A roots compressor fault diagnosis method and system

Abstract

The application provides a Roots compressor fault diagnosis method, comprising the following steps: acquiring vibration signals of a power shaft, current of a motor input end, compressor outlet pressure, compressor outlet flow and liquid temperature signals collected by multiple sensors on the Roots compressor to form a two-dimensional matrix; introducing the two-dimensional matrix into a trained noise reduction model to obtain a multi-scale signal feature matrix after noise reduction; introducing the multi-scale signal feature matrix into a trained Roots compressor fault classification model to diagnose a fault category; wherein the noise reduction model is a new auto-encoding model fusing multi-scale hollow convolution and attention, and the Roots compressor fault classification model is a residual network model with residual connection. By removing the noise of multi-sensor fusion data under complex working conditions, the application improves the fault diagnosis precision of the deep learning model and improves the accuracy of the model in complex working condition fault recognition.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of mechanical equipment fault diagnosis, and particularly relates to a Roots compressor fault diagnosis method and system. BACKGROUND

[0002] The Roots compressor is a kind of rotary machine, which relies on the relative rotation of two impellers to continuously compress and transport gas. Due to its stable working characteristics, high volumetric efficiency and strong adaptability, it is widely used in the evaporation link of electronic, petroleum, steel, building material, chemical, food, medical and papermaking industries.

[0003] The Roots compressor is prone to damage failure due to long-term operation under complex conditions such as high speed and high temperature. Common failure modes of the Roots compressor include damage of key vulnerable parts such as impellers, bearings, gears and transmission shafts, loosening of foundation bolts, motor winding open circuit, winding ground or inter-phase short circuit, rotor cage bar breakage and abnormal leakage. The Roots compressor has a complex structure and operates in a harsh environment, which makes it prone to failure during operation, causing economic losses and even safety accidents. Therefore, it is necessary to monitor and diagnose the working state and performance of the Roots compressor to timely detect the above-mentioned main failure modes and prevent accidents.

[0004] One of the key points of correct or incorrect identification of the Roots compressor fault is how to obtain the fault features of the Roots compressor from its complex signals. For example, Susanto et al. proposed a method combining fast Fourier transform and Hilbert transform, which can analyze the frequency spectrum to determine the processing condition. However, this method has a large artificial judgment component, which may misjudge the running state of the machine. Therefore, some researchers proposed machine learning methods such as MLP, SVM and decision tree to identify the fault of rotating machinery, which improved the accuracy of mechanical fault identification. However, these machine learning methods have some problems. Since the model structure is relatively simple, the self-learning ability is weak when dealing with complex data, and the model establishment and feature extraction need to be completed separately.

[0005] At present, deep learning is an "end-to-end model" with strong self-adaptive learning ability and strong advantages in mining potential features of data, and has been widely applied in fault diagnosis. For example, Lei Guo applied a deep learning model to the method of state monitoring of mechanical equipment big data; for another example, Lv Zhe et al. proposed a rotating machinery fault diagnosis method based on a generative adversarial network and a personalized mechanism model; for another example, Yang Shengkang et al. proposed a mechanical fault diagnosis method based on multi-source domain deep transfer learning.

[0006] However, the above deep learning method has the following problems, including: first, the training set features are relatively single, and the vibration acceleration signal is basically taken as the test data, which may not achieve the required recognition accuracy under complex working conditions; second, although deep learning can learn potential features during feature extraction, noise is also easily learned as signal features.

[0007] Therefore, how to effectively remove the noise of multi-sensor fusion data under complex working conditions and then perform fault recognition through a deep learning method is a problem to be solved. Therefore, it is necessary to use a new deep learning method in the fault diagnosis of Roots-type compressors, which can remove the noise of multi-sensor fusion data under complex working conditions and improve the accuracy of model fault recognition under complex working conditions. SUMMARY

[0008] The technical problem to be solved by the embodiments of the present application is to provide a Roots-type compressor fault diagnosis method and system, which improves the fault diagnosis precision of a deep learning model by removing the noise of multi-sensor fusion data under complex working conditions, and improves the accuracy of model fault recognition under complex working conditions.

[0009] To solve the above technical problem, the embodiments of the present application provide a Roots-type compressor fault diagnosis method, which comprises the following steps:

[0010] Obtaining vibration signals of a power shaft, current of a motor input end, compressor outlet pressure, compressor outlet flow and liquid temperature signals collected by multiple sensors on a Roots-type compressor to form a two-dimensional matrix;

[0011] The two-dimensional matrix is input into a trained noise reduction model for noise reduction to obtain a multi-scale signal feature matrix after noise reduction; wherein the noise reduction model is a new auto-encoding model fusing multi-scale hollow convolution and attention;

[0012] The obtained multi-scale signal feature matrix is input into a trained Roots-type compressor fault classification model to diagnose the fault category of the Roots-type compressor; wherein the Roots-type compressor fault classification model is a residual network model with residual connection.

[0013] The multi-scale hollow convolution attention auto-encoding model comprises an encoder and a decoder; wherein,

[0014] The encoder is first inputted with the two-dimensional matrix, and after adding a noise damage signal to the two-dimensional matrix, a linear layer and an activation function ReLU are used for feature extraction, and the extracted features are put into a multi-scale hollow convolution block to capture semantic features, and after the captured semantic features are decomposed into three matrices of query, key and value through a separation layer, they are put into a three-layer Transformer block to learn noise distribution features and signal features themselves;

[0015] The decoder extracts the noise distribution features and the signal features themselves to obtain a multi-scale signal feature matrix; wherein the decoder includes a linear layer for predicting the added noise and a linear layer for reorganizing the signal features; after calculating the gradients of the noise prediction loss and the reconstruction loss, the decoder uses the stochastic gradient descent method to update the weights.

[0016] Each layer of the Transformer block in the encoder is composed of two sub-layers; wherein,

[0017] One of the two sub-layers is the convergence of multi-head attention, and the other is a position-based feedforward network; wherein the calculation process between the two sub-layers is shown as And F i And Residual connection is used between the two sub-layers, and layer normalization is applied after the residual connection.

[0018] Wherein, the residual network takes the multi-scale signal feature matrix as the input of Resnet, and after feature extraction by a 7*7 convolution layer and a batch normalization layer, key features are extracted by a 3*3 max pooling layer, then four residual blocks are used to extract features from the feature matrix again, and finally, after the global average pooling layer aggregates the features of all channels, the extracted signal is classified by the full connection layer.

[0019] Wherein, the fault categories of the Roots compressor include impeller failure, bearing failure, gear failure, transmission shaft damage, foundation bolt loosening, motor winding open circuit, winding ground and phase-to-phase short circuit, rotor cage strip breakage and abnormal leakage.

[0020] The embodiment of the application also provides a Roots compressor fault diagnosis system, comprising:

[0021] A multi-signal fusion unit is used to acquire vibration signals of a power shaft, current of a motor input end, compressor outlet pressure, compressor outlet flow and liquid temperature signals collected by multiple sensors on the Roots compressor to form a two-dimensional matrix.

[0022] a fusion signal denoising unit, configured to import the two-dimensional matrix into a trained denoising model to perform denoising, so as to obtain a denoised multi-scale signal feature matrix; wherein the denoising model is a novel auto-encoding model fusing multi-scale hollow convolution and attention;

[0023] a fault category diagnosis unit, configured to import the obtained multi-scale signal feature matrix into a trained Roots compressor fault classification model to diagnose a fault category of the Roots compressor; wherein the Roots compressor fault classification model is a residual network model with residual connection.

[0024] The fault category of the Roots compressor includes impeller failure, bearing failure, gear failure, transmission shaft damage, foundation bolt loosening, motor winding open circuit, winding ground and phase-to-phase short circuit, rotor cage bar breakage and abnormal leakage.

[0025] The embodiment of the present application has the following beneficial effects:

[0026] The denoising model constructed based on the multi-scale hollow convolution attention auto-encoding model has strong denoising capability, and in combination with the Roots compressor fault classification model constructed based on the residual network, the problem of low accuracy of fault discrimination of the Roots compressor due to complex working environment is overcome, the accuracy of the model is further improved, so that the noise of the multi-sensor fusion data under complex working conditions can be removed to improve the fault diagnosis precision of the deep learning model, and the accuracy of fault recognition of the model under complex working conditions is improved. BRIEF DESCRIPTION OF DRAWINGS

[0027] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings obtained according to these drawings without creative labor are still within the scope of the present application.

[0028] Figure 1 A flow chart of a Roots compressor fault diagnosis method provided by an embodiment of the present application;

[0029] Figure 2 A framework diagram of a multi-scale hollow convolution attention auto-encoding model in a Roots compressor fault diagnosis method provided by an embodiment of the present application;

[0030] Figure 3 A framework diagram of a Transformer block contained in the multi-scale hollow convolution attention auto-encoding model in the embodiment; Figure 2 ​

[0031] Figure 4 For Figure 3 The framework diagram of the residual network in the Roots compressor fault diagnosis method provided by the embodiment of the application;

[0032] Figure 5 The framework diagram of the residual network in the Roots compressor fault diagnosis method provided by the embodiment of the application;

[0033] Figure 6 The overall framework diagram of the combination of the multi-scale hollow convolution attention auto-encoding model and the residual network in the Roots compressor fault diagnosis method provided by the embodiment of the application;

[0034] Figure 7 The result diagram of the analysis of the Roots compressor fault categories by using simulation data in the application scenario of the Roots compressor fault diagnosis method provided by the embodiment of the application;

[0035] Figure 8 The result diagram of the 5-fold cross-validation accuracy rate of the analysis of the Roots compressor fault categories by using simulation data in the application scenario of the Roots compressor fault diagnosis method provided by the embodiment of the application;

[0036] Figure 9 The result diagram of the analysis of the Roots compressor fault categories by using experimental data in the application scenario of the Roots compressor fault diagnosis method provided by the embodiment of the application;

[0037] Figure 10 The result diagram of the 5-fold cross-validation accuracy rate of the analysis of the Roots compressor fault categories by using experimental data in the application scenario of the Roots compressor fault diagnosis method provided by the embodiment of the application;

[0038] Figure 11 The structural schematic diagram of the Roots compressor fault diagnosis system provided by the embodiment of the application. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical scheme and advantages of the application clearer, the application will be further described in detail below with reference to the drawings.

[0040] As Figure 1 shown, the Roots compressor fault diagnosis method provided by the embodiment of the application includes the following steps:

[0041] Step S1, acquiring the vibration signal of a power shaft, the current of a motor input end, the outlet pressure of a compressor, the outlet flow of the compressor and the temperature signal of a liquid collected by a plurality of sensors on a Roots compressor to form a two-dimensional matrix;

[0042] Step S2, introducing the two-dimensional matrix into the trained noise reduction model for noise reduction to obtain a multi-scale signal feature matrix after noise reduction; wherein the noise reduction model is a new auto-encoding model fusing multi-scale hollow convolution and attention;

[0043] Step S3, introducing the obtained multi-scale signal feature matrix into the trained Roots compressor fault classification model to diagnose the fault category of the Roots compressor; wherein the Roots compressor fault classification model is a residual network model with residual connection.

[0044] The specific process is that, before step S1, the first step is to construct a noise reduction model based on a multi-scale hollow convolution and attention auto-encoding model (DAAE), so that the multi-scale hollow convolution attention auto-encoding has stronger noise reduction capability and can obtain deeper fault features and more global related features, i.e. the noise reduction model is a new auto-encoding model fusing multi-scale hollow convolution and attention.

[0045] The second step is to obtain the training samples of the noise reduction model, and the process is as follows:

[0046] (1) Set the sampling frequency (such as 60HZ), sample the power shaft, motor, compressor and liquid under multiple fault categories through multiple sensors preset on the Roots compressor, and obtain multiple sampling data under each fault category; wherein the fault categories include but are not limited to impeller failure, bearing failure, gear failure, transmission shaft damage, foundation bolt loosening, motor winding open circuit, winding ground and phase-to-phase short circuit, rotor cage broken strip and abnormal leakage; each sampling data includes the vibration signal of the power shaft (X-axis direction, Y-axis direction and Z-axis direction), the current of the motor input end, the outlet pressure of the compressor, the outlet flow of the compressor and the temperature signal of the liquid;

[0047] (2) Each sampling data is composed into a two-dimensional matrix and output as a sample to obtain a large number of training samples of the noise reduction model.

[0048] The third step is to determine the framework of the multi-scale hollow convolution attention auto-encoding model, as shown in Figure 2

[0049] The encoder first takes the two-dimensional matrix as input, adds noise damage signal to the two-dimensional matrix, uses a linear layer with activation function ReLU to extract features, puts the extracted features into a multi-scale hollow convolution block to capture semantic features, further separates the captured semantic features into three matrices of query, key and value through a separation layer, and then puts them into three Transformer blocks to learn noise distribution features and signal features.

[0050] ​It should be noted that, as shown in Figure 3 Each layer of the Transformer block in the encoder is composed of two sub-layers; one of the two sub-layers is the convergence of multi-head attention, and the other is a position feed-forward network; the calculation process between the two sub-layers is shown as and F i and respectively represent the input and output of the i-th layer; residual connection is used between the two sub-layers, and layer normalization is applied after the residual connection.

[0051] It should be noted that the multi-head attention in the sub-layer is obtained by concatenating multiple attention heads and then performing linear transformation, as shown in the framework diagram Figure 4 The mathematical expression of multi-head attention is: given the query key and value Each attention head can be represented as: h i =f(W i (q) q,W i (k) k,W i (v) v)∈R pv ; wherein, and and f represents the function of attention convergence, which are all learnable parameters. At this time, the output of the multi-head attention corresponds to the result of the combination of h heads, and the learnable parameters are which can be represented as:

[0052] The decoder extracts the noise distribution features and the signal itself features to obtain a multi-scale signal feature matrix; wherein the decoder includes a linear layer for predicting the added noise and a linear layer for reorganizing the signal features; after calculating the gradients of the noise prediction loss and the reconstruction loss, the decoder uses the method of stochastic gradient descent to update the weights. It should be noted that the loss L is used in the algorithm in the embodiment of the present application to supervise the denoising result of the multi-scale dilated convolution attention auto-encoding model, and the loss L is obtained by adding the noise loss L mask and the reorganization loss L rec L mask is a cross-entropy loss, and L rec is a mean square error loss, which can be mathematically expressed as and L=L mask +L rec; wherein c is the type of added noise position. Because the loss function is continuously derivable, the embodiment of the present application selects a stochastic gradient optimization (SGD) optimization to update the weight parameters.

[0053] In the fourth step, the training samples obtained in the third step are divided into a training set and a test set, and the denoising model is trained through the training set and tested through the test set to obtain the trained denoising model. At this time, the multi-scale signal feature matrix output by the denoising model during training is also used as a training sample for the subsequent Roots compressor fault classification model.

[0054] In the fifth step, after the denoising model is trained, a Roots compressor fault classification model is constructed based on a residual network, the features after denoising are classified by the residual network, the fault category of the Roots compressor is obtained, and the Roots compressor fault classification model is a residual network model with residual connection.

[0055] In the sixth step, the multi-scale signal feature matrix output by the denoising model during training is used as a training sample for the Roots compressor fault classification model.

[0056] In the seventh step, as shown in Figure 5 , the residual network framework is determined. The multi-scale signal feature matrix output by the denoising model is used as the input of the Resnet, after feature extraction by the 7*7 convolution layer and the batch normalization layer, the key features are extracted by the 3*3 maximum pooling layer, then the feature matrix is extracted by the four residual blocks, and finally the features of all channels are aggregated by the global average pooling layer, and the extracted signals are classified by the fully connected layer.

[0057] In the eighth step, the multi-scale signal feature matrix output by the denoising model in the sixth step is divided into a training set and a test set, the Roots compressor fault classification model is iteratively trained through the training set until the iteration number is reached, and then saved, and tested through the test set and 5-fold cross-validation, and finally the trained Roots compressor fault classification model is obtained.

[0058] Based on the above first to eighth steps, the overall framework diagram of the denoising model and the Roots compressor fault classification model can be known, as shown in Figure 6 .

[0059] In step S1, the relevant data required to be measured by the Roots compressor at present is obtained, including the vibration signal of the power shaft, the current of the motor input end, the outlet pressure of the compressor, the outlet flow of the compressor, and the temperature signal of the liquid, and the relevant data required to be measured at present is formed into a two-dimensional matrix.

[0060] In step S2, the two-dimensional matrix in step S1 is imported into the trained denoising model for denoising to obtain a multi-scale signal feature matrix after denoising.

[0061] In step S3, the multi-scale signal feature matrix obtained in step S2 is introduced into the trained Roots compressor fault classification model, and the fault category of the Roots compressor can be quickly diagnosed.

[0062] As Figure 7 to Figure 10 shown, the application scenario of the Roots compressor fault diagnosis method provided in the embodiment of the application is further described, and the specific process is as follows:

[0063] In one application scenario, the data set generated by the admas software simulation model in the laboratory is selected as sample data. The simulation data contains 7 groups of data, each group of data representing a fault data of a Roots compressor, and each group of data containing 1000 samples with a size of 1000*6 training set and 200 with a size of 1000*6 test set. Each sample of the Roots compressor is a two-dimensional matrix composed of six one-dimensional simulation fault features, such as x-axis direction vibration of the transmission shaft, y-axis direction vibration of the transmission shaft, z-axis direction vibration of the transmission shaft, compressor outlet pressure, compressor outlet flow and liquid temperature signal. At this time, the simulated fault categories include impeller failure, bearing failure, gear failure, transmission shaft damage, foundation bolt loosening, rotor cage broken strip and abnormal leakage.

[0064] First, the training sample with a size of 8400*1000*6 is taken as the input of the noise reduction model constructed by the multi-scale hollow convolution residual attention auto-encoding model (DAAE) to train the model. The batch size in training is 32, and the noise reduction model weight is updated by the SGD optimization algorithm, and the learning rate of the optimization algorithm is 0.001. After 400 times of training, the model is saved, the feature layer of the last layer of the DAAE model is extracted, and a multi-scale signal feature matrix with a size of 8400*1000*1536 is obtained.

[0065] The multi-scale signal feature matrix is split into 8400 samples with a size of 1000*1536. During the training of the Roots compressor fault classification model, 7000 samples are randomly taken as the training set, and 1400 samples are taken as the test set. The training batch size is 32, the iteration times are 30, the learning rate of the SGD optimization algorithm is 0.05, and the first training result is as shown in Figure 7 , and the 5-fold cross-validation diagram is as shown in Figure 8 . From Figure 7 and Figure 8 , it can be seen that the classification accuracy of the method on the Roots compressor fault data generated by the admas software simulation model is very high, and the method can accurately identify the fault category of the Roots compressor.

[0066] In another application scenario, a dataset collected by laboratory experiments is selected as sample data. The experimental data contains 10 groups of data, each group of data representing a fault data of a Roots compressor, and each group of data containing 1000 samples with a size of 1000*7 training set and 1000 samples with a size of 1000*7 test set. Each sample of each group of Roots compressor is a two-dimensional matrix composed of seven one-dimensional simulation fault features, such as x-axis direction vibration of the transmission shaft, y-axis direction vibration of the transmission shaft, z-axis direction vibration of the transmission shaft, current of the motor input end, compressor outlet pressure, compressor outlet flow and liquid temperature signal. At this time, the simulated fault categories include impeller failure, bearing failure, gear failure, transmission shaft damage, foundation bolt loosening, motor winding open circuit, winding ground and phase-to-phase short circuit, rotor cage broken strip and abnormal leakage.

[0067] First, the obtained three-dimensional data with a sample size of 12000*1000*7 is used as the input of the denoising model constructed by the multi-scale hollow convolution residual attention auto-encoding model (DAAE) to train the model. The batch size in training is 32, and the denoising model weight is updated by the SGD optimization algorithm, and the learning rate of the optimization algorithm is 0.001. After 400 training, the model is saved, the feature layer of the last layer of the DAAE model is extracted, and a multi-scale signal feature matrix with a size of 12000*400*1536 is obtained.

[0068] The multi-scale signal feature matrix is split into 12000 samples with a size of 1000*1536. During the training of each classification model, 10000 samples are randomly taken as the training set and 200 samples are taken as the test set. The training batch size is 32, the iteration times are 30, the learning rate of the SGD optimization algorithm is 0.05, and the training result is as shown in Figure 9 The 5-fold cross-validation diagram is as shown in Figure 10 From Figure 9 and Figure 10 , it can be seen that the classification accuracy of the method on the experimental Roots compressor fault data is very high, and the method can accurately identify the fault categories of the Roots compressor.

[0069] As shown in Figure 11 , it is a Roots compressor fault diagnosis system provided in an embodiment of the application, which comprises:

[0070] The multi-signal fusion unit 110 is configured to acquire vibration signals of a power shaft, a current of a motor input end, a compressor outlet pressure, a compressor outlet flow and a liquid temperature signal collected by a plurality of sensors on the Roots compressor to form a two-dimensional matrix.

[0071] The fusion signal denoising unit 120 is configured to input the two-dimensional matrix into a trained denoising model to perform denoising, and obtain a multi-scale signal feature matrix after denoising; wherein the denoising model is a new auto-encoding model fusing multi-scale hollow convolution and attention.

[0072] The fault category diagnosis unit 130 is configured to input the obtained multi-scale signal feature matrix into a trained Roots compressor fault classification model to diagnose a fault category of the Roots compressor; wherein the Roots compressor fault classification model is a residual network model with residual connection.

[0073] The fault category of the Roots compressor includes impeller failure, bearing failure, gear failure, transmission shaft damage, foundation bolt loosening, motor winding open circuit, winding ground and phase-to-phase short circuit, rotor cage strip breakage and abnormal leakage.

[0074] The embodiment of the present application has the following beneficial effects:

[0075] The denoising model constructed based on the multi-scale hollow convolution attention auto-encoding model has strong denoising capability, and in combination with the Roots compressor fault classification model constructed based on the residual network, the problem of low accuracy of fault discrimination of the Roots compressor due to complex working environment is overcome, the accuracy of the model is further improved, so that the noise of the multi-sensor fusion data under complex working conditions can be removed to improve the fault diagnosis precision of the deep learning model, and the accuracy of the model in complex working condition fault recognition is improved.

[0076] It should be noted that in the above system embodiment, each unit included is only divided according to functional logic, but is not limited to the above division, as long as the corresponding function can be realized; in addition, the specific name of each functional unit is only for easy differentiation, and does not limit the protection scope of the present application.

[0077] Those skilled in the art can understand that all or part of the steps of the above-mentioned embodiment methods can be completed by programs instructing related hardware, and the programs can be stored in a computer readable storage medium, such as ROM / RAM, magnetic disk, optical disc, etc.

[0078] The above disclosure is only the preferred embodiment of the present application, and of course cannot limit the scope of the present application, so equivalent changes made according to the claims of the present application are still within the scope of the present application.

Claims

1. A fault diagnosis method for a Roots compressor, characterized in that, The method includes the following steps: The vibration signal of the power shaft, the current at the motor input end, the compressor outlet pressure, the compressor outlet flow rate, and the liquid temperature signal are collected by multiple sensors on the Roots compressor to form a two-dimensional matrix; The two-dimensional matrix is ​​imported into the trained denoising model for denoising to obtain the denoised multi-scale signal feature matrix; wherein, the denoising model is a novel autoencoder model that integrates multi-scale dilated convolution and attention. The obtained multi-scale signal feature matrix is ​​imported into the trained Roots compressor fault classification model to diagnose the fault category of the Roots compressor; wherein, the Roots compressor fault classification model is a residual network model with residual connections added. The noise reduction model includes an encoder and a decoder; wherein... The encoder first takes the two-dimensional matrix as input, adds a noise-damaged signal to the two-dimensional matrix, extracts its features using a linear layer with the ReLU activation function, and puts the extracted features into a multi-scale dilated convolution block to capture semantic features. The captured semantic features are further decomposed into three matrices—queries, keys, and values—through a separation layer, and then put into a three-layer Transformer block to learn the noise distribution features and the features of the signal itself. The decoder extracts the noise distribution features and the signal features themselves to obtain a multi-scale signal feature matrix; wherein, the decoder includes a linear layer for predicting the added noise and a linear layer for reconstructing the signal features; the decoder updates the weights by backpropagating using stochastic gradient descent after calculating the gradients of the noise prediction loss and the reconstruction loss.

2. The fault diagnosis method for a Roots compressor as described in claim 1, characterized in that, Each Transformer block in the encoder consists of two sub-layers; wherein... One of the two sub-layers is a multi-head attention convergence layer, and the other is a location-based feedforward network; the computation process between the two sub-layers is shown below. and ; and They represent The layers have inputs and outputs; residual connections are used between the two sub-layers, and layer normalization is applied after the residual connections.

3. The fault diagnosis method for a Roots compressor as described in claim 2, characterized in that, The residual network uses the multi-scale signal feature matrix as input to ResNet. After feature extraction by a 7*7 convolutional layer and a batch normalization layer, key features are extracted by a 3*3 max pooling layer. Then, four residual blocks are used to extract features from the feature matrix again. Finally, a global average pooling layer is used to aggregate the features of all channels, and a fully connected layer is used to classify the extracted signals.

4. The fault diagnosis method for a Roots compressor as described in claim 1, characterized in that, The fault categories of the Roots compressor include impeller failure, bearing failure, gear failure, drive shaft damage, loose foundation bolts, open circuit in motor winding, winding grounding and phase-to-phase short circuit, broken rotor cage bars, and abnormal leakage.

5. A fault diagnosis system for a Roots compressor, characterized in that, include: The multi-signal fusion unit is used to acquire the vibration signal of the power shaft, the current at the motor input end, the compressor outlet pressure, the compressor outlet flow rate, and the liquid temperature signal collected by multiple sensors on the Roots compressor, in order to form a two-dimensional matrix; The fusion signal denoising unit is used to import the two-dimensional matrix into the trained denoising model for denoising to obtain a denoised multi-scale signal feature matrix; wherein, the denoising model is a novel autoencoder model that integrates multi-scale dilated convolution and attention. The fault category diagnosis unit is used to import the obtained multi-scale signal feature matrix into the trained Roots compressor fault classification model to diagnose the fault category of the Roots compressor; wherein, the Roots compressor fault classification model is a residual network model with residual connections added. The noise reduction model includes an encoder and a decoder; wherein... The encoder first takes the two-dimensional matrix as input, adds a noise-damaged signal to the two-dimensional matrix, extracts its features using a linear layer with the ReLU activation function, and puts the extracted features into a multi-scale dilated convolution block to capture semantic features. The captured semantic features are further decomposed into three matrices—queries, keys, and values—through a separation layer, and then put into a three-layer Transformer block to learn the noise distribution features and the features of the signal itself. The decoder extracts the noise distribution features and the signal features themselves to obtain a multi-scale signal feature matrix; wherein, the decoder includes a linear layer for predicting the added noise and a linear layer for reconstructing the signal features; the decoder updates the weights by backpropagating using stochastic gradient descent after calculating the gradients of the noise prediction loss and the reconstruction loss.

6. The Roots compressor fault diagnosis system as described in claim 5, characterized in that, The fault categories of the Roots compressor include impeller failure, bearing failure, gear failure, drive shaft damage, loose foundation bolts, open circuit in motor winding, winding grounding and phase-to-phase short circuit, broken rotor cage bars, and abnormal leakage.

Images