CN114841196A

CN114841196A - Mechanical equipment intelligent fault detection method and system based on supervised learning

Info

Publication number: CN114841196A
Application number: CN202210356687.7A
Authority: CN
Inventors: 邵海霞; 孟真; 邵世聪
Original assignee: Qiteng Technology Beijing Co ltd
Current assignee: Qiteng Technology Beijing Co ltd
Priority date: 2022-04-06
Filing date: 2022-04-06
Publication date: 2022-08-02

Abstract

The embodiment of the invention discloses a mechanical equipment intelligent fault detection method and system based on supervised learning, wherein a mechanical equipment intelligent fault diagnosis model based on supervised learning combines a stacked two-dimensional and one-dimensional convolutional neural network CNN, a residual error duration memory network LSTM and an LSTM based on supervised learning, can detect abnormal data generated by mechanical equipment, can use a CNN model to extract spatial characteristics of data, and detects abnormal states in a mechanical vibration data set based on time sequences under various environments by combining the residual error LSTM and the LSTM, so as to evaluate the health state of the equipment; original signal data based on a time sequence is converted into a Mel frequency spectrum image, and image-based analysis is performed, so that better performance is obtained in a fault diagnosis system applying data enhancement; the data set is added to solve the problem of data imbalance, so that the accuracy of the model is improved.

Description

Mechanical equipment intelligent fault detection method and system based on supervised learning

Technical Field

The embodiment of the invention relates to the technical field of fault diagnosis of mechanical equipment, in particular to an intelligent fault detection method and system of mechanical equipment based on supervised learning.

Background

As mechanical equipment technology advances, the complexity of the industrial environment and the uncertainty associated with productivity also increases. If aged mechanical equipment is ignored and the damage caused by it is not adequately repaired, the equipment will be defective and productivity will be reduced. In addition, the damaged device may cause damage to other devices. Therefore, it is necessary to develop advanced technologies to improve the device security. Machine equipment components such as valves, fans, slide rails, rolling bearings and the like are core components of machine equipment systems and play a key role in determining the performance and the service life of the machine systems. These mechanical components may fail for a variety of reasons, the most serious of which is defects in the bearings of the electromagnetic drive system.

Various condition monitoring methods have been used to detect faults in components of industrial machines. Conventional methods include distance-based K-nearest neighbor algorithms; a local anomaly factor for detecting a local anomaly; and a connectivity-based outlier factor that is an improved version of the local anomaly factor and uses radius to detect anomalies. The conventional method based on distance and density has a disadvantage in that anomaly detection requires a considerable time as the number of data points increases. A deep learning method capable of overcoming this limitation has appeared in recent years and exhibits higher performance than the conventional method. In addition, the development of the internet of things industry has facilitated large-scale data collection. Therefore, the importance of abnormality detection based on supervised learning is increasing.

According to the recent trend, the anomaly detection method based on machine learning consists of three main steps: data preprocessing, extracting important characteristics of normal and abnormal data from a raw signal based on a time sequence; selecting a deep learning based model, i.e. selecting a model for a fault diagnosis system; anomaly detection-a deep learning based model uses extracted features and detects anomalies through a learning process. Currently, machine learning has been used to develop a method for time-series data-based diagnosis, aiming at abnormality detection of a malfunction of a component of a mechanical apparatus using data of the mechanical apparatus. This establishes deep learning for anomaly detection of mechanical devices using both time and frequency domains. However, the current technology is limited in that data from the components themselves are not considered. Furthermore, the fault diagnosis model based on intelligent data is not tested over a range of loads, durations, and noise; therefore, the data used is not sufficient to fully test the fault diagnosis model.

Disclosure of Invention

Therefore, the embodiment of the invention provides an intelligent fault detection method and system for mechanical equipment based on supervised learning, so as to adapt to abnormal detection for various environments, enhance the generalization capability of an abnormal detection algorithm and construct a fault diagnosis system with robustness and universality.

In order to achieve the above object, the embodiments of the present invention provide the following technical solutions:

according to a first aspect of the embodiments of the present invention, a method for detecting an intelligent fault of a mechanical device based on supervised learning is provided, where the method includes:

converting the time series based original data signal into Mel frequency spectrogram data;

and inputting the Mel frequency spectrogram data into a constructed mechanical equipment intelligent fault diagnosis model based on supervised learning to detect abnormal states in a mechanical vibration data set based on a time sequence, wherein the model comprises a two-dimensional convolutional neural network CNN, a one-dimensional convolutional neural network CNN, a residual error long-time memory network LSTM and a long-time memory network LSTM which are stacked.

Further, an output of the two-dimensional convolutional neural network CNN is connected to an input of the one-dimensional convolutional neural network CNN, an output of the one-dimensional convolutional neural network CNN is connected to an input of the residual error long-short term memory network LSTM, and an output of the residual error long-short term memory network LSTM is connected to an input of the long-short term memory network LSTM.

Further, the output of the long-time and short-time memory network LSTM is connected with a full connection layer, and the full connection layer is connected with a classification output layer.

Further, the method further comprises:

The combined convolutional neural network is used for extracting normal and abnormal signal characteristics based on the Mel frequency spectrum image data, the two-dimensional convolutional neural network CNN is used for maintaining the spatial information of the Mel frequency spectrum image through multiple convolution filters, extracting and learning the characteristics of adjacent low-frequency images, and then performing characteristic collection and enhancement on the extracted characteristics through a pooling layer and using the characteristics as the input of the one-dimensional convolutional neural network.

Further, the method further comprises:

features extracted by the convolutional neural network are used as inputs to the remaining LSTM layer, enabling detection of abnormal states in the time-series based mechanical vibration dataset by the combination of residual LSTM and LSTM.

Further, the method further comprises:

the classification output layer converts it into a probability distribution between 0 and 1 corresponding to classes using a softmax activation function, enabling detection of anomalies by separating normal and anomalous data into two or more classes.

According to a second aspect of the embodiments of the present invention, a mechanical equipment intelligent fault detection system based on supervised learning is provided, the system includes:

the data preprocessing module is used for converting the original data signals based on the time sequence into Mel frequency spectrogram data;

And the anomaly detection module is used for inputting the Mel frequency spectrogram data into a constructed mechanical equipment intelligent fault diagnosis model based on supervised learning to detect an abnormal state in a time sequence-based mechanical vibration data set, and the model comprises a two-dimensional convolutional neural network CNN, a one-dimensional convolutional neural network CNN, a residual error long-time memory network LSTM and a long-time memory network LSTM which are stacked.

According to a third aspect of embodiments of the present invention, a computer storage medium containing one or more program instructions for executing the method of any one of the above items by a supervised learning based mechanical equipment intelligent fault detection system is provided.

The embodiment of the invention has the following advantages:

the mechanical equipment intelligent fault diagnosis model based on supervised learning combines a two-dimensional convolutional neural network CNN, a residual error long-time memory network LSTM and a LSTM based on supervised learning, can detect abnormal data generated by mechanical equipment, can extract spatial characteristics of data by using the CNN model, and detects abnormal states in a mechanical vibration data set based on time sequences in various environments by combining the residual error LSTM and the LSTM, so as to evaluate the health state of the equipment; original signal data based on a time sequence is converted into a Mel frequency spectrum image, and image-based analysis is performed, so that better performance is obtained in a fault diagnosis system applying data enhancement; the data set is added to solve the problem of data imbalance, so that the accuracy of the model is improved.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below. It should be apparent that the drawings in the following description are merely exemplary, and that other embodiments can be derived from the drawings provided by those of ordinary skill in the art without inventive effort.

Fig. 1 is a schematic flowchart of a mechanical equipment intelligent fault detection method based on supervised learning according to embodiment 1 of the present invention;

fig. 2 is a schematic structural diagram of a one-dimensional convolutional neural network in a mechanical device intelligent fault detection method based on supervised learning according to embodiment 1 of the present invention;

fig. 3 is a schematic structural diagram of a two-dimensional convolutional neural network in a mechanical device intelligent fault detection method based on supervised learning according to embodiment 1 of the present invention;

fig. 4 is a schematic structural diagram of an LSTM network in the intelligent fault detection method for mechanical equipment based on supervised learning according to embodiment 1 of the present invention;

fig. 5 is a schematic structural diagram of a residual error LSTM network in the intelligent fault detection method for mechanical equipment based on supervised learning according to embodiment 1 of the present invention;

Fig. 6 is a schematic structural diagram of a mechanical equipment intelligent fault diagnosis model based on supervised learning in a mechanical equipment intelligent fault detection method based on supervised learning according to embodiment 1 of the present invention.

Detailed Description

The present invention is described in terms of particular embodiments, other advantages and features of the invention will become apparent to those skilled in the art from the following disclosure, and it is to be understood that the described embodiments are merely exemplary of the invention and that it is not intended to limit the invention to the particular embodiments disclosed. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

Example 1

As shown in fig. 1, the present embodiment provides a method for detecting an intelligent fault of a mechanical device based on supervised learning, where the method includes:

s100, converting the original data signal based on the time sequence into Mel frequency spectrogram data.

This embodiment proposes the use of a mel-frequency spectrogram to pre-process the raw signal in order to convert the time-series signal in the form of raw output data from the machine into an appropriate input to the convolutional network. In the preprocessing stage, a Mel frequency spectrum is adopted to extract the time-frequency domain signal characteristics generated by mechanical equipment, and the frequency spectrum characteristics of abnormal signals doped in the original signals are identified.

S200, inputting the Mel frequency spectrogram data into a constructed mechanical equipment intelligent fault diagnosis model based on supervised learning to detect abnormal states in a time sequence-based mechanical vibration data set, wherein the model comprises a two-dimensional convolutional neural network CNN, a one-dimensional convolutional neural network CNN, a residual error long-time memory network LSTM and a long-time memory network LSTM which are stacked.

The present embodiment adopts a multiple filter of a 2D convolutional neural network to maintain spatial information of a mel-frequency spectrum image and extract and learn features of adjacent low-frequency images. The extracted features are then feature collected and enhanced through the pooling layer, serving as input to the 1D convolutional neural network. And finally, detecting abnormal states in the mechanical vibration data set based on the time sequence through the combination of residual LSTM and LSTM, and further evaluating the health state of the equipment.

The main contributions of the mechanical equipment intelligent fault detection method based on supervised learning of the embodiment are as follows:

1) anomalies are determined using a mel-frequency spectrogram image used to extract normal and anomalous features of a time-series based mechanical device vibration data set. 2) The data set is added to solve the problem of data imbalance, so that the accuracy of the model is improved. 3) In addition, the robustness of various models to noise was tested in consideration of the signal-to-noise ratio (SNR) prevailing in an actual industrial field. 4) Supervised learning based models were validated using time series based vibration data of different lengths. In addition, the data set is divided according to loads, and various environments are considered through transition learning from one load to another. 5) Through 2D and 1D convolution neural networks, a characteristic diagram representing the spectral image spatial relationship of the bearing and the industrial machine component can be extracted. In addition, temporal profiles of time series analysis can be extracted by LSTM. 6) Supervised learning-based intelligent diagnostic models developed using various loads, noise, and time series-based machine data sets exhibit superior generalization capability that effectively addresses the task of diagnosing machine faults.

1) Convolutional Neural Network (CNN)

The CNN is used for analyzing the original signals based on time series, converting the original signals into images of Mel frequency spectrums, and extracting normal and abnormal signal characteristics. CNN has been widely used in the fields of image and signal processing, and the range of further using CNN in these fields has been determined. Recently, a combination of one-dimensional and two-dimensional CNNs has been used for image classification and pattern recognition, and such a combination shows good possibilities in processing structured data. It has been demonstrated that the combined use of 1D and 2D CNNs provides higher performance than the use of only one CNN. CNN is a deep learning architecture created by mimicking the human optic nerve. Unlike multilayer perceptrons, it can learn unique features of an image regardless of the position and orientation of the object. 1D and 2D CNN are implemented using the same principles. CNN is performed by combining convolution with neural networks. Convolution operation keeps the space consistency of the image, greatly reduces the calculation amount compared with a completely connected neural network, and shows good performance in image classification. The convolution stages shown in fig. 2 and 3 represent a learnable set of convolution filters, including the pooling task. CNN uses such convolution filters instead of pixels to solve the problems encountered by multi-layered perceptrons. The convolution filter extracts features of the input data (corners and curves of the image) by calculating weights and applying activation functions while moving through the window at intervals from the top left to the bottom right of the image. If the portion passed by the operation filter coincides with the feature extracted by the convolution filter, a high output can be obtained, thereby increasing the likelihood of achieving good image classification. The main features extracted by learning the image using such an operation filter are input to a pooling layer, which reduces the image size by reducing the dimensionality. Pixels in a specific area in an image are grouped and reduced to a representative value. This pooling layer reduces the amount of computations and prevents overfitting. When the convolution operation starts from the left side of each layer, the network learns by extracting more features.

2) RNN variant network Long-short memory network (LSTM)

A Recurrent Neural Network (RNN) is a deep learning based model that processes inputs and outputs in sequence units. The RNN performs learning by remembering and using information about past events. In theory, RNNs can process sequence data well by considering the sequence, but as the sequence time range increases, RNNs are difficult to learn successfully. FIG. 4 shows the architecture of the LSTM long-short memory neural network. LSTM can overcome RNN learning limitations by introducing two memories and three gates. LSTM has the ability to learn from long input sequences. Memory cells using LSTM may solve the gradient depletion problem and multiple LSTM layers may be stacked. Three gates select the amount of information contained in the LSTM: 1) forgetting to record the door: it determines the amount of past information contained in the LSTM. The value passing through the sigmoid function is the output value of the forgetting gate. The output of the sigmoid function ranges between 0 and 1. A value of 0 indicates forgetting information of a previous state, and a value of 1 indicates memorizing information of a previous state. 2) An input gate: it determines the amount of new information stored. The sigmoid is used to determine which values are used for updating, and the second is a tanh layer used to generate new candidate values to be added to obtain the candidate values. 3) An output gate: firstly, obtaining an initial output through a sigmoid layer, then using tanh to scale the value to-1 to 1, and then multiplying the value by the output obtained by sigmoid pair by pair to obtain the output of the model.

3) Residual long and short memory network LSTM

Fig. 5 shows the residual long and short memory network LSTM. As described above, LSTM provides greater accuracy than RNN because the former has better learning capabilities. However, even if the number of layers of the LSTM increases, only a certain number of layers are running, and the network becomes too slow to learn. Thus, as with RNNs, deep LSTM networks may cause gradient depletion problems. To solve this problem, a remaining LSTM layer is used in the set, which simulates the difference between the middle layer and the next layer. The residual LSTM has an advantage in that accuracy and calculation speed in the learning process can be improved by using residual concatenation. In addition, it helps to solve the problem of gradient loss during back-propagation computation and can greatly improve gradient flow. In addition, it can learn feature mapping from the spatial feature relationship of time series-based melbourne image data acquired by CNN. In the last layer, the temporal feature map between the channels is extracted from the output values of the remaining LSTM layers by the LSTM layer and the fully connected layer, which has the temporal sequence features of the mechanical device elements. These features are then provided to the output classification layer, which converts them into a probability distribution between 0 and 1 corresponding to the class using the softmax activation function. Thus, anomalies can be detected by separating normal and anomalous data into two or more categories.

4) Mechanical equipment intelligent fault diagnosis model based on supervised learning

The invention provides a new model, which combines two-dimensional and one-dimensional CNN with residual LSTM and LSTM to detect the abnormality in the vibration data set of the mechanical equipment based on time sequence. Fig. 6 shows a constructed supervised learning-based intelligent diagnosis model of mechanical equipment. In the input layer of the model, the spatial information of the spectrogram image is maintained by using a multiple filter of a two-dimensional convolution layer, and the low-frequency image of the head part of the feature software is extracted and learned. The extracted features then pass through the pooling layer where they are collected and enhanced and reused for data entry into the one-dimensional convolutional layer. The one-dimensional convolutional layer improves the operation efficiency by reducing the size of previously extracted important information about the mechanical device. The natural values extracted by the convolutional layer are used as input values for the residual LSTM layer. And finally, detecting abnormal states in the time-series-based mechanical vibration data set under various environments through the combination of the residual LSTM and the LSTM.

According to the intelligent fault detection method for the mechanical equipment based on supervised learning, the problem of data imbalance in the bearing and industrial mechanical data based on time series is solved through data expansion. Further, by considering various types of noise in an actual industrial environment, a failure can be detected early. In addition, it facilitates the development of data-based intelligent fault diagnosis systems that improve the productivity of the above-described equipment by using mel-frequency spectrograms containing various features obtained from the raw signals. The proposed supervised learning based ensemble model combining CNN and LSTM not only provides the most advanced classification performance, but also solves many problems encountered in existing intelligent data-based fault diagnosis techniques, such as time-series based data and load variations and robustness to noise. In various practical environments, the performance of the model is superior to that of the existing fault diagnosis method based on supervised learning. The model can extract the spatial features of a Mel frequency spectrogram image with obvious noise features through two-dimensional and one-dimensional CNN. In addition, the characteristics of the vibration data set based on the time sequence are extracted through the residual LSTM and LSTM layers, so that normal and abnormal data can be effectively detected, and the health state evaluation of mechanical equipment is realized.

Example 2

Corresponding to the above embodiment 1, this embodiment proposes an intelligent fault detection system for mechanical equipment based on supervised learning, where the system includes:

The functions executed by each component in the supervised learning based intelligent fault detection system for mechanical equipment provided by the embodiment of the present invention are described in detail in embodiment 1 above, and therefore, redundant description is not repeated here.

Example 3

In accordance with the above embodiments, the present embodiment proposes a computer storage medium containing one or more program instructions for executing the method of embodiment 1 by a supervised learning based intelligent fault detection system for mechanical equipment.

Although the invention has been described in detail above with reference to a general description and specific examples, it will be apparent to one skilled in the art that modifications or improvements may be made thereto based on the invention. Accordingly, such modifications and improvements are intended to be within the scope of the invention as claimed.

Claims

1. A mechanical equipment intelligent fault detection method based on supervised learning is characterized by comprising the following steps:

2. The supervised learning based intelligent fault detection method for mechanical equipment according to claim 1, wherein an output of the two-dimensional convolutional neural network CNN is connected with an input of a one-dimensional convolutional neural network CNN, an output of the one-dimensional convolutional neural network CNN is connected with an input of a residual error long-short term memory network LSTM, and an output of the residual error long-short term memory network LSTM is connected with an input of the long-short term memory network LSTM.

3. The supervised learning-based intelligent fault detection method for mechanical equipment as recited in claim 2, wherein the output of the long-term memory network (LSTM) is connected with a full connection layer, and the full connection layer is connected with a classification output layer.

4. The supervised learning based intelligent fault detection method for mechanical equipment according to claim 3, further comprising:

5. The supervised learning based intelligent fault detection method for mechanical equipment according to claim 3, further comprising:

6. The supervised learning based intelligent fault detection method for mechanical equipment according to claim 3, further comprising:

7. A mechanical equipment intelligent fault detection system based on supervised learning, characterized in that the system comprises:

8. A computer storage medium comprising one or more program instructions embodied in the computer storage medium for execution by a supervised learning based mechanical equipment intelligence failure detection system to perform the method of any of claims 1-6.