Air compressor fault detection method based on deep learning

Abstract

The present invention provides an air compressor fault detection method based on deep learning, which can effectively detect the fault types that occur in the air compressor and is more accurate in classifying the fault types than the current method. The present invention includes the following steps: S1, obtaining a training data set, that is, using sensors in different places in the air compressor to detect different input signals, processing and splicing the sensor input signals detected in different places, and finally converting the signal sequence into a one-dimensional tensor, and the obtained one-dimensional vector constitutes the final training data set; S2, constructing a fault detection model based on deep learning, the fault detection model input is the signal sequence that has been converted into a one-dimensional tensor in S1; S3, using the training data set obtained in S1 to train the model in the fault detection model established in S2; S4, using the trained model to detect the fault type of the air compressor and output the detected fault category.

Description

Technical Field

[0001] The present invention relates to an air compressor fault detection method based on deep learning. Background Art

[0002] Air compressors, devices used to compress air, have numerous failure modes, and thus the diagnosis of their fault types has long been a hot topic among fault diagnosis researchers both domestically and internationally. Early fault diagnosis of air compressors relied primarily on expert field experience, a method with high artifactual limitations. With the rapid advancement of sensor and computer technology, fault diagnosis methods have emerged that combine air compressor vibration, pressure, motor current, and other signals with deep learning algorithms, achieving promising diagnostic results in laboratory studies.

[0003] Currently, existing equipment failure case dictionaries are not very reliable for quickly locating faults and diagnosing them after they occur. Power equipment fault diagnosis methods are relatively simple, often relying on the maintenance experience of professional technicians and susceptible to subjective factors. In recent years, with the rise of deep learning technology, deep learning models, leveraging their powerful learning capabilities, can directly extract features from raw signals and perform data classification through training. This allows for rapid and accurate fault detection, which is crucial for subsequent maintenance and repair. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the purpose of the present invention is to provide an air compressor fault detection method based on deep learning, which can effectively detect the fault types of air compressors and classify fault types more accurately than current methods.

[0005] The purpose of the present invention can be achieved by the following technical solution: an air compressor fault detection method based on deep learning, characterized in that it includes the following steps:

[0006] S1, obtaining a training data set, that is, using sensors at different locations in the air compressor to detect different input signals, processing and splicing the sensor input signals detected at different locations, and finally converting the signal sequence into a one-dimensional tensor. The obtained one-dimensional vector constitutes the final training data set;

[0007] S2, builds a fault detection model based on deep learning. The input of this fault detection model is the signal sequence that has been converted into a one-dimensional tensor in S1, and the output is the detected fault type;

[0008] S3, using the training data set obtained in S1, trains the fault detection model established in S2;

[0009] S4, uses the trained model to detect the fault type of the air compressor and outputs the detected fault category.

[0010] In S2 of the present invention, the fault detection model includes a backbone network, a semantic information guidance module and a multi-scale fusion module; the backbone network is composed of two modules: a residual channel attention module and a convolutional attention information interaction module; at the same time, a branch is added next to the backbone network, and a semantic information guidance module is added to the branch; the multi-scale fusion module serves as the output part of the fault detection model.

[0011] In S2 of the present invention, building a backbone network includes the following steps:

[0012] The residual channel attention module is a residual structure module, which consists of a channel attention module and convolution. The calculation process of the residual channel attention module is as follows:

[0013] M i =σ(LN((f 1×3 (X i ))), X i+1 =X i +SE(LN(f 1×3 (M i ))),

[0014] where X i is the output of the previous layer, X i+1 is the output of the current layer, f 1×m (.) represents an over-parameterized convolution operation with a kernel size of 1×m, LN(.) represents a normalization layer, σ(.) represents an activation function, and SE(.) represents a channel attention module;

[0015] The convolutional attention information interaction module includes a lightweight attention module and a bottleneck structure. The overall calculation process of the convolutional attention information interaction module is as follows:

[0016] TR=Edg(f 1×1 (X)), CO=Bottle(TR), Y=Avg 1×3 (Avg 1×3 (Cat(TR,CO))),

[0017] Where X represents input, Y represents output, and f 1×1 (.) indicates a convolution operation with a convolution kernel size of 1×1, Edg(.) indicates the operation after the lightweight attention module, Bottle(.) indicates the operation after the bottleneck structure, Avg 1×m (.) represents the average pooling operation with a pooling kernel size of 1×m, and Cat(.) represents the splicing operation based on the channel dimension.

[0018] In S2 of the present invention, constructing a semantic information guidance module includes the following steps:

[0019] The semantic information guidance module is used for the interaction of feature information between different layers. The input of this module is the extracted features of different layers. First, a 1×3 convolution is used to extract the information from the adjacent feature layer as the query vector. Then, a 1×1 convolution is used in the current layer to filter the information as the key and value. The query vector extracted from the adjacent layer is multiplied by the key extracted from the current layer and then passed through the activation function (Softmax) to obtain the guidance weight. Then, after the product operation is performed with the value obtained from the current layer, it is fused with the information extracted from the adjacent layer. Finally, the output features after the interaction between different layers can be obtained. The calculation process of the semantic information guidance module SIG is as follows:

[0020] Q=LN(f 1×3 (X1)), K=LN(f 1×1 (X2)), V=LN(f 1×1 (X2)),

[0021] Y=Cat(σ((Q+P h +P w )·K)·V,(Q+P h +P w )),

[0022] Where X2 represents the feature information of the current layer; X1 represents the feature information of the adjacent layer; Q, K, and V represent the query vector, key, and value respectively; Cat(.) represents the splicing operation based on the channel dimension; f 1×m (.) represents the convolution operation with a convolution kernel size of 1×m; LN(.) represents the normalization layer; σ(.) represents the activation function; P w and P h They are two prior parameters based on different directions, representing the prior of spatial position in height and width dimensions, and are two learnable vectors; Y represents the final output.

[0023] In S2 of the present invention, constructing a multi-scale fusion module includes the following steps:

[0024] The multi-scale fusion module is used to fuse the features of the branches. It accepts inputs from two branches. The left branch is the feature processing mechanism for the main network, and the right branch is the feature processing mechanism for the auxiliary branch. In the right branch, 1×3 depth-separable convolution is used to process the information in the shallow features, where 1×1 convolution facilitates the adjustment of the number of feature channels. Different methods are used to process the features of different branches. The information extracted from the auxiliary branch is used to guide the classification of the main network. Then they are added and fused. Finally, a layer of 1×3 depth-separable convolution is used to filter the fused features, which can effectively reduce the loss of semantic information and maintain relatively rich feature representation during the detection process. The calculation process of constructing the multi-scale fusion module is as follows:

[0025] H1=σ(LN(f 1×1 (σ(LN(f 1×3 (X1)))))), L1=σ(LN(f 3×3 (X2))), L2=σ(LN(f 1×1 (σ(LN(f 1×3 (Up(X2))))))), Y=σ(LN(f 1×3 ((L1·H1)+(L2·H2)))),

[0026] Where X1 and X2 represent the features of different branches, LN(.) represents the normalization layer, σ(.) represents the activation function, and f 1×k (.) represents a depth-wise separable convolution with a kernel size of 1×k. represents a dilated convolution with a kernel size of 1×k, Cat(.) represents a concatenation operation based on the channel dimension, and Y represents the final output.

[0027] In S3 of the present invention, the model is trained using a supervised training mode, and the specific steps are:

[0028] The signals detected by the sensors are processed into a one-dimensional tensor as the input sequence of the fault detection model. The output and label are used to calculate the loss function. The batch size of each input is set to 128, and an equal-interval learning rate adjustment strategy is adopted to make the learning rate gradually decay with the increase of training times. The initial learning rate is set to 0.0005, the decay coefficient is 0.98, and the learning rate is updated every 3 rounds, for a total of 300 rounds of training.

[0029] The learning rate calculation formula is as follows:

[0030] lr N =lr0·β N / s ,

[0031] where lrN is the learning rate of the Nth training, lr0 is the initial learning rate, β is the decay coefficient, and s is the update interval;

[0032] The optimizer used in the training process is the Adam algorithm, and cross entropy is selected as the loss function. The formula is expressed as:

[0033]

[0034] Where x is the predicted fault category output by the model, class is the label, and Loss represents the loss function.

[0035] In S4 of the present invention, during detection, the model parameters with the highest accuracy of the validation set during the training iteration are used for detection.

[0036] In S4 of the present invention, a signal sequence is input, and after a detection operation is performed using trained weights to extract features, the fault type of the air compressor is output.

[0037] Beneficial effects of the present invention: The detection method of the present invention can effectively detect the fault type of the air compressor. Compared with the current method, it is more accurate in classifying the fault type, and can effectively reduce the interference of complex information and enhance the ability to distinguish different fault types. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 Schematic diagram of the positions of various sensors in the air circuit system of an air compressor in a specific embodiment of the present invention;

[0039] Figure 2 Schematic diagram of the positions of various sensors in the oil circuit system of an air compressor in a specific embodiment of the present invention;

[0040] Figure 3 It is a schematic diagram of the overall structure of the fault detection model in the present invention;

[0041] Figure 4a Schematic diagram of the residual channel attention module in the present invention;

[0042] Figure 4b Schematic diagram of the convolutional attention information interaction module in the present invention;

[0043] Figure 5 Schematic diagram of the semantic information guidance module in the present invention;

[0044] Figure 6 Schematic diagram of the multi-scale fusion module in the present invention;

[0045] Figure 7 In combination with the specific embodiment of the present invention Figure 1 、 Figure 2Table diagram of parameter models of each sensor of the structure;

[0046] Figure 8 In combination with the specific embodiment of the present invention Figure 1 、 Figure 2 A tabular diagram of the required inspection results for each fault of the structure. DETAILED DESCRIPTION

[0047] The present invention will be further described in detail below with reference to the accompanying drawings and through examples. The following examples are intended to explain the present invention but the present invention is not limited to the following examples.

[0048] like Figure 1 、 Figure 2 、 Figure 7 、 Figure 8 As shown, it is a parameter input schematic diagram of an embodiment of this solution in which the specific object is an air compressor of model STC-GC (210-3).

[0049] in, Figure 1 and Figure 2 This is a schematic diagram of the sensor installation positions in the gas and oil lines of the STC-GC (210-3) air compressor. Figure 1 、 2 Modules 1, 2, and 3 are 1st-stage air compression, 2nd-stage air compression, and 3rd-stage air compression.

[0050] Figure 7 for Figure 1 and Figure 2 The various models and parameters of the various sensors in the illustrated embodiment are the data inputs of the embodiment of the present invention.

[0051] Figure 8 for Figure 1 and Figure 2 Examples of faults detected by the illustrated embodiment are outputs of fault detection by an embodiment of the invention.

[0052] The air compressor fault detection method based on deep learning includes the following steps:

[0053] S1, obtain the training data set and divide the training data set into a training set and a validation set;

[0054] Sensors are used at different locations in the air compressor to detect different input signals. The sensors can be distributed at different locations on the air and oil circuits of the air compressor where failures are likely to occur. Figure 1 、 Figure 2 The sensors are distributed in different places of the air compressor. The different detected signals are processed and spliced, and finally the signal sequence is converted into a one-dimensional tensor. The one-dimensional vectors obtained after processing the sensor input signals at different places constitute the final training data set. Figure 7 Figure 2 shows an example of sensor condition monitoring in an air compressor.

[0055] S2: Build a fault detection model based on deep learning. The input of this fault detection model is the signal sequence that has been converted into a one-dimensional tensor in step S1, and the output is the detected fault type.

[0056] like Figure 3 As shown, the entire fault detection model includes a backbone network, a semantic information guidance module (SIG), and a multi-scale fusion module (MSF). The backbone network consists of two modules: the residual channel attention module (RCA) and the convolutional attention information interaction module (CAII). These modules are used to extract feature information at different levels. Integrating the attention mechanism into the feature extraction process helps the model focus on meaningful information when extracting features, avoiding interference from useless information. A branch is added to the backbone network of the model, incorporating the semantic information guidance module (SIG). This allows feature information at different levels to guide each other, addressing the limitations of features at a single level and providing more accurate global and semantic information for the subsequent classification process. In the output, this application uses the multi-scale fusion module (MSF) to fuse features from different branches, incorporating multi-scale convolution to strengthen the connections between feature information at different scales. Finally, a classifier is added to gradually reduce the number of channels in the network, acting as a buffer and avoiding the problem of inaccurate detection caused by direct output. The combination of the semantic information guidance module (SIG) and the multi-scale fusion module (MSF) gives the fault detection model proposed in this application a powerful ability to extract the intrinsic connections between different features. Due to the mutual guidance between feature information at different levels, the fault detection model proposed in this application can adapt to various types of fault detection tasks. In addition, the final output head Classifier uses two convolutional layers to gradually reduce the number of network channels to act as a buffer, avoiding the problem of inaccurate detection caused by direct output.

[0057] The specific steps to build a deep learning-based fault detection model are:

[0058] S21, building a backbone network;

[0059] like Figure 4a and Figure 4b As shown in Figure 1, the backbone network consists of two modules: the residual channel attention module RCA and the convolutional attention information interaction module CAII.

[0060] Existing networks usually use existing convolutional networks for feature extraction. The most common one is the residual network ResNet, which contains a residual connection structure that can make the network structure deeper while avoiding the problem of network degradation. However, ResNet uses a pure convolutional structure, which has limitations and is not conducive to modeling global information in the input sequence. Here, the backbone network of this application consists of two modules: the residual channel attention module RCA and the convolutional attention information interaction module CAII.

[0061] The residual channel attention module is a residual structure module. Unlike ResNet, it incorporates an attention mechanism. The residual channel attention module consists of a channel attention module SEModule and convolution. SEModule is a classic channel attention module. Since different channel levels have different spatial information, the addition of SEModule helps to accurately locate the channel level where important information is located, making the network more efficient in extracting feature information. The calculation process of the residual channel attention module is as follows:

[0062] M i =σ(LN((f 1×3 (X i ))),

[0063] X i+1 =X i +SE(LN(f 1×3 (M i ))),

[0064] where X i is the output of the previous layer, X i+1 is the output of the current layer, f 1×m (.) represents an over-parameterized convolution operation with a convolution kernel size of 1×m, LN(.) represents a normalization layer, σ(.) represents an activation function, and SE(.) represents the SEModule module.

[0065] The convolutional attention information interaction module includes a lightweight attention module EdgViT_block and a bottleneck structure Bottleneck. In the convolutional attention information interaction module, the present invention uses 1×3 and 3×1 over-parameterized convolutions instead of the original 3×3 ordinary convolution structure, which helps reduce the number of model parameters and accelerate the network training process.

[0066] As the number of network layers continues to increase, the convolutional attention information interaction module CAII is used in the deep layers of the network for feature extraction. The lightweight attention module EdgViT_block comes from a lightweight attention network EdgeViTs. It decomposes self-attention into continuous modules to process spatial information in different ranges, and further reduces the overhead of the self-attention mechanism through a sparse attention module, thereby achieving a better accuracy-latency balance. The bottleneck structure Bottleneck is the bottleneck module in the residual network ResNet. Since the lightweight attention module EdgViT_block can learn features in a global range, while convolution only captures local features, the global information obtained by the lightweight attention module EdgViT_block can be extracted locally through the local refinement of the bottleneck structure Bottleneck under the guidance of global information. The final splicing operation merges the global features with the local features. After adjustment through two layers of average pooling layers, the number of output parameters is reduced and the translation invariance of the features is improved. The overall calculation process of the convolutional attention information interaction module is as follows:

[0067] TR=Edg(f 1×1 (X)),

[0068] CO=Bottle(TR),

[0069] Y=Avg 1×3 (Avg 1×3 (Cat(TR,CO))),

[0070] Where X represents input, Y represents output, and f 1×1 (.) indicates the convolution operation with a convolution kernel size of 1×1, Edg(.) indicates the operation after EdgViT_block, Bottle(.) indicates the operation after Bottleneck, Avg 1×m (.) represents the average pooling operation with a pooling kernel size of 1×m, Cat(.) represents the splicing operation based on the channel dimension, and Y represents the final output.

[0071] S22, constructing semantic information guidance module;

[0072] The semantic information guidance module SIG is used to interact with feature information between different layers and strengthen features. The input of this module is the extracted features of different layers. By interactively guiding the feature information of adjacent layers, it is helpful to better distinguish different categories of ground objects. First, a 1×3 convolution is used to extract information from the adjacent feature layer as the query vector query. Then, a 1×1 convolution is used in the current layer to filter information as the key key and value val. w and Ph These are two prior parameters based on different directions, representing the priors of spatial position in the height and width dimensions. They are two learnable vectors. The query vector query extracted from the adjacent layer is multiplied by the key key extracted from the current layer and then passed through the activation function Softmax to obtain the guidance weight. This is then multiplied with the value val obtained from the current layer and fused with the information extracted from the adjacent layer. Finally, the output features of the interaction between different layers are obtained. The calculation process of the semantic information guidance module SIG is as follows:

[0073] Q=LN(f 1×3 (X1)),

[0074] K=LN(f 1×1 (X2)),

[0075] V=LN(f 1×1 (X2)),

[0076] Y=Cat(σ((Q+P h +P w )·K)·V,(Q+P h +P w )),

[0077] Where X2 represents the feature information of the current layer, X1 represents the feature information of the adjacent layer, Q, K, and V represent the query vector, key, and value obtained after a series of operations, respectively, and Cat(.) represents the splicing operation based on the channel dimension. 1×m (.) represents a convolution operation with a kernel size of 1×m, LN(.) represents a normalization layer, σ(.) represents an activation function, and Y represents the final output.

[0078] S23, building a multi-scale fusion module;

[0079] The multi-scale fusion module is used to fuse the features of the branches. It accepts inputs from two branches. The left branch is the feature processing mechanism for the main network, and the right branch is the feature processing mechanism for the auxiliary branch. In the right branch, 1×3 depthwise separable convolutions are used to process the information in the shallow features. The 1×1 convolution facilitates the adjustment of the number of feature channels. Multi-scale convolution is introduced in the right branch because information in different ranges may be associated during the fault detection classification process. The use of multi-scale convolution is conducive to extracting the relationship between features at different distances. The dilated convolution has a larger receptive field and is more beneficial for detail extraction. The features of different branches are processed in different ways. The information extracted from the auxiliary branch is used to guide the classification of the main network. Then they are added and fused. Finally, a layer of 1×3 depthwise separable convolution is used to filter the fused features. This can effectively reduce the loss of semantic information and maintain a relatively rich feature representation during the detection process. The calculation process for constructing the multi-scale fusion module is as follows:

[0080] H1=σ(LN(f 1×1 (σ(LN(f 1×3 (X1)))))),

[0081]

[0082] L1=σ(LN(f 3×3 (X2))),

[0083] L2=σ(LN(f 1×1 (σ(LN(f 1×3 (Up(X2))))))),

[0084] Y=σ(LN(f 1×3 ((L1·H1)+(L2·H2)))),

[0085] Where X1 and X2 represent the features of different branches, LN(.) represents the normalization layer, σ(.) represents the activation function, and f 1×k (.) represents a depth-wise separable convolution with a kernel size of 1×k. represents a dilated convolution with a kernel size of 1×k, and Cat(.) represents a splicing operation based on the channel dimension.

[0086] S3, using the data set obtained in S1, trains the fault detection model established in S2;

[0087] The training of the model of the present invention adopts a supervised training mode, and the specific steps are:

[0088] The signals detected by the sensors are processed into a one-dimensional tensor as the input sequence of the fault detection model. The output and label are used to calculate the loss function. The batch size of each input is set to 128, and an equal-interval learning rate adjustment strategy is adopted to make the learning rate gradually decay with the increase of training times. The initial learning rate is set to 0.0005, the decay coefficient is 0.98, and the learning rate is updated every 3 rounds, for a total of 300 rounds of training.

[0089] The learning rate calculation formula is as follows:

[0090] lr N =lr0·β N / s ,

[0091] where lr N is the learning rate of the Nth training, lr0 is the initial learning rate, β is the decay coefficient, and s is the update interval.

[0092] The optimizer used in the training process is the Adam algorithm, and the cross entropy is selected as the loss function. The formula is expressed as follows:

[0093]

[0094] Where x is the predicted fault category output by the model, class is the label, and Loss represents the loss function.

[0095] S4, use the trained model to detect the fault type of the air compressor, and output the detected fault category. During the detection, the model parameters with the highest accuracy in the validation set during the training iteration are used for detection; a signal sequence is input, and the features are extracted through the trained weights for detection operation, and then the fault type of the air compressor is output.

[0096] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention, and such changes and modifications fall within the scope of the invention as claimed.

Claims

1. Air compressor fault detection method based on deep learning, characterized by: The following steps are involved: S1, obtaining a training data set, that is, using sensors at different locations in the air compressor to detect different input signals, processing and splicing the sensor input signals detected at different locations, and finally converting the signal sequence into a one-dimensional tensor. The obtained one-dimensional vector constitutes the final training data set; S2, builds a fault detection model based on deep learning. The input of this fault detection model is the signal sequence that has been converted into a one-dimensional tensor in S1, and the output is the detected fault type; S3, using the training data set obtained in S1, trains the fault detection model established in S2; S4, uses the trained model to detect the fault type of the air compressor and outputs the detected fault category; In S2, the fault detection model includes a backbone network, a semantic information guidance module, and a multi-scale fusion module; the backbone network is composed of two modules: a residual channel attention module and a convolutional attention information interaction module; at the same time, a branch is added next to the backbone network, and the semantic information guidance module is added to the branch; The multi-scale fusion module serves as the output of the fault detection model; Building a backbone network includes the following steps: The residual channel attention module is a residual structure module, which consists of a channel attention module and convolution. The calculation process of the residual channel attention module is as follows: M i σ(LN((f 1×3 (X i ))),X i+1 DX i +SE(LN(f 1×3 (M i ))), where X i is the output of the previous layer, X i+1 is the output of the current layer, f 1×m (.) denotes an over-parameterized convolution operation with a kernel size of 1×m, LN(.) denotes a normalization layer, σ(.) denotes an activation function, and SE(.) denotes a channel attention module; The convolutional attention information interaction module includes a lightweight attention module and a bottleneck structure. The overall calculation process of the convolutional attention information interaction module is as follows: TR=Edg(f 1×1 (X)),CO=Bottle(TR),Y=Avg 1×3 (Avg 1×3 (Cat(TR,CO))), Where X represents input, Y represents output, and f 1×1 (.) indicates a convolution operation with a convolution kernel size of 1×1, Edg(.) indicates the operation after the lightweight attention module, Bottle(.) indicates the operation after the bottleneck structure, Avg 1×m (.) represents the average pooling operation with a pooling kernel size of 1×m, and Cat(.) represents the splicing operation based on the channel dimension; Building a semantic information guidance module includes the following steps: The semantic information guidance module is used for the interaction of feature information between different layers. The input of this module is the extracted features of different layers. First, a 1×3 convolution is used to extract the information from the adjacent feature layer as the query vector. Then, a 1×1 convolution is used in the current layer to filter the information as the key and value. The query vector extracted from the adjacent layer is multiplied by the key extracted from the current layer and then passed through the activation function (Softmax) to obtain the guidance weight. Then, after the product operation is performed with the value obtained from the current layer, it is fused with the information extracted from the adjacent layer. Finally, the output features after the interaction between different layers can be obtained. The calculation process of the semantic information guidance module SIG is as follows: Q=LN(f 1×3 (X1)),K=LN(f 1×1 (X2)),V=LN(f 1×1 (X2)), Y=Cat(σ((Q+P h +P w )·K)·V,(Q+P h +P w )), Where X2 represents the feature information of the current layer; X1 represents the feature information of the adjacent layer; Q, K, and V represent the query vector, key, and value respectively; Cat(.) represents the splicing operation based on the channel dimension; f 1×m (.) represents the convolution operation with a convolution kernel size of 1×m; LN(.) represents the normalization layer; σ(.) represents the activation function; P w and P h They are two prior parameters based on different directions, representing the prior of spatial position in height and width dimensions, and are two learnable vectors; Y represents the final output; Building a multi-scale fusion module includes the following steps: The multi-scale fusion module is used to fuse the features of the branches. It receives input from two branches. The left branch is the feature processing mechanism of the main network, and the right branch is the feature processing mechanism of the auxiliary branch. In the right branch, 1×3 depth-wise separable convolution is used to process the information in the shallow features. The information extracted from the auxiliary branch is used to guide the classification of the main network. Then, they are added and fused. Finally, a layer of 1×3 depth-wise separable convolution is used to filter the fused features. The calculation process of constructing the multi-scale fusion module is as follows: L1=σ(LN(f 3×3 (X2))),L2=σ(LN(f 1×1 (σ(LN(f 1×3 (Up(X2))))))), And=σ(LN(f 1×3 ((L1·H1)+(L2·H2)))), Where X1 and X2 represent the features of different branches, LN(.) represents the normalization layer, σ(.) represents the activation function, and f 1×k (.) represents a depth-wise separable convolution with a kernel size of 1×k. represents a dilated convolution with a kernel size of 1×k, Cat(.) represents a concatenation operation based on the channel dimension, and Y represents the final output.

2. The air compressor fault detection method based on deep learning according to claim 1, characterized in that: In S3, the model is trained using a supervised training mode. The specific steps are as follows: The signals detected by the sensors are processed into a one-dimensional tensor as the input sequence of the fault detection model. The output and label are used to calculate the loss function. The batch size of each input is set to 128, and an equal-interval learning rate adjustment strategy is adopted to make the learning rate gradually decay with the increase of training times. The initial learning rate is set to 0.0005, the decay coefficient is 0.98, and the learning rate is updated every 3 rounds, for a total of 300 rounds of training. The learning rate calculation formula is as follows: lr N ＝lr0·β N / s ， where lr N is the learning rate of the Nth training, lr0 is the initial learning rate, β is the decay coefficient, and s is the update interval; The optimizer used in the training process is the Adam algorithm, and cross entropy is selected as the loss function. The formula is expressed as: Where x is the predicted fault category output by the model, class is the label, and Loss represents the loss function.

3. The air compressor fault detection method based on deep learning according to claim 1, characterized in that: In S4, during detection, the model parameters with the highest accuracy of the validation set during the training iteration are used for detection.

4. The air compressor fault detection method based on deep learning according to claim 1, characterized in that: In S4, a signal sequence is input, and after the trained weights are used to extract features for detection, the fault type of the air compressor is output.

Images