Joint assembly equipment health condition evaluation method based on MC-IWDCNN-LSTM

Abstract

A joint assembly equipment health condition evaluation method based on MC-IWDCNN-LSTM is characterized by comprising the following steps: (1) collecting and processing noisy vibration signals related to the joint assembly equipment health condition according to joint assembly equipment health key indexes; (2) an MC-IWDCNN-LSTM feature extraction model is established; (3) performing performance evaluation and parameter adjustment on the feature extraction model by adopting a test set to realize feature extraction of noisy vibration signals of the joint assembly equipment; and (4) in combination with a fuzzy mathematics theory, based on an output quantity and a membership function of the MC-IWDCNN-LSTM model, calculating to obtain a health degree score. Decision support can be provided for a maintenance team, the health condition of noisy vibration signals in the joint assembly posture adjusting mechanism is evaluated, and the safety and stability of equipment operation are ensured.

Description

Technical Field

[0001] The present invention relates to a health management technology for joint assembly equipment, in particular to a health status assessment technology for joint assembly equipment, specifically to a health status assessment method for joint assembly equipment based on MC-IWDCNN-LSTM. Background Art

[0002] With the rapid development of industrial intelligence, joint assembly is the core link in production and manufacturing. The accurate evaluation of the health status of noisy vibration signals in joint assembly equipment has become an important research direction to ensure production efficiency and work safety. However, the existing technology still has the following deficiencies in the field of health status assessment: under the actual working conditions of the joint assembly equipment, the environment where the noisy vibration signal source is located is very complex, and many factors will cause errors in the original vibration signal of the bearing, thereby affecting the bearing fault diagnosis effect. It mainly includes mechanical noise and structural resonance caused by contact, friction and vibration between the moving parts of the mechanism, which usually have the characteristics of high frequency and randomness; vibration, sound and electromagnetic radiation from the surrounding environment, which have the characteristics of a wide frequency range; electrical noise from motors, sensors, electronic equipment in the posture adjustment mechanism and amplitude fluctuations caused by pressure pulses of the hydraulic system; and measurement noise introduced by the acquisition sensor itself, including nonlinear characteristics, noise of the amplifier circuit and sampling rate limitations.

[0003] In order to obtain a health status assessment model of noisy vibration signals with strong anti-noise interference ability and good robustness, the present invention proposes a multi-channel processing neural network model (MC-IWDCNN-LSTM) based on the attention mechanism. The model inputs noisy data simulating actual working conditions, uses dual channels to extract features from the noise data, and each channel learns different features respectively. The attention mechanism is used to adaptively assign weights to the two processing channels to highlight effective features and suppress invalid features. Finally, through feature combination and parameter sharing at the merging layer, more comprehensive, rich, and deep feature information is obtained, which increases the complexity and flexibility of the model, thereby reducing the impact of noise interference in actual working conditions and improving the accuracy of model prediction. Summary of the invention

[0004] The purpose of the present invention is to address the problem that in the current field of health status assessment, under the actual operating conditions of the docking assembly equipment, the environment where the noisy vibration signal source is located is very complex, and many factors will cause errors in the original vibration signal of the bearing, thereby affecting the bearing fault diagnosis effect. A method for health status assessment of docking assembly equipment based on MC-IWDCNN-LSTM is invented. By establishing an MC-IWDCNN-LSTM model, the accuracy of health assessment is improved, and decision support is provided for equipment health management and maintenance decisions.

[0005] The technical solution of the present invention is:

[0006] A method for evaluating the health status of joint assembly equipment based on MC-IWDCNN-LSTM is characterized in that the method collects and processes noisy vibration signals related to the health status of joint assembly equipment around the key health indicators of joint assembly equipment, establishes an MC-IWDCNN-LSTM feature extraction model, uses a test set to evaluate the performance and adjust the parameters of the feature extraction model, realizes the feature extraction of noisy vibration signals of joint assembly equipment, and finally combines fuzzy mathematics theory to calculate the health score based on the output eigenvalue and membership function of the MC-IWDCNN-LSTM model. Through this method, the health status of joint assembly equipment can be evaluated, decision support can be provided for the maintenance team, and the safety and stability of equipment operation can be ensured.

[0007] Step 1: Focusing on the key health indicators of the bonding and assembly equipment, collect and process the noisy vibration signals related to the health status of the bonding and assembly equipment;

[0008] Step 2: Establish the MC-IWDCNN-LSTM feature extraction model;

[0009] Step 3: Use the test set to evaluate the performance and adjust the parameters of the feature extraction model to achieve feature extraction of noisy vibration signals of the bonding assembly equipment;

[0010] Step 4: Combining fuzzy mathematics theory, the health score is calculated based on the output eigenvalues ​​and membership functions of the MC-IWDCNN-LSTM model.

[0011] Furthermore, in step 1, the joint assembly equipment is composed of a chassis motion system, an active sensing system and a posture adjustment mechanism system to realize the intelligent operation of the "storage, transportation, posture adjustment, docking, separation and replacement" process. The chassis motion system mainly includes wheel sets, shock absorbers of the suspension system, motor drive units, and rolling bearings connected between wheel sets; the active sensing system includes alignment measurement modules, laser radars, RFC sensing devices, magnetic navigation sensing devices and various sensors; the core components of the posture adjustment mechanism system include wheel sets, reducers, controllers, motor drive units, rolling bearings connected between wheel sets and their controller signals. The health status of these components needs to be monitored and diagnosed based on specific health indicator parameters, many of which are noisy vibration signals. For example, the health status of the motor drive unit can be judged by monitoring its vibration signal, which may contain high-frequency components with noise. Especially when the motor temperature is too high, the accompanying abnormal vibration signal may indicate motor overload or heat dissipation failure;

[0012] The vibration signals of the joint assembly equipment are collected. Since these data may have abnormal values, missing data, inconsistent dimensions, etc., they need to be preprocessed. For example, a sudden increase or decrease in current within an unreasonable range in a short period of time may be caused by external interference; at the same time, data may be missing due to communication problems. Data preprocessing includes cleaning abnormal values, filling missing data, and normalization to ensure data consistency and integrity;

[0013] Furthermore, in the step (2), a multi-channel processing neural network model (MC-IWDCNN-LSTM) based on an attention mechanism is adopted. The model inputs a vibration signal with noise under actual working conditions, uses dual channels for feature extraction, and each channel learns different features respectively. Finally, the attention mechanism is used to adaptively assign weights to the two processing channels, highlight effective features and suppress invalid features, and obtain more comprehensive, rich, and in-depth feature information through feature combination and parameter sharing in the merging layer, thereby increasing the complexity and flexibility of the model, thereby reducing the impact of noise interference in actual working conditions and improving the accuracy of model prediction.

[0014] The MC-IWDCNN-LSTM model consists of the first-layer feature extraction module, the attention module, the multi-channel processing module, and the fully connected module. When designing the model structure, some reference was made to the widely used two-dimensional convolutional neural networks, such as VGGnet, ResNet, and Google's InceptionV4. These three network structures all use stacked 3×3 convolution kernels multiple times, which can achieve the effect of suppressing overfitting while obtaining a larger receptive field and using only a few parameters. However, for one-dimensional data, since the two-layer 3×1 convolution structure pays the price of 6 weights but only obtains a 5×1 receptive field, the original advantage in two dimensions becomes a disadvantage in one dimension, so it cannot be completely copied according to the structure of the two-dimensional convolutional network. At the same time, the idea of ​​the original WDCNN model was also partially referred to. A large convolution kernel was set in the first layer to achieve the effect of short-time Fourier transform to extract short-time features; except for the first layer, the convolution kernel size of the remaining layers was set to 3×1, which is conducive to deepening the network and suppressing overfitting.

[0015] The specific neural network structure of the MC-IWDCNN-LSTM model is as follows: the large convolution kernel selected in the first-layer feature extraction module is conducive to extracting temporal features and enhancing the robustness of the model to noise interference; the attention module adaptively adjusts the attention weight and the extracted features to improve the training rate and classification effect of the model; the multi-channel processing module uses the IWDCNN model and the LSTM model for feature extraction, combining the advantages of the two; the four-layer small convolution kernel of the left channel is conducive to deepening the network, extracting deep-level features, and suppressing overfitting; the LSTM of the right channel can efficiently utilize temporal features and improve prediction accuracy; finally, the features are merged through the merging layer, and overfitting is suppressed by dropout. After full connection, softmax is used for classification, and the number n of outputs is the number of health status categories to be classified; in the model of the present invention, a batch normalization layer (BN) is also used after each convolution activation layer to reduce the transfer of internal covariates, improve the generalization ability of the model, and enhance the training speed and efficiency of the neural network. The number of nodes in the fully connected Dense layer is usually set to 5 to 20 times the number of softmax, which is better.

[0016] Furthermore, in step 3, a test set is used to evaluate the prediction accuracy and training speed of the model, and the model parameters are reasonably adjusted to improve the model performance and realize feature extraction of noisy vibration signals.

[0017] Furthermore, in step 4, the health score is calculated based on the output eigenvalues ​​and membership functions of the MC-IWDCNN-LSTM model in combination with fuzzy mathematics theory.

[0018] The beneficial effects of the present invention are:

[0019] 1. Enhanced robustness: First, input a one-dimensional vibration signal with noise, and extract features through the first-layer wide convolution kernel to enhance the model's robustness to noise signals.

[0020] 2. Improve accuracy: Use the attention mechanism to adaptively assign weights to the dual channels, highlight important features, and suppress useless features; the dual channels use the improved WDCNN model and LSTM model respectively, combining the advantages of the two to prevent overfitting and improve resolution efficiency and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 Flowchart for equipment health assessment for bonding assembly.

[0022] Figure 2 Parameters related to equipment health for the main functional components of the bonding assembly equipment.

[0023] Figure 3 This is the neural network structure diagram of the MC-IWDCNN-LSTM model.

[0024] Figure 4 This is the model training curve graph. DETAILED DESCRIPTION

[0025] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0026] like Figure 1-Figure 4 shown.

[0027] A method for assessing the health status of joint assembly equipment based on MC-IWDCNN-LSTM is proposed. It collects and processes noisy vibration signals related to the health status of joint assembly equipment based on key health indicators of joint assembly equipment, establishes an MC-IWDCNN-LSTM feature extraction model, uses a test set to evaluate the performance and adjust the parameters of the feature extraction model, and realizes the feature extraction of noisy vibration signals of joint assembly equipment. Finally, combined with fuzzy mathematics theory, the health score is calculated based on the output eigenvalues ​​and membership functions of the MC-IWDCNN-LSTM model. Through this method, the health status of joint assembly equipment can be evaluated, decision support can be provided for the maintenance team, and the safety and stability of equipment operation can be ensured.

[0028] Figure 1 The health status evaluation flow chart for the joint assembly equipment includes the following steps:

[0029] Step 1: Focusing on the key health indicators of the bonding and assembly equipment, collect and process the noisy vibration signals related to the health status of the bonding and assembly equipment;

[0030] Figure 2 Parameters related to equipment health for the main functional components of the bonding assembly equipment.

[0031] The joint assembly equipment is composed of a chassis motion system, an active sensing system, and a posture adjustment mechanism system to realize the intelligent operation of the "storage, transportation, posture adjustment, docking, separation, and replacement" process. The chassis motion system mainly includes wheel sets, shock absorbers of the suspension system, motor drive units, and rolling bearings connected between wheel sets; the active sensing system includes alignment measurement modules, laser radars, RFC sensing devices, magnetic navigation sensing devices, and various sensors; the core components of the posture adjustment mechanism system include wheel sets, reducers, controllers, motor drive units, rolling bearings connected between wheel sets, and their controller signals. The health status of these components needs to be monitored and diagnosed based on specific health indicator parameters, many of which are noisy vibration signals. For example, the health status of the motor drive unit can be judged by monitoring its vibration signal, which may contain high-frequency components with noise. Especially when the motor temperature is too high, the accompanying abnormal vibration signal may indicate motor overload or heat dissipation failure;

[0032] The vibration signals of the joint assembly equipment are collected. Since these data may have abnormal values, missing data, inconsistent dimensions, etc., they need to be preprocessed. For example, a sudden increase or decrease in current within an unreasonable range in a short period of time may be caused by external interference; at the same time, data may be missing due to communication problems. Data preprocessing includes cleaning abnormal values, filling missing data, and normalization to ensure data consistency and integrity;

[0033] Step 2: Establish the MC-IWDCNN-LSTM feature extraction model;

[0034] The specific neural network structure of the MC-IWDCNN-LSTM model is as follows Figure 3 As shown, the large convolution kernel selected in the first-layer feature extraction module is conducive to extracting temporal features and enhancing the robustness of the model to noise interference; the attention module adaptively adjusts the attention weight and the extracted features to improve the training rate and classification effect of the model; the multi-channel processing module uses the IWDCNN model and the LSTM model for feature extraction, combining the advantages of the two; the four-layer small convolution kernel of the left channel is conducive to deepening the network, extracting deep-level features, and suppressing overfitting; the LSTM of the right channel can efficiently utilize temporal features and improve prediction accuracy; finally, the features are merged through the merging layer, and overfitting is suppressed by dropout. After full connection, softmax is used for classification, and the number n of outputs is the number of health status categories to be classified; in the model of the present invention, a batch normalization layer (BN) is also used after each convolution activation layer to reduce the transfer of internal covariates, improve the generalization ability of the model, and enhance the training speed and efficiency of the neural network. The number of nodes in the fully connected Dense layer is usually set to 5 to 20 times the number of softmax, which is better.

[0035] Convolutional Attention Module (CBAM) is an attention module for convolutional networks, which is divided into channel attention and spatial attention modules.

[0036] The channel attention mechanism first aggregates the spatial feature information using the maximum pooling layer and the average pooling layer, and then merges the data of the average pooling layer and the maximum pooling layer through a multi-layer perceptron (MLP) with a hidden layer to obtain the channel attention weight. The specific formula is as follows.

[0037]

[0038] in, are the average and maximum pooling features respectively; W0 and W1 are weight matrices, and W0 is obtained by the ReLU activation function; AvgPool and MaxPool are average pooling and maximum pooling; MLP is a shared multilayer perceptron; σ is a sigmoid activation function.

[0039] The spatial attention mechanism is a complement to the channel attention mechanism, focusing on the features that need to be focused on. The features of the input channel attention module are applied to the average and maximum pooling operations, the two results are combined, and the spatial attention weight M is obtained through convolution and sigmoid activation function. S , the specific formula is as follows.

[0040]

[0041] in, is the generated average pooling feature and maximum pooling feature; σ is the sigmoid activation function; p n is a convolution kernel of size n.

[0042] The multi-channel processing module is divided into a left-channel IWDCNN model and a right-channel LSTM model, which respectively use CNN to extract features and LSTM to perform temporal analysis on features.

[0043] The convolution layer has the characteristics of local connection and weight sharing. It uses a one-dimensional convolution kernel to perform convolution operations on local areas and output the extracted features, reducing the number of features. The convolution calculation formula is as follows.

[0044]

[0045] in, It is the feature subgraph of the j-th convolution output of the L-th layer; * indicates the convolution operation; represents an additional bias matrix; f(·) is a nonlinear activation function that performs nonlinear activation for each feature map value.

[0046] The LSTM model first uses two layers of small convolution kernels to extract features, and then inputs the extracted features into the LSTM unit to learn the dependencies between temporal features, thereby speeding up training time and obtaining more accurate prediction results.

[0047] Step 3: Use the test set to evaluate the performance and adjust the parameters of the feature extraction model to achieve feature extraction of noisy vibration signals of the bonding assembly equipment;

[0048] The optimizer used in the model is the adam algorithm that can adaptively adjust the learning rate, with the parameter set to 0.001; the cross-entropy loss function is used as the loss function; the early stop and optimal model saving mechanism is introduced during the training process, and the training is automatically stopped when it is detected that the accuracy has not improved for multiple consecutive rounds. According to the results of multiple model training, the classification accuracy of the test set is 98.1% ± 0.5%. The accuracy and loss value curves of the training set and validation set during the model training process are shown in the figure below. Figure 4shown.

[0049] Step 4: Combining fuzzy mathematics theory, the health score is calculated based on the output eigenvalues ​​and membership functions of the MC-IWDCNN-LSTM model.

[0050] Combined with fuzzy mathematics theory, the health score is used to quantify the degree of deviation between the operating parameters of the health index and the rated standard interval. The interval range is [0,1]. The larger the value, the better the health status of the equipment, and the more suitable it is for long-term joint assembly tasks. According to the health score results of equipment maintenance personnel and experts for different health status levels, combined with fuzzy mathematics theory, the equipment health score is linked to the membership degree, and the health level is further quantitatively analyzed to make the fuzzy quantity precise. The membership function indicates the degree to which the current health indicator belongs to a certain health state. The closer it is to 0, the lower the degree to which the indicator belongs to the health state; conversely, the higher the degree to which it belongs to the health state.

[0051] According to the actual situation of the joint assembly equipment, the ridge distribution function is selected as the membership function of the present invention to represent the fuzzy relationship between the health states. The characteristics of the membership function are that the transition area is gentle and the main value interval is wide, and the effect in terms of symmetry and stability is also good, as shown below.

[0052]

[0053] The Zadeh representation of the health score HD in fuzzy theory is shown below.

[0054] HD=0.8×m1+0.6×m2+0.4×m3+0.2×m4

[0055] m1~m4 represent the degree of membership of four health states, namely excellent, good, fair, and observation, and the health scores are 0.8, 0.6, 0.4, and 0.2.

[0056] The present invention does not involve parts that are the same as the prior art or can be implemented by using the prior art.

Claims

1. A method for assessing the health status of joint assembly equipment based on MC-IWDCNN-LSTM, characterized in that: Focusing on the key health indicators of joining and assembly equipment, the noisy vibration signals related to the health status of joining and assembly equipment are collected and processed, and the MC-IWDCNN-LSTM feature extraction model is established. The test set is used to evaluate the performance and adjust the parameters of the feature extraction model to achieve the feature extraction of noisy vibration signals of joining and assembly equipment. Finally, based on the output eigenvalues ​​and membership functions of the MC-IWDCNN-LSTM model, the health score is calculated by combining fuzzy mathematics theory, thereby evaluating the health status of the joining and assembly equipment, providing decision support for the maintenance team, and ensuring the safety and stability of equipment operation.

2. The method according to claim 1, characterized in that It includes the following steps: Step 1: Focusing on the key health indicators of the bonding and assembly equipment, collect and process the noisy vibration signals related to the health status of the bonding and assembly equipment; Step 2: Establish the MC-IWDCNN-LSTM feature extraction model; Step 3: Use the test set to evaluate the performance and adjust the parameters of the feature extraction model to achieve feature extraction of noisy vibration signals of the bonding assembly equipment; Step 4: Combining fuzzy mathematics theory, the health score is calculated based on the output eigenvalues ​​and membership functions of the MC-IWDCNN-LSTM model.

3. The method according to claim 2, characterized in that In step 1: The joint assembly equipment is composed of a chassis motion system, an active sensing system and a posture adjustment mechanism system to realize the intelligent operation of the "storage, transportation, posture adjustment, docking, separation and replacement" process; the chassis motion system mainly includes wheel sets, shock absorbers of the suspension system, motor drive units and rolling bearings connected between wheel sets; the active sensing system includes an alignment measurement module, a laser radar, an RFC sensing device, a magnetic navigation sensing device and a sensor; the core components of the posture adjustment mechanism system include wheel sets, reducers, controllers, motor drive units, rolling bearings connected between wheel sets and their controller signals; the health status of these components needs to be monitored and diagnosed based on specific health indicator parameters, and the health indicator parameters are noisy vibration signals, including motor The health status of the drive unit can be determined by monitoring its vibration signal. The vibration signal contains high-frequency components with noise. Especially when the motor temperature is too high, the accompanying abnormal vibration signal indicates motor overload or heat dissipation failure. The noisy vibration signal of the joint assembly equipment is collected. Under actual working conditions, the vibration signal is inevitably affected by mechanical noise, environmental interference and sensor errors. Since these data have abnormal values, missing data and inconsistent dimensions, they need to be preprocessed. A sudden increase or decrease in current to an unreasonable range in a short period of time is caused by external interference. At the same time, data loss may also occur due to communication problems. Data preprocessing includes cleaning abnormal values, filling missing data and normalization to ensure data consistency and integrity.

4. The method according to claim 2, characterized in that In step 2: A multi-channel processing neural network model MC-IWDCNN-LSTM based on the attention mechanism is adopted. The model inputs the vibration signal with noise under actual working conditions, uses dual channels for feature extraction, and each channel learns different features respectively. Finally, the attention mechanism is used to adaptively assign weights to the two processing channels, highlighting effective features and suppressing invalid features. In the merging layer, feature combination and parameter sharing are used to obtain more comprehensive, rich, and deep feature information, increase the complexity and flexibility of the model, thereby reducing the impact of noise interference in actual working conditions and improving the accuracy of model prediction. The MC-IWDCNN-LSTM model consists of a first-layer feature extraction module, an attention module, a multi-channel processing module, and a fully connected module. When designing the model structure, one part uses a two-dimensional convolutional neural network: VGGnet, ResNet, and InceptionV4. These three network structures use stacked 3×3 convolution kernels multiple times. While obtaining a larger receptive field, only a few parameters are used to achieve the effect of suppressing overfitting. However, for one-dimensional data, since the two-layer 3×1 convolution structure pays the price of 6 weights but only obtains a 5×1 receptive field, the original advantage in two dimensions becomes a disadvantage in one dimension, so it cannot be completely copied according to the structure of the two-dimensional convolutional network. At the same time, the other part adopts the idea of ​​the DCNN model, and sets a large convolution kernel in the first layer to achieve the effect of short-time Fourier transform to extract short-time features. Except for the first layer, the convolution kernel size of the remaining layers is set to 3×1, which is conducive to deepening the network and suppressing overfitting. The specific neural network structure of the MC-IWDCNN-LSTM model is as follows: the large convolution kernel selected in the first-layer feature extraction module is conducive to extracting temporal features and enhancing the robustness of the model to noise interference; the attention module adaptively adjusts the attention weight and the extracted features to improve the training rate and classification effect of the model; the multi-channel processing module uses the IWDCNN model and the LSTM model for feature extraction respectively; the four-layer small convolution kernel of the left channel is conducive to deepening the network, extracting deep-level features, and suppressing overfitting; the LSTM of the right channel can efficiently utilize temporal features and improve prediction accuracy; finally, the features are merged through the merging layer, and dropout is used to suppress overfitting. After full connection, softmax is used for classification, and the number of outputs n is the number of health status categories to be classified; a batch normalization layer (BN) is also used after each convolution activation layer in the model to reduce the transfer of internal covariates, improve the generalization ability of the model, and enhance the training speed and efficiency of the neural network; the number of nodes in the fully connected Dense layer is usually set to 5 to 20 times the number of softmax.

5. The method according to claim 4, characterized in that The channel attention mechanism first uses the maximum pooling layer and the average pooling layer to aggregate spatial feature information, and then merges the data of the average pooling layer and the maximum pooling layer through a multi-layer perceptron (MLP) with a hidden layer to obtain the channel attention weight. The specific formula is as follows: in, are average and maximum pooling features respectively; W0 and W1 are weight matrices, and W0 is obtained by the ReLU activation function; AvgPool and MaxPool are average pooling and maximum pooling; MLP is a shared multi-layer perceptron; σ is a sigmoid activation function; The spatial attention mechanism is a supplement to the channel attention, focusing on the features that need to be focused on; the features of the channel attention module are input, the average and maximum pooling operations are applied to them, the two results are merged, and the spatial attention weight M is obtained through convolution and sigmoid activation function. S , the specific formula is as follows: in, is the generated average pooling feature and maximum pooling feature; σ is the sigmoid activation function; p n is a convolution kernel of size n.

6. The method according to claim 4, characterized in that The multi-channel processing module is divided into a left-channel IWDCNN model and a right-channel LSTM model; the two respectively use CNN to extract features and use LSTM to perform temporal analysis on the features; The convolution layer has the characteristics of local connection and weight sharing. It uses a one-dimensional convolution kernel to perform convolution operations on the local area and outputs the extracted features, reducing the number of features. The convolution calculation formula is as follows. in, It is the feature subgraph of the j-th convolution output of the L-th layer; * indicates the convolution operation; represents an additional bias matrix; f(·) is a nonlinear activation function that performs nonlinear activation for each feature map value; The LSTM model first uses two layers of small convolution kernels to extract features, and then inputs the extracted features into the LSTM unit to learn the dependencies between the temporal features, thereby speeding up the training time and obtaining more accurate prediction results.

7. The method according to claim 2, characterized in that: In step 3: a test set is used to evaluate the prediction accuracy and training speed of the model, and the model parameters are reasonably adjusted to improve the model performance and realize feature extraction of noisy vibration signals.

8. The method according to claim 2, characterized in that: In step 4: the output value obtained by inputting the noisy vibration signal into the model is combined with fuzzy mathematics theory to obtain a health score.

Images