diagnostic effect. The model trained by the meta-learning strategy can have strong generalization ability. When encountering new tasks, only a small number of samples need to be used to update the model to adapt to the new task requirements.

[0110] It is easy for those skilled in the art to understand that, under the premise of no conflict, the above-mentioned advantageous methods can be freely combined and superimposed.

Claims

1. An elevator fault diagnosis method based on one-dimensional convolutional neural network and meta-learning, characterized by: The steps include: Step 1: Obtain vertical acceleration data generated by the elevator cars during operation, and perform data preprocessing on the acceleration signals of the elevator cars; Step 2: Perform noise reduction on the pre-processed acceleration signal data; The pre-processed acceleration signal data is subjected to wavelet decomposition, and the acceleration signal obtained after wavelet decomposition is subjected to soft threshold noise reduction; Step 3: Extract the contour features of the acceleration signal data after noise reduction and resample it to construct a complete data set; Step 4: Divide the complete dataset into training set and test set; Step 5: According to the requirements of the meta-learning strategy, the existing training set is divided into a training set support set and a training set query set, and the existing test set is divided into a test set support set and a test set query set; Step 6: Construct a feature extraction network and a feature classification network based on a one-dimensional convolutional neural network. Use the divided training set support set, training set query set, test set support set, and test set query set as network inputs, and use a meta-learning strategy to train the feature extraction network and feature classification network based on the one-dimensional convolutional neural network. Wherein, step 6 includes the following steps: Step 6.1: The process of building a one-dimensional convolutional neural network is to build a feature extraction network and a feature classification network in sequence; Step 6.2: The feature extraction network consists of four convolutional layers and four regularization layers. The parameters of the convolutional layers and regularization layers are the same. The convolutional layer has 64 convolution kernels, the convolution kernel size is 2, the stride is 2, and the activation function is ReLU. Step 6.3: The feature classification network uses a fully connected layer and a Softmax function to complete the classification of existing features. The features extracted by the convolution layer are stretched and combined into a column vector, which is the input of the feature extraction and classification network. The feature classification network contains two fully connected layers. The first fully connected layer is connected to the last layer of the feature extraction network and the activation function is Relu. The second fully connected layer is connected to the first fully connected layer and its activation function is the Softmax function. The first fully connected layer has 256 neurons, and the second fully connected layer has N neurons.

2. The elevator fault diagnosis method based on one-dimensional convolutional neural network and meta-learning according to claim 1 is characterized in that: In step 1, the vertical acceleration data generated by the elevator cars during operation are obtained, and the acceleration signals of the elevator cars are preprocessed using the zero-mean normalization method. The normalization formulas are as follows: Among them, μ is the mean of the sample data; σ is the standard deviation of the sample data; is the amplitude of the sample point; N is the number of sample points, and z is the standardized sample data.

3. The elevator fault diagnosis method based on one-dimensional convolutional neural network and meta-learning according to claim 1 is characterized in that: Step 2 includes: Step 2.1: The pre-processed acceleration signal data is recorded as S(t), and the acceleration signal S(t) is decomposed by wavelet for n times, where the acceleration signal after the j-th decomposition is S j (w), S j (w) is composed of detail component CD j and approximate component CA j The calculation formulas of wavelet transform are as follows: S j (w)=CD j +CA j j=1…n(4) in, is the wavelet basis function, a is the scaling scale of the wavelet function, b is the translation of the wavelet function, w is the angular frequency, and n is the number of layers of wavelet decomposition; Step 2.2: Decompose the acceleration signal S obtained by wavelet decomposition j (w) Perform soft threshold denoising. Wavelet soft threshold denoising is performed on the basis of wavelet decomposition. j The soft threshold noise reduction method is used for noise reduction. The calculation formulas of the soft threshold noise reduction method are as follows: in, is the jth wavelet decomposition detail component after denoising, CA j is the approximate component of the jth wavelet decomposition, S j ′(w) is the recombined signal, λ is the noise reduction threshold, N is the signal length, sign(x) is the sign function: when x>0, sign(x)=1, when x=0, sign(x)=0, when x<0, sign(x)=-1, and n is the number of layers of wavelet decomposition; Step 2.3: All the detail components after denoising With all the untreated approximate components CA j Reconstruction is performed to obtain the acceleration signal S′(t) generated in the vertical direction of the elevator car after noise reduction.

4. The elevator fault diagnosis method based on one-dimensional convolutional neural network and meta-learning according to claim 1 is characterized in that: Step 3 includes: Step 3.1: Collect the status signals generated during the operation of the elevator. The status signals reflect the operating status of the elevator, including: up signal, down signal and door closing signal. The status signals and acceleration signals are consistent in time. According to the characteristics of acceleration changes, the operation process of the elevator can be divided into four processes: ascending acceleration, ascending deceleration, descending acceleration, and descending deceleration. Step 3.2: Extract the index corresponding to the door closing signal and the upward signal, intercept and save the vertical acceleration signal of the elevator car when it moves upward from s′(t) and record it as S′ up (t); extract the index corresponding to the door closing signal and the downward signal, intercept and save the vertical acceleration signal of the elevator car when it moves downward from s′(t) and record it as S′ down (t); Step 3.3: Read the intercepted acceleration signal S′ up (t) and S′ down (t), respectively obtain S′ up (t) and S′ down (t) The index corresponding to the maximum amplitude point and the minimum amplitude point in the signal, calculated as formula (9): Among them, Index is the index acquisition function, min(S′ up (t)) is the minimum amplitude of the uplink signal, max(S′ up (t)) is the maximum amplitude of the uplink signal, min(S′ down (t)) is the minimum amplitude of the downlink signal, max(S′ down (t)) is the maximum amplitude value of the downlink signal; Step 3.4: From S′ up (t) starts searching from the maximum and minimum amplitude indexes of the index, respectively, to both sides of the index until the amplitude reaches 0, stops searching, records the final index range, intercepts the corresponding acceleration signal and calculates the amplitude sum, where the process where the amplitude sum is greater than 0 is the rising acceleration process and its corresponding acceleration signal is S′ up,speed (t), the process where the amplitude and value are less than 0 is the rising and decelerating process, and the corresponding acceleration signal is S′ up,slow (t), save the corresponding process; from S′ down (t) starts searching from the maximum and minimum amplitude indexes of the index, respectively, and stops searching when the amplitude reaches 0. The final index range is recorded, the corresponding acceleration signal is intercepted, and the amplitude sum is calculated. The process where the amplitude sum is less than 0 is a descending acceleration process, and its corresponding acceleration signal is S′. down,speed (t), the process where the amplitude and value are greater than 0 is the descending deceleration process, and the corresponding acceleration signal is S′ down,slow (t), save the corresponding process; Step 3.5: For S′ up,speed (t), S′ up,slow (t), S′ down,speed (t), S′ down,slow (t) The corresponding acceleration signal is resampled to 800 dimensions and a curve is drawn. The data is manually classified to produce a complete data set.

5. The elevator fault diagnosis method based on one-dimensional convolutional neural network and meta-learning according to claim 1 is characterized in that: When dividing the training set and test set in step 4, W categories of data are extracted from the complete data set without replacement to form the training set, and the remaining categories of data constitute the test set. The data categories of the training set and the test set do not overlap.

6. The elevator fault diagnosis method based on one-dimensional convolutional neural network and meta-learning according to claim 1 is characterized in that: Step 5 includes: Step 5.1: Each time, randomly extract N types of data from the training set data without replacement, and then randomly extract K + M samples from the N types of data. The first K samples form the training set support set, and the remaining M samples are the training set query set. All the data extracted each time constitute a task; Step 5.2: Each time, randomly extract N types of data from the test set data without replacement, and then randomly extract K + M samples from the N types of data. The first K samples constitute the support set of the test set, and the remaining M samples constitute the query set of the test set. All the data extracted each time constitute a task.

7. The elevator fault diagnosis method based on one-dimensional convolutional neural network and meta-learning according to claim 1 is characterized in that: Step 6 also includes: Step 6.4: Set the learning rate of the base model to 0.1, divide the data of the training set support set into batches of 10 tasks, feed the data into the feature extraction network in batches, feed the features extracted by the feature extraction network into the feature classification network for classification, use the cross entropy function to obtain the sample loss of the batch training set support set, use the backpropagation algorithm to update the parameters of the feature extraction network and feature classification network, and complete the update of the base model; Step 6.5: Set the learning rate of the meta-learning model to 0.001, divide the training set and query set data into batches of 10 tasks, feed the data batches into the updated model for feature extraction and classification, use the cross-entropy function to calculate the average loss of all samples in the training set and query set, and use the backpropagation algorithm to fine-tune the parameters of the feature extraction network and feature classification network to complete the meta-learning model update; Step 6.6: After every 10 iterations, the model's generalization ability needs to be tested. The actual classification model learning rate is set to 0.

01. The test set and query set data are divided into batches of 10 tasks. The data is fed into the meta-learning model for calculation. The cross-entropy function is used to obtain the loss of the batch samples. The back-propagation algorithm is used to update the parameters of the feature extraction network and the feature classification network to complete the actual classification model training. Step 6.7: Feed the data in the test set query set into the actual classification model in batches for calculation to obtain the loss and accuracy of the actual classification model. The loss and accuracy are used as indicators to evaluate the classification and generalization capabilities of the model. Step 6.8: Repeat steps 6.1 to 6.7 to complete the training of the elevator fault diagnosis model.

Images