Rotating machine real-time fault diagnosis method and system based on edge computing terminal

Abstract

The invention discloses a rotating machine real-time fault diagnosis method and system based on an edge computing terminal, and relates to the field of mechanical equipment health detection.The method comprises the following steps that an acceleration sensor is used for obtaining a vibration signal of mechanical equipment, and median filtering is used for conducting filtering processing on the collected vibration signal on the edge computing terminal; a frequency domain signal is obtained by using fast Fourier transform, a time domain feature is extracted by using Transform, a frequency domain feature is extracted by using CNN, and a calculation amount is reduced by using one-dimensional depth separable convolution; time domain and frequency domain features are efficiently fused by using a designed feature fusion module, and SoftMax is used for classified output; gPU training is used locally, and the trained model is deployed on an edge computing device to realize real-time fault diagnosis. According to the method, the collected vibration signals are processed in real time by using an efficient and lightweight model in an edge calculation mode, so that the fault diagnosis time can be effectively shortened, and the problem that mechanical faults are difficult to find in time in actual engineering is solved.

Description

Technical Field

[0001] The present invention relates to the field of rotating machinery health detection, and in particular to a rotating machinery real-time fault diagnosis method and system based on an edge computing terminal. Background Art

[0002] Rotating machinery is widely used in large-scale machinery such as automobiles, ships, and industrial production equipment. However, when a rotating machinery fault occurs, if it is not promptly addressed, it may cause equipment downtime, production interruptions, and even safety accidents. With the development of technology, rotating machinery fault diagnosis has evolved from manual inspections to intelligent fault diagnosis. However, diagnostic methods mainly rely on high-performance computers or cloud servers for centralized diagnosis, which wastes a lot of time in data transmission and diagnosis. Intelligent fault diagnosis based on edge computing terminals is characterized by its short time and high efficiency, which is crucial for fault detection.

[0003] In recent years, intelligent fault diagnosis technology based on deep learning has made significant progress. Early intelligent fault diagnosis methods primarily relied on CNN models. Due to the difficulty of extracting features from one-dimensional vibration signals, the signals were typically converted into two-dimensional images. This significantly increases the model's computational complexity. CNNs also primarily extract local features and fail to capture global dependencies within the data sequence. To overcome this limitation, fault diagnosis methods based on Transformers have gradually begun to be explored. Based on a full self-attention mechanism, Transformers can effectively capture long-range dependencies within time series signals when extracting features, which is crucial for feature extraction from vibration signals. Although existing deep learning-based fault diagnosis models demonstrate excellent accuracy, these methods are mostly designed for high-performance computing terminals and lack compatibility with resource-constrained edge computing terminals. This leads to the widespread use of inference models on local computers or cloud servers. This approach requires pre-collecting data and then uniformly feeding it into the model for fault diagnosis. This wastes significant time during device acquisition and data transmission, leading to delayed fault resolution, which can result in significant losses.

[0004] With the improvement of embedded chip performance, edge computing is increasingly being used in engineering. Edge computing involves migrating data processing and computing tasks from traditional data centers or cloud servers to locations closer to the data source. This approach typically involves performing processing on edge nodes in the network. Compared to traditional cloud computing, edge computing offers real-time and security advantages. Therefore, it is necessary to design an efficient and lightweight FT-CNN-Transformer algorithm model for fault diagnosis at the edge to address the aforementioned issues. Summary of the Invention

[0005] The purpose of this invention is to design a lightweight and efficient fault diagnosis model for real-time fault diagnosis on edge computing terminals. This model overcomes the computational complexity and the fact that existing fault diagnosis models can only be run on computers or servers, thereby reducing the time required for fault diagnosis. Therefore, combining these features, a real-time fault diagnosis method and system for rotating machinery based on edge computing terminals are proposed.

[0006] A real-time fault diagnosis method and system for rotating machinery based on edge computing terminals is specifically completed in the following steps:

[0007] (1) Install the acceleration sensor on the rotating machinery and collect the vibration signal of the equipment when the equipment is running;

[0008] (2) Inputting the vibration signal collected in step (1) into the edge computing terminal, removing the sudden abnormal noise in the signal by means of median filtering, and collecting the vibration signals under different fault conditions and the vibration signals when the equipment is operating normally according to the above method;

[0009] (3) Design a lightweight neural network model FT-CNN-Transformer, which includes a Transformer model for time domain feature extraction, a CNN model for frequency domain signal feature extraction, a fast Fourier transform module, a feature fusion module, and a classification output module;

[0010] (4) Dividing the vibration signal collected in step (2) into a training set and a test set, and sending them into the FT-CNN-Transformer model built in step (3) for training and testing;

[0011] (5) Build the Pytorch virtual environment required for model inference on the edge computing terminal, and deploy the trained FT-CNN-Transformer model to the edge computing terminal;

[0012] (6) Connect the data acquisition link to the edge computing terminal, use MobaXterm software to remotely connect to the edge computing terminal and run the program to achieve real-time fault diagnosis on the edge computing terminal with a diagnosis time of less than 100 milliseconds.

[0013] Compared with existing fault diagnosis methods, it has the following beneficial effects:

[0014] 1. Taking the one-dimensional signal of the sensor as input, it greatly reduces the amount of calculation and complexity, and improves the reasoning speed.

[0015] 2. A fast Fourier transform module embedded in the model was designed, enabling the model to extract time domain and frequency domain features simultaneously.

[0016] 3. Using the Transformer encoder to extract one-dimensional time domain signal features and using CNN to extract one-dimensional frequency domain signal features can simultaneously capture the long-term dependencies of time series and the local adjacent features of frequency domain signals, improving feature extraction capabilities.

[0017] 4. A module for fusing time domain and frequency domain features is designed to align multi-domain features, highlight important features, weaken irrelevant features, and efficiently fuse time domain and frequency domain features, thereby improving the generalization ability and accuracy of the model.

[0018] 5. The final designed model has only 9.962KB of parameters and 1.277MB of floating-point computation. It achieves real-time fault diagnosis with an inference time of 68 milliseconds on a Raspberry Pi 4B terminal and 76 milliseconds on a Feiteng Pi terminal, and can be deployed on most edge computing chips.

[0019] Principle of the present invention:

[0020] The proposed method and system for real-time fault diagnosis of rotating machinery based on an edge computing terminal achieves edge-based fault diagnosis primarily due to the model's minimal computational effort. Using an efficient Fast Fourier Transform (FFT) algorithm, the discrete time-domain signals acquired by a vibration sensor are converted into frequency-domain signals. A CNN model is then used to extract features from the frequency-domain signals, while a Transformer model is used to extract features from the time-domain signals. This allows the model to simultaneously consider both time-domain and frequency-domain features. The combination of the CNN and Transformer models captures both local features of the vibration signal and global dependencies, significantly reducing the difficulty of extracting features from one-dimensional signals. A feature fusion module for both time and frequency domains is designed to deeply fuse the extracted time-domain and frequency-domain features. Finally, a SoftMax function is used for classification and output of diagnostic results. The designed model, deployed on a Raspberry Pi 4B and a Feiteng Pi, achieved a diagnostic accuracy of 99.8% on the widely recognized CWRU open-source dataset. The inference time on the Raspberry Pi was only 68 milliseconds, and on the Feiteng Pi, it was only 76 milliseconds.

[0021] Figures in the specification

[0022] Figure 1 It is a structural flow chart of the edge terminal real-time fault diagnosis system.

[0023] Figure 2 This is a structural diagram of the FT-CNN-Transformer model of the present invention.

[0024] Figure 3 This is a schematic diagram of the one-dimensional depth-separable convolution principle.

[0025] Figure 4This is the schematic diagram of the feature fusion module.

[0026] Figure 5 This is a comparison chart of the model's accuracy on different devices.

[0027] Figure 6 This is a comparison chart of the model's test results on the Raspberry Pi 4B4G terminal. DETAILED DESCRIPTION

[0028] The present invention will be further described below with reference to the embodiments and drawings. The specific embodiments are only used to further illustrate the present invention and do not limit the scope of protection of the claims of this application.

[0029] Example 1: Bearing fault diagnosis based on edge computing terminal:

[0030] (1) The data on the drive end of the Case Western Reserve University (CWRU) bearing dataset with a sampling frequency of 12 kHz is divided into ten types, and these data are divided into training and test data with a sample number of 2048 sampling points. The fault types include normal bearing, bearing inner ring fault, bearing outer ring fault and rolling element fault. Each fault type has three fault defects.

[0031] (2) The fast Fourier transform module is used to convert the time domain signal of the acceleration sensor into a frequency domain signal. The fast Fourier transform module is placed at the beginning of the CNN model. Its specific implementation steps include: first, a discrete sequence with a length of 2048 sampling points is converted to the frequency domain through fast Fourier transform. Since the frequency domain sequence after the one-dimensional real number Fourier transform is symmetrical, in order to reduce unnecessary calculations, half of the non-repeated frequency domain data of the discrete sequence is taken, and then the discrete sequence is normalized to obtain the frequency domain signal for input into the CNN model. A represents the normalized signal, and X represents the frequency domain signal after the fast Fourier transform. The normalization calculation formula is as follows:

[0032] A=2|X| / N, N=2048

[0033] (3) The model structure of FT-CNN-Transformer is as follows Figure 2 As shown in Figure 2, the frequency domain data obtained by fast Fourier transform in step (2) is used as input in the CNN part. In the structural design of the CNN model, the first layer of convolution uses one-dimensional conventional convolution with a convolution kernel size of 5. From the first layer onwards, all layers use one-dimensional depthwise separable convolution instead of conventional convolution. The structure of one-dimensional depthwise separable convolution includes one-dimensional depthwise convolution, point-by-point convolution layer and activation function layer. The specific structure is as follows: Figure 3As shown. In the second layer of the CNN model, a depth-separable convolution with a convolution kernel size of 3 is used, in the third layer, a depth-separable convolution with a convolution kernel size of 5 is used, and in the fourth layer, a depth-separable convolution with a convolution kernel size of 3 is used. Then, the feature map output by the first layer of conventional convolution is concatenated with the feature map output by the fourth layer of convolution in the channel dimension through maximum pooling to retain more shallow features. A residual block consisting of two layers of depth-separable convolution with a convolution kernel size of 3 is then used for further feature extraction to obtain the output of the CNN branch. X is the input feature map, W is the convolution kernel, and Y is the output feature map. The one-dimensional convolution calculation formula is:

[0034]

[0035] (4) In the Transformer part, the time domain signal divided in step (1) is used as the input of the model. At the beginning of the Transformer model, a convolution layer is used to downsample the input feature map of length 2048. The first layer uses a conventional convolution with a convolution kernel size of 9, and the second layer uses a depth-wise separable convolution with a convolution kernel size of 7. After two layers of convolution downsampling and expanding the number of channels, a fully connected layer is used to positionally encode the feature map. The feature map is then input into the Transformer model composed of a 5-layer Transformer encoder to extract features. Finally, a layer of depth-wise separable convolution with a convolution kernel size of 3 is used to further fuse the features to obtain the output of the Transformer module.

[0036] (5) Use the feature fusion module to fuse the feature maps extracted in steps (3) and (4). The feature fusion module consists of a feature normalization layer, a feature concatenation layer, a global average pooling layer, a fully connected layer, and a weight function layer, such as Figure 4 As shown in the figure, the main purpose of the feature normalization layer is to align the time domain and frequency domain features, ensure that the feature values of the two dimensions are under the same standard, and then splice the time and frequency domain feature maps in the channel dimension, so that the time and frequency domain features have been preliminarily fused. In order to make the features more deeply integrated, retain and enhance the features with large contributions, and weaken irrelevant features, the global average pooling layer is used to compress the features to the channel dimension, and the fully connected layer is used to automatically learn the importance of the features. Then, the weight function is used to map the importance to the actual weight value. Finally, the feature map is multiplied by the weight value to achieve deep fusion, and the fused feature map is output. X is the output of the fully connected layer, W is the output weight, α and β are hyperparameters, and the weight mapping function is:

[0037] W(X FC )=max(0,min(1,α|X FC |+β))

[0038] (6) After the feature fusion is completed in step (5), the fused features are further extracted using a residual block consisting of two layers of depth-wise separable convolution with a kernel size of 3. The feature map after feature extraction is then adaptively global average pooled, and finally, after passing through a fully connected layer, the diagnosis result is output by the Soft Max function classification.

[0039] (7) The FT-CNN-Transformer model built in steps (2) to (6) is trained using the data from step (1). The training device uses an NVIDIA GTX 1650Ti GPU, the learning rate is 0.001, the batch size is 96, the optimizer is Adam, α is set to 1.2, β is set to 0.2, the loss function uses cross entropy loss, and the training is carried out for 300 rounds.

[0040] (7) Deploy the trained model in step (7) to the Raspberry Pi 4B, and put the test set divided by step (1) into the Raspberry Pi, and test the accuracy and inference time of the model on the Raspberry Pi. After testing, the average accuracy of the CWRU dataset is 99.86%. The accuracy comparison between the model proposed in this patent and the general benchmark model is as follows: Figure 5 As shown in the figure, the average inference time of the model is 68 milliseconds. Figure 6 shown.

[0041] Example 2: Real-time fault diagnosis of oil pumps based on edge computing terminals:

[0042] (1) The real-time fault diagnosis system structure based on edge computing terminal is as follows Figure 1 As shown, vibration signal data is first collected on an oil pump fault simulation test bench. Oil pump fault types are categorized into four types: normal oil pump, oil pump screw failure, pump shaft misalignment, and shock absorber failure. The laboratory fault simulation test bench consists of a drive system, electronic control system, hydraulic system, oil temperature control, and filtration system, simulating typical oil pump failure modes. CA-YD-188 accelerometers are installed at all four corners of the oil pump to comprehensively monitor the pump's vibration status. Both training and test data are collected from the sensor fixed at the 1 o'clock position (upper right corner of the oil pump) with a sampling frequency of 20kHz. An A / D converter circuit board is used to collect the accelerometer vibration signal. During the acquisition process, the signal is pre-processed using a median filter to eliminate outliers. After acquisition, the data is converted into data samples of length 2048 and collected under loads of 0 MPa, 0.5 MPa, and 1 MPa. x represents the input signal, y represents the filtered signal, and the window value W is 3. The median filter formula is:

[0043] y i =Median(x i-(W-1) / 2 ,...,xi+(W-1) / 2 )

[0044] (2) After data preparation is completed, the model is built according to steps (2) to (6) in Example 1, and then the model is trained using an NVIDIA GTX1650Ti GPU with a learning rate of 0.001, a batch size of 96, an Adam optimizer, α set to 1.2, β set to 0.2, and a cross entropy loss loss as the loss function. The model is trained for 50 rounds.

[0045] (3) The trained model was deployed on a Raspberry Pi 4B, which was then connected to the oil pump simulation test bench in step (1). Data was collected and tested in real time. The collection method was the same as in step (1). The sample points were selected using a sliding window mechanism when inputting the model. That is, the window size was 2048, the sliding step size was 1024, and the two adjacent groups of samples had a semi-overlap rate, which ensured data continuity. After real-time testing, the model accuracy was above 99%, and the average inference time of the model was 68 milliseconds.

[0046] Obviously, the above-described embodiments of the present invention are intended to simply and clearly illustrate the present invention and are not intended to limit the embodiments of the present invention. Those skilled in the art will appreciate that other variations can be made based on the above description. It is not necessary to enumerate all embodiments here. Therefore, any modifications made within the spirit and principles of the present invention are intended to be included within the scope of protection of the claims of the present invention.

Claims

1. A real-time fault diagnosis method and system for rotating machinery based on edge computing terminal, characterized in that The steps include: (1) Install the acceleration sensor on the rotating machinery and collect the vibration signal of the equipment when the equipment is running; (2) Inputting the vibration signal collected in step (1) into the edge computing terminal, removing the sudden abnormal noise in the signal by means of median filtering, and collecting the vibration signals under different fault conditions and the vibration signals when the equipment is operating normally according to the above method; (3) Design a lightweight neural network model FT-CNN-Transformer, which includes a Transformer model for time domain feature extraction, a CNN model for frequency domain signal feature extraction, a fast Fourier transform module, a feature fusion module, and a classification output module; (4) Dividing the vibration signal collected in step (2) into a training set and a test set, and sending them into the FT-CNN-Transformer model built in step (3) for training and testing; (5) Build the Pytorch virtual environment required for model inference on the edge computing terminal, and deploy the trained FT-CNN-Transformer model to the edge computing terminal; (6) Connect the data acquisition link to the edge computing terminal, use MobaXterm software to remotely connect to the edge computing terminal and run the program to achieve real-time fault diagnosis on the edge computing terminal with a diagnosis time of less than 200 milliseconds.

2. A method and system for real-time fault diagnosis of rotating machinery based on edge computing terminal according to claim 1, characterized in that The digital or analog accelerometer used in steps (1) to (2) has a sampling frequency of 4kHz to 20kHz and is fixed on the rotating machinery. The edge computing terminal refers to a terminal with a computing power of 0.4TOPS or above and can be installed with the Pytorch environment, including Raspberry Pi, Feiteng Pi, Orange Pi, etc.

3. A method and system for real-time fault diagnosis of rotating machinery based on edge computing terminal according to claim 1, characterized in that In step (3), a fast Fourier transform module embedded in the model is designed. Compared with the discrete Fourier algorithm, the fast Fourier transform can greatly reduce the amount of calculation and perform energy normalization on the converted frequency domain signal.

4. A method and system for real-time fault diagnosis of rotating machinery based on edge computing terminal according to claim 1, characterized in that In step (3), the encoder part of the Transformer model is used to extract time domain signal features, and the feature map is embedded in an automatically learnable positional encoding before feature extraction.

5. A method and system for real-time fault diagnosis of rotating machinery based on edge computing terminal according to claim 1, characterized in that In step (3), the residual structure and one-dimensional depth-wise separable convolution layer are used in the CNN model design. Depth-wise separable convolution is used instead of conventional convolution. Its convolution structure includes one-dimensional depth-wise convolution, batch normalization, Hard-Swish activation function, one-dimensional point-wise convolution, batch normalization, and Hard-Swish activation function.

6. A method and system for real-time fault diagnosis of rotating machinery based on edge computing terminal according to claim 1, characterized in that The feature fusion module in step (3) includes time domain and frequency domain feature map alignment, time domain and frequency domain feature map splicing, and feature weight calculation. The classification and output module includes an adaptive global average pooling layer, a fully connected layer, and a SoftMax classification layer.

7. A method and system for real-time fault diagnosis of rotating machinery based on edge computing terminal according to claim 1, characterized in that In step (4), the collected data is divided into a training set and a test set and trained using a GPU. The cross entropy loss function is used to solve the model classification loss. The learning rate is set to 0.01 to 0.0001 and the batch size is 16 to 256.

Images