A vibration fault diagnosis method for elevator door machine system

Abstract

The present invention discloses a vibration fault diagnosis method for an elevator door machine system, which is performed according to the following steps: step S1: arranging multiple acceleration sensors on an elevator door machine test platform to collect vibration signals of the acceleration sensors at multiple scales; step S2: constructing a VMD-MCNN model, wherein the VMD-MCNN model includes an input layer, a VMD decomposition module, an MCNN network training layer, and an output layer; the vibration signal is input to the input layer; the VMD decomposition module is used for VMD of the vibration signal; the MCNN network training layer is used for network iterative training; and the output layer is used to output a fault classification. The present invention combines VMD and MCNN for data processing and calculation to achieve efficient and accurate gearbox fault detection.

Description

Technical Field

[0001] The present invention relates to the field of gear box fault detection, and in particular to a vibration fault diagnosis method for an elevator door machine system. Background Art

[0002] With the acceleration of modern urbanization and the prevalence of high-rise buildings, elevators, as a vital means of vertical transportation, have become an indispensable part of people's daily lives. Elevator door operator systems, crucial for elevator safety and efficiency, have a direct impact on passenger experience and operational reliability. Failures can lead to abnormal elevator speeds, even passenger trapping accidents, harsh noises during door opening and closing, and collisions caused by the lack of a buffered door.

[0003] In this context, improving the diagnostic performance of elevator door machine system faults is of great significance to the normal operation and maintenance of elevators. Traditional fault diagnosis usually relies on expert systems, but expert systems need to be led by domain experts and design relevant judgment rules. They are often only applicable to a small number of elevator door machine faults. They are poorly portable and inflexible, which is not conducive to the popularization of the method and also affects the accuracy of the diagnostic results. In addition, due to the complex structure and operating conditions of elevator door machine systems, they are greatly affected by environmental factors, which makes it difficult to extract vibration signal features. So far, there are few reports on the diagnosis of elevator door machine system faults or failures. Therefore, it is necessary to strengthen the diagnosis research of common elevator door machine faults.

[0004] In recent years, fault diagnosis has progressed towards intelligent and automated approaches. Deep learning, leveraging its powerful feature extraction capabilities and efficient data analysis, uses multi-layer neural networks to extract and learn deep features from vibration signals end-to-end, enabling fault diagnosis. This overcomes the previous over-reliance on expert prior knowledge for key feature extraction. Currently, methods such as convolutional neural networks and recurrent neural networks have been widely applied to diagnose bearing faults and high-end manufacturing applications such as aircraft engines, axial piston pumps, and planetary gearboxes, achieving significant success. However, for elevator door machine systems, relevant fault datasets are scarce, and research on door machine fault classification and diagnosis is still in its infancy. Therefore, accurately identifying fault types in elevator door machine systems, reducing maintenance personnel workload, lowering elevator operation and maintenance costs, and promoting high-level intelligent operation and maintenance for elevator systems has become a pressing issue. Summary of the Invention

[0005] The present invention aims to provide a vibration fault diagnosis method for an elevator door machine system. The present invention combines VMD and MCNN for data processing and calculation to achieve efficient and accurate gearbox fault detection.

[0006] The technical solution of the present invention is a method for diagnosing vibration faults of an elevator door machine system, which is performed according to the following steps:

[0007] Step S1: multiple acceleration sensors are set on the elevator door machine test platform to collect vibration signals of the acceleration sensors at multiple scales;

[0008] Step S2: Constructing a VMD-MCNN model, the VMD-MCNN model includes an input layer, a VMD decomposition module, an MCNN network training layer, and an output layer; the vibration signal is input to the input layer; the VMD decomposition module is used for VMD of the vibration signal; the MCNN network training layer is used for network iterative training; and the output layer is used to output fault classification;

[0009] Step S3: performing VMD decomposition on the vibration signal and reconstructing the decomposed vibration signal based on the fault frequency of the elevator door machine, and normalizing the reconstructed vibration signal;

[0010] The reconstruction process is carried out in the following steps:

[0011] Step S3.1: Calculate the input meshing frequency f of the elevator door machine driven gear pair c and the variable frequency motor input frequency f m ;

[0012] Step S3.2: Search for the maximum center frequency close to f c The VMD decomposes the signal components and records them as Z1, searching for the maximum center frequency close to f m The VMD decomposition signal component is recorded as Z2; the remaining VMD decomposition signal components are accumulated and summed to obtain Z3;

[0013] Step S3.3: Use Z1, Z2 and Z3 as reconstructed signal components for normalization processing;

[0014] Step S4: Input the vibration signal from step S3 as training sample data into the VMD-MCNN model for network iterative training, calculate the difference between the actual sample category and the predicted result, and optimize and adjust the network parameters until the VMD-MCNN model reaches the convergence standard;

[0015] Step S5: Use the trained VMD-MCNN model to perform elevator door machine system fault diagnosis.

[0016] In the above-mentioned elevator door machine system vibration fault diagnosis method, the VMD decomposition algorithm in step S2 is:

[0017]

[0018] Where S represents the original signal; u krepresents the kth modal component; t represents time; u k (t) represents the mode function; ω k represents the center frequency of the kth modal component; k represents the number of modal decompositions; K represents the maximum number of modal decompositions; represents the partial derivative with respect to t; δ(t) represents the impulse function; represents the estimated center frequency of the mode.

[0019] In the above-mentioned elevator door machine system vibration fault diagnosis method, the normalization processing expression in step S2 is:

[0020]

[0021] Where a norm represents the normalized data, a min Indicates the corresponding minimum characteristic value, a max Indicates the maximum value of the corresponding feature.

[0022] In the aforementioned elevator door machine system vibration fault diagnosis method, the MCNN network training layer includes multiple multi-scale feature extraction layers, parallel feature fusion layers, high-dimensional feature extraction layers and flattening layers arranged in sequence; the multi-scale feature extraction layer includes multiple stacked convolution layers and multiple pooling layers for feature extraction and optimization of vibration signals; the VMD decomposition module is connected before the convolution layer of the multi-scale feature extraction layer and connected to the input layer; the parallel feature fusion layer includes a series layer and a fully connected layer for feature fusion; the high-dimensional feature extraction layer includes multiple convolution layers and an overlay layer for further processing features; the flattening layer is used to flatten three-dimensional features into one-dimensional vectors; the output layer is a fully connected layer and is connected to the flattening layer.

[0023] In the aforementioned elevator door machine system vibration fault diagnosis method, the convolutional layers all use the ReLU activation function and add regularization.

[0024] In the aforementioned elevator door machine system vibration fault diagnosis method, a random deactivation layer is provided between the convolution layer and the pooling layer of the multi-scale feature extraction layer and between the flattening layer and the output layer for optimizing features.

[0025] In the aforementioned elevator door machine system vibration fault diagnosis method, the output layer calculates the recognition probability of each fault category through the SoftMax startup function, and uses the category label with the highest probability as the diagnosis result.

[0026] Compared with the existing technology, the present invention uses VMD to decompose signals. VMD is suitable for non-stationary sequences, and the decomposition obtains subsequences containing multiple different frequency scales and relatively stable. It does not need to rely on prior knowledge of the fault cycle and can accurately decompose mechanical fault information, especially in complex faults. In the vibration signal, a specific frequency range may show more significant features; the use of MCNN can comprehensively extract rich and complementary feature information, which can effectively improve the accuracy of fault diagnosis. By capturing low-level to high-level features of data at multiple levels, the learning and generalization capabilities of the model are greatly improved; the VMD-MCNN model constructed by combining the two accurately captures the fault features in the signal, thereby having high classification accuracy and performance in various working condition tests. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 Flowchart of the present invention;

[0028] Figure 2 This is a flowchart of the reconstruction of signal components based on VMD decomposition of fault frequency characteristics of the present invention;

[0029] Figure 3 This is a diagram showing the arrangement of measurement points of the acceleration sensor according to an embodiment of the present invention;

[0030] Figure 4 The VMD decomposition diagram and corresponding spectrum diagram of the embodiment based on the normal condition of the fault frequency characteristics;

[0031] Figure 5 This is the VMD signal reconstruction result based on the normal condition of the fault frequency characteristic in the embodiment;

[0032] Figure 6 This is a diagram of the VMD-MCNN model training process of the present invention;

[0033] Figure 7 This is a graph showing the accuracy change during the VMD-MCNN training process.

[0034] Figure 8 This is a graph showing the change in loss function during VMD-MCNN training.

[0035] Figure 9 A visual confusion matrix diagram of VMD-MCNN fault classification in the embodiment;

[0036] Figure 10 2 is a comparison chart of fault classification accuracy of different algorithms in the embodiment. DETAILED DESCRIPTION

[0037] The present invention will be further described below with reference to the accompanying drawings and examples, but they are not intended to limit the present invention.

[0038] Embodiment: A method for diagnosing vibration faults in an elevator door machine system, as shown in the attached Figure 1 As shown, follow the steps below:

[0039] Step S1: As shown in the attached Figure 3 As shown, six acceleration sensors are set on the elevator door machine test platform to collect vibration signals of acceleration sensors of multiple scales. In this embodiment, the elevator door machine test platform is mainly composed of a door machine, an upper sill, a lower sill, a floor door panel, a pulley group and a car door panel, wherein the door machine is the main control and power mechanism, including a variable frequency asynchronous motor, a synchronous belt, a synchronous pulley, an encoder, a belt splint and a door limit switch. Acceleration sensors with a sampling frequency of 50,000 Hz are installed at a total of six positions on the elevator door machine test platform, including the upper and lower axial, horizontal and radial positions, to detect vibration signals. In order to simulate the working conditions of the elevator during peak power consumption and unstable power grid, three working conditions are set for each fault. The configuration of each working condition is shown in Table 1 below.

[0040]

[0041] Table 1

[0042] According to the statistics and investigation of elevator fault maintenance records, it is found that the main faults include abnormal door opening and closing speed, abnormal working sound, abnormal synchronous belt transmission, incorrect installation position and the location of the fault, which can be specifically divided into mechanical component failure, transmission device failure and external blockage failure. In this embodiment, faults are injected into different parts of the elevator door machine to simulate elevator operation failures under actual working conditions: 1. By adjusting the door machine controller, it is simulated that the door opening and closing speed becomes faster and the door closing is accompanied by a loud collision sound; 2. By inserting a thin copper wire into the driven synchronous wheel bearing ball, a bearing jam failure is simulated; 3. By sticking transparent tape on the upper sill, it is simulated that the upper sill is stained with dust and other stains; 4. By pressing the roller against the middle edge of the floor door panel, it is simulated that the floor door and door frame, fireproof edge friction failure; 5. By sticking two layers of tape on the tooth surface of the synchronous belt, it is simulated that the tooth surface wear failure of the synchronous belt is simulated; 6. The rubber on both sides of the normal slider is cut off to simulate the serious wear failure of the car door slider, as shown in Table 2 below

[0043]

[0044] Table 2

[0045] 600 groups of six fault signals and normal signals of the elevator door machine system were collected respectively, and 3 working conditions were set, totaling 12,600 groups; the normal condition of the elevator door machine, the fault of opening and closing the door too quickly, the fault of the synchronous pulley bearing blocking, the fault of the upper sill being stained with dust and stains, the fault of the friction between the floor door and the door cover, the fault of severe wear of the synchronous belt tooth surface, and the fault of severe wear of the car door slider were divided into seven categories: a, b, c, d, e, f, and g. Each type of fault had 600 samples, and the training set and test set were divided into 5:1, as shown in Table 3 below.

[0046]

[0047] Table 3

[0048] Step S2: Construct a VMD-MCNN model, which includes an input layer, three parallel multi-scale feature extraction layers, a parallel feature fusion layer, a high-dimensional feature extraction layer, a flattening layer, and an output layer. The vibration signal is input to the input layer. The multi-scale feature extraction layer includes two stacked convolution layers and two pooling layers, and the convolution layer is connected to a VMD decomposition module for feature extraction and optimization of the vibration signal. The parallel feature fusion layer includes a series layer and a fully connected layer, and the extracted multi-scale features are operated in parallel in the direction of the sensor channel. The high-dimensional feature extraction layer includes two equal-width convolution layers and a stacked pooling layer. Layer, used to further process features; the flattening layer is used to flatten the three-dimensional features into a one-dimensional vector; the output layer is a fully connected layer, which uses the SoftMax activation function to calculate the recognition probability of each fault category, and takes the category label with the highest probability as the diagnosis result of the model; each convolution layer uses the ReLU activation function, and to prevent overfitting, L2 regularization and Dropout optimization layers are added to each convolution layer. In addition, random inactivation layers are added between the convolution layer and the pooling layer, and between the flattening layer and the output layer to further optimize the model. The optimizer selects the Adam optimizer, and the learning rate is set to 0.0001, as shown in the attached figure. Figure 6 As shown, the cross entropy loss function is selected as the loss function of the model. The ultimate goal of model training is to minimize the loss function of the validation set. The specific model results are shown in Table 4 below.

[0049]

[0050]

[0051] Table 4

[0052] Step S3: The original signal contains a large amount of environmental noise. In addition, the faults set in this embodiment are all early-stage faults, which are weak faults. Due to noise interference, the fault characteristics are partially hidden, making extraction more difficult. Therefore, the original vibration signals collected from the six measurement points of the elevator door machine are subjected to variational mode decomposition (VMD). The length of the original signal is N, denoted as S. VMD becomes the solution to a constrained problem, which is expressed in the following formula:

[0053]

[0054] Where S represents the original signal; u k represents the kth modal component; t represents time; u k (t) represents the mode function; ωk represents the center frequency of the kth modal component; k represents the number of modal decompositions; K represents the maximum number of modal decompositions; represents the partial derivative with respect to t; δ(t) represents the impulse function; represents the estimated center frequency of the mode.

[0055] As attached Figure 2 As shown in FIG, the decomposed signal components (IMF components) are reconstructed according to the motor rotation frequency and the synchronous belt gear meshing frequency characteristics, and the weak fault characteristics are enhanced; first, the penalty factor α in the VMD decomposition is empirically taken as 1.5-2.0 times the sample length, and α=2000; secondly, since the decomposition layer number K has a great effect on the VMD modal decomposition, in order to avoid under-decomposition or over-decomposition of the vibration signal, this embodiment uses an enumeration method to determine the final optimal decomposition number based on the center frequency of each IMF corresponding to different K values, as shown in Table 5 below

[0056] K <![CDATA[u1]]> <![CDATA[u2]]> <![CDATA[u3]]> <![CDATA[u4]]> <![CDATA[u5]]> <![CDATA[u6]]> <![CDATA[u n <!-- 7 -->]]> 3 2 301 301 — — — — 6 2 11 301 301 301 396 — 10 2 4 119 239 303 301 479 13 2 4 14 239 301 301 479 14 2 4 14 119 239 301 479 15 2 4 14 119 239 303 479

[0057] Table 5

[0058] It can be seen from Table 5 that when K = 3 and K = 6, the center frequencies of the decomposed IMF components are concentrated at 11 and 301 Hz, and there is an under-decomposition phenomenon. When K = 10, there are still very few low-frequency components in the IMF. Finally, considering the low-frequency decomposition of K = 13, K = 14, and K = 15, the final decomposition layer K is determined to be 14; according to the attached Figure 4 As shown in the figure, the center frequency of IMF3 is close to the output frequency of the variable frequency asynchronous motor (15 Hz), and the center frequency of IMF4 is close to the meshing frequency of the driven synchronous gear pair (144 Hz). After signal reconstruction, the multidimensional signal containing only three IMFs is obtained, as shown in the attached figure. Figure 5 shown.

[0059] After that, Min-max normalization is performed, and the expression is:

[0060]

[0061] Where a norm represents the normalized data, a min Indicates the corresponding minimum characteristic value, a max Indicates the maximum value of the corresponding feature.

[0062] Step S4: Input the vibration signal from step S3 as training sample data into the VMD-MCNN model for network iterative training, calculate the difference between the actual sample category and the predicted result, and optimize and adjust the network parameters until the VMD-MCNN model reaches the convergence standard;

[0063] Taking all 6 sensor combinations as input, a VMD-MCNN model with 6 parallel multi-scale feature extraction layers is established and learning, training, verification and testing are carried out. Accuracy and loss values ​​are often used to measure the diagnostic effect of the VMD-MCNN model. The accuracy compares the model's prediction results with the actual fault category of the sample in the form of percentage. The higher the accuracy, the better the model effect. The loss value represents the sum of errors calculated by a given loss function. The lower the loss value, the better the model effect. Figure 7 and attached Figure 8 As shown in the figure, the model training is overall stable without drastic fluctuations, and it has converged around the 8th iteration. The final accuracy of the training set and validation set reaches 99.29%, and the loss value is reduced to 0.001.

[0064] The untrained test sample data is input into the trained VMD-MCNN model to output the diagnosis results, which are then compared with the actual fault categories to calculate the accuracy of the VMD-MCNN model. The corresponding confusion matrix is ​​as follows: Figure 9 As shown in the figure, each column of the confusion matrix represents the category predicted by the model, and each row represents the actual fault category of the test set. The value in each column represents the number of times this sample is predicted to be of that category. The value on the main diagonal is the number of correct classifications for each category. Only a few samples are classified outside the main diagonal, which proves that the proposed model can correctly classify the six fault types and normal conditions of the elevator door machine with a high accuracy. A total of seven types.

[0065] To further evaluate the performance of the proposed VMD signal decomposition and MCNN diagnostic method, it was compared with support vector machine (SVM), K-nearest neighbor method (KNN), random forest (RF), one-dimensional convolutional network (1D-CNN), and empirical mode decomposition-multiscale convolutional network (EMD-MCNN). To avoid random errors, all results are the mean of 10 repeated experiments. The results are shown in the attached figure. Figure 10 shown.

[0066] Attachment Figure 10 The left subgraph is a traditional machine learning model without signal processing, and the right subgraph is a neural network model, and the latter two are combined with signal processing methods. In terms of accuracy, when the original signal is used as input, the accuracy of the one-dimensional convolutional network is higher than that of other single models, but lower than that of the latter two hybrid methods after signal processing. This shows that the neural network's ability to extract and analyze features is stronger than traditional machine learning methods. The VMD-MCNN proposed in this paper has an average accuracy of 97.72%, which is much higher than that of a single one-dimensional convolutional network. Although its robustness is slightly lower than that of the EMD-MCNN, which is also a hybrid method, its overall fault recognition rate is better than the former.

[0067] Step S5: Use the trained VMD-MCNN model to perform helicopter accessory gearbox fault diagnosis.

[0068] In summary, the VMD-MCNN model combines the unique advantages of variational mode decomposition (VMD) and multi-channel convolutional neural network (MCNN). VMD decomposes the original signal into several sub-modes with central frequencies by constructing an iterative model. First, it can self-determine the number of modal decompositions and second, reduce the non-stationarity of time series with high complexity and strong nonlinearity. Therefore, VMD is suitable for non-stationary sequences, decomposing them into relatively stationary sub-sequences containing multiple different frequency scales. Moreover, VMD does not rely on prior knowledge of fault cycles and can accurately decompose mechanical fault information, especially in vibration signals of complex faults, where specific frequency ranges may show more significant features. MCNN can comprehensively extract rich and complementary feature information, which can effectively improve the accuracy of fault diagnosis. By capturing low-level to high-level features of data at multiple levels, it greatly enhances the learning and generalization capabilities of the model. At the same time, the efficient parallel processing capabilities demonstrated by MCNN when processing large-scale data sets make it of great application value in real-time fault monitoring and diagnosis.

Claims

1. A method for diagnosing vibration faults in an elevator door machine system, characterized by: Follow these steps: Step S1: multiple acceleration sensors are set on the elevator door machine test platform to collect vibration signals of the acceleration sensors at multiple scales; Step S2: Constructing a VMD-MCNN model, the VMD-MCNN model includes an input layer, a VMD decomposition module, an MCNN network training layer, and an output layer; the vibration signal is input to the input layer; the VMD decomposition module is used for VMD of the vibration signal; the MCNN network training layer is used for network iterative training; and the output layer is used to output fault classification; Step S3: performing VMD decomposition on the vibration signal and reconstructing the decomposed vibration signal based on the fault frequency of the elevator door machine, and normalizing the reconstructed vibration signal; The reconstruction process is carried out in the following steps: Step S3.1: Calculate the input meshing frequency f of the elevator door machine driven gear pair c and the variable frequency motor input frequency f m ; Step S3.2: Search for the maximum center frequency close to f c The VMD decomposes the signal components and records them as Z1, searching for the maximum center frequency close to f m The VMD decomposition signal component is recorded as Z2; the remaining VMD decomposition signal components are accumulated and summed to obtain Z3; Step S3.3: Use Z1, Z2 and Z3 as reconstructed signal components for normalization processing; Step S4: Input the vibration signal from step S3 as training sample data into the VMD-MCNN model for network iterative training, calculate the difference between the actual sample category and the predicted result, and optimize and adjust the network parameters until the VMD-MCNN model reaches the convergence standard; Step S5: Use the trained VMD-MCNN model to perform elevator door machine system fault diagnosis.

2. The elevator door machine system vibration fault diagnosis method according to claim 1, characterized in that: The VMD decomposition algorithm in step S2 is: Where S represents the original signal; u k represents the kth modal component; t represents time; u k (t) represents the mode function; ω k represents the center frequency of the kth modal component; k represents the number of modal decompositions; K represents the maximum number of modal decompositions; represents the partial derivative with respect to t; δ(t) represents the impulse function; represents the estimated center frequency of the mode.

3. The elevator door machine system vibration fault diagnosis method according to claim 1, characterized in that: The normalization processing expression in step S2 is: Where a norm represents the normalized data; a min Indicates the minimum value of the corresponding feature; a max Indicates the maximum value of the corresponding feature.

4. The elevator door system vibration fault diagnosis method according to claim 1, characterized in that: The MCNN network training layer includes multiple multi-scale feature extraction layers, parallel feature fusion layers, high-dimensional feature extraction layers, and flattening layers arranged in sequence; the multi-scale feature extraction layer includes multiple stacked convolution layers and multiple pooling layers for feature extraction and optimization of vibration signals; the VMD decomposition module is connected before the convolution layer of the multi-scale feature extraction layer and connected to the input layer; the parallel feature fusion layer includes a series layer and a fully connected layer for feature fusion; the high-dimensional feature extraction layer includes multiple convolution layers and stacking layers for further feature processing; The flattening layer is used to flatten the three-dimensional features into a one-dimensional vector; the output layer is a fully connected layer and is connected to the flattening layer.

5. The elevator door machine system vibration fault diagnosis method according to claim 4, characterized in that: The convolutional layers all use the ReLU activation function and add regularization.

6. The elevator door system vibration fault diagnosis method according to claim 4, characterized in that: A random inactivation layer is provided between the convolution layer and the pooling layer of the multi-scale feature extraction layer, and between the flattening layer and the output layer, for optimizing features.

7. The elevator door system vibration fault diagnosis method according to claim 4, characterized in that: The output layer calculates the recognition probability of each fault category through the SoftMax startup function, and takes the category label with the highest probability as the diagnosis result.

Images