Industrial equipment fault prediction method and device based on deep learning

Abstract

The invention discloses an industrial equipment fault prediction method and device based on deep learning, and belongs to the technical field of industrial automation and intelligent detection. The method comprises the following steps: mapping a one-dimensional time sequence signal into a time sequence angle, and calculating the time sequence angle through a dynamic weighting Grubrum matrix to obtain a two-dimensional image; constructing a deep learning network structure to extract local features and global features of the two-dimensional image; splicing the local features and the global features to obtain fusion features; and carrying out dimensionality reduction on the fusion features through a self-adaptive average pooling layer, then carrying out full-connection layer classification, and outputting a fault prediction result. The device comprises a signal acquisition module, an edge calculation module, a signal transmitting module, a signal receiving module and a server. According to the invention, a multi-module architecture integrating hardware acquisition, edge calculation and a software algorithm is adopted, the spatial-temporal characteristics of a one-dimensional time sequence signal are effectively captured by improving a signal processing and deep learning network structure, local details and a global structure are considered, and the fault detection precision of industrial equipment is effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of industrial automation and intelligent detection technology, and more specifically, to a method and device for predicting industrial equipment failures based on deep learning. Background Art

[0002] In recent years, fault prediction and diagnosis technology for industrial equipment has received increasing attention. With the increasing complexity of industrial equipment, traditional signal processing methods such as time domain, frequency domain and time-frequency domain analysis have gradually exposed their limitations in dealing with complex signals. Especially when dealing with nonlinear and non-stationary signals, traditional methods often find it difficult to extract deep features, resulting in low accuracy in fault identification. This limitation is particularly prominent in modern industry and cannot meet the needs of efficient and accurate equipment health monitoring.

[0003] With the rise of deep learning technology, convolutional neural networks (CNNs) have gradually emerged in the field of fault detection. Convolutional neural networks greatly simplify the complexity of feature engineering through automated feature extraction, avoiding a lot of manual intervention in traditional methods. However, convolutional neural networks are not omnipotent. They perform well in extracting local features, but have certain shortcomings in processing global features. The translation invariance of convolutional neural networks makes them insensitive to changes in the position of features in the signal when processing features. This is particularly evident in one-dimensional time series signals, because the relative position of features is often crucial to the meaning of the signal. Therefore, convolutional neural networks, especially in complex industrial environments, may not be able to fully reflect the operating status of equipment by relying solely on local information.

[0004] After searching, the Chinese patent application, application number 201710469573.2, discloses a method for predicting industrial equipment faults based on deep learning. The method first collects sensor data of industrial equipment through sensors, then obtains a spectrum diagram based on the time series wave of the sensor data within a fixed time, and finally uses a deep learning algorithm based on the convolutional neural network framework to predict the faults of industrial equipment based on the spectrum diagram, so as to accurately predict whether the industrial equipment is faulty or not. However, when this method obtains the spectrum diagram through the time series wave of the sensor data within a fixed time, it uses Fourier transform operation to decompose the time series wave and then obtain the time series wave spectrum diagram, which cannot flexibly focus on different areas of the time series wave signal, and there is a problem of insufficient capture of time series wave information. Summary of the invention

[0005] 1. Technical problems to be solved

[0006] In view of the problems in the prior art of insufficient feature extraction and low fault recognition rate of one-dimensional time series signal fault detection, the present invention provides a method and device for industrial equipment fault prediction based on deep learning. By improving signal processing technology, the one-dimensional time series signal is converted into a two-dimensional image, and a deep learning network structure is constructed to extract feature information from the two-dimensional image. The fault prediction of industrial equipment is achieved through the acquired feature information, which effectively improves the accuracy and robustness of one-dimensional time series signal fault detection and is suitable for intelligent monitoring and preventive maintenance of various types of rotating machinery and equipment. 2. Technical Solution

[0007] The purpose of the present invention is achieved through the following technical solutions.

[0008] A method for predicting industrial equipment failure based on deep learning, comprising the following steps:

[0009] Collect one-dimensional time series signals of industrial equipment;

[0010] Preprocessing one-dimensional time series signals;

[0011] Convert the preprocessed one-dimensional time series signal into a two-dimensional image;

[0012] Deep learning networks predict faults in industrial equipment based on 2D images.

[0013] As a further improvement of the present invention, the specific steps of converting the preprocessed one-dimensional time series signal into a two-dimensional image are as follows:

[0014] A one-dimensional time series signal is set, and normalization processing is performed on the one-dimensional time series signal;

[0015] Map the normalized one-dimensional time series signal to a time series angle;

[0016] The two-dimensional image is obtained by calculating the time series angle through the dynamic weighted Gram matrix.

[0017] As a further improvement of the present invention, the normalized one-dimensional time series signal is mapped to a time series angle, and the mapping formula is:

[0018] φ=arccos(x′)

[0019] Among them, φ represents the timing angle, x represents the one-dimensional timing signal, arccos represents the trigonometric function, and x′ represents the normalized one-dimensional timing signal.

[0020] As a further improvement of the present invention, the dynamically weighted Gram matrix includes a Gram angle and difference field matrix and a Gram angle and field matrix; the two-dimensional image is calculated by the Gram angle and difference field matrix or the Gram angle and field matrix.

[0021] As a further improvement of the present invention, a two-dimensional image is generated by calculating the time series angle through the Gram angle and the difference field matrix. The calculation formula is:

[0022] GAF i，j =DTW i，j ×sin(φ i +φ j )

[0023] Among them, GAF i,j Represents a two-dimensional image, i represents the row index of the Gram angle and difference field matrix, j represents the column index of the Gram angle and difference field matrix, DTW i,j represents the cumulative distance matrix, φ i Indicates the angle value at the i-th moment of the timing angle, φ j Indicates the angle value at the jth moment of the timing angle;

[0024] The two-dimensional image is generated by calculating the time series angle through the Gram angle and field matrix. The calculation formula is:

[0025] GAF i，j =DTW i，j ×cos(φ i +φ j )

[0026] Among them, GAF i,j Represents a two-dimensional image, i represents the row index of the Gram angle and field matrix, j represents the column index of the Gram angle and field matrix, DTW i,j represents the cumulative distance matrix, φ i Indicates the angle value at the i-th moment of the timing angle, φ j Indicates the angle value at the jth moment of the timing angle.

[0027] As a further improvement of the present invention, the deep learning network includes a SwinTransformer model and a two-dimensional convolutional network based on a global attention mechanism;

[0028] The SwinTransformer model extracts local features of a two-dimensional image through a window attention mechanism, and the calculation formula is:

[0029] F local =SwinTransformer(Linear(GAF i，j )+PE)

[0030] Among them, F local Represents local features, PE represents position encoding, and Linear represents linear layer.

[0031] As a further improvement of the present invention, the two-dimensional convolutional network based on the global attention mechanism extracts the global features of the two-dimensional image through the GAM-Attention module, and the calculation formula is:

[0032] F conv =CNN2d(GAF i，j )

[0033] Attention GAM =σ(W2·ReLU(W1·F conv +b1)+b2)

[0034] F global =Attention GAM (F conv )

[0035] Among them, CNN2d represents a two-dimensional convolutional neural network, F conv Represents the features extracted by the convolutional layer, Attention GAM represents the global attention mechanism module, σ represents the activation function, W1 and W2 represent weight matrices, b1 and b2 represent bias terms, ReLU represents the ReLU activation function, and F global Represents global features.

[0036] As a further improvement of the present invention, the local features and the global features are spliced ​​to obtain fused features; the fused features are expressed as:

[0037] F fused =Concat(F local , F global )

[0038] Among them, F fused represents the fusion feature, and Concat(·) represents the concatenation operation in the channel dimension.

[0039] As a further improvement of the present invention, the fusion features are reduced in dimension by an adaptive average pooling layer, and then classified by a fully connected layer to output a fault prediction result; the fault prediction result is expressed as:

[0040]

[0041] in, represents the fault prediction result, GAP(·) represents the global average pooling layer, and FC(·) represents the fully connected layer.

[0042] A deep learning-based industrial equipment fault prediction device, comprising: a signal acquisition module, an edge computing module, a signal transmission module, a signal receiving module and a server;

[0043] The signal acquisition module acquires one-dimensional time series signals of industrial equipment;

[0044] The edge computing module pre-processes the one-dimensional time series signal;

[0045] The signal transmitting module transmits the preprocessed one-dimensional time series signal to the signal receiving module through a transmission medium;

[0046] The signal receiving module transmits the preprocessed one-dimensional time series signal to the server;

[0047] The server converts the preprocessed one-dimensional time series signal into a two-dimensional image, and the deep learning network predicts faults of the industrial equipment based on the two-dimensional image.

[0048] 3. Beneficial effects

[0049] Compared with the prior art, the advantages of the present invention are:

[0050] (1) The present invention provides a method and device for predicting industrial equipment faults based on deep learning. By converting a one-dimensional time series signal into a two-dimensional image, the temporal and spatial characteristics of the one-dimensional time series signal are effectively captured. This not only retains the original information of the one-dimensional time series signal, but also extracts the dynamic changes and overall trends in the one-dimensional time series signal, significantly enhancing the comprehensiveness and accuracy of feature extraction.

[0051] (2) The present invention provides a method and device for predicting industrial equipment faults based on deep learning. The window attention mechanism of the SwinTransformer model is used to extract local features, and the GAM-Attention module is used to enhance the ability to capture global features. Through the combination of this dual attention mechanism, the deep learning network can take into account both local details and global structures when processing complex fault modes, thereby improving the detection accuracy of one-dimensional time series signal faults.

[0052] (3) The present invention provides a method and device for predicting industrial equipment faults based on deep learning, which combines local features and global features to obtain fused features. Through the design of the feature fusion layer, it can effectively integrate feature information from local and global sources, thereby reducing missed detections and false detections caused by feature loss during the detection process. This fusion method ensures the sensitivity of the deep learning network in detecting small fault features, thereby improving the accuracy and reliability of detection.

[0053] (4) The present invention provides a method and device for industrial equipment fault prediction based on deep learning. By constructing a deep learning network, better results can be achieved for industrial equipment fault prediction by using only lightweight detectors. The use of detectors with simpler structures in the process of industrial equipment fault prediction can greatly reduce the resources required for calculation, improve the detection speed of the deep learning network and reduce the memory usage required in the detection process, avoid unnecessary waste, and effectively utilize limited computing resources, thereby realizing the practical application of deep learning networks in industrial equipment fault prediction. The method and device have strong practicality and wide applicability. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 This is a flow chart of a method for predicting industrial equipment failures according to an embodiment of the present invention;

[0055] Figure 2 A schematic diagram of converting a one-dimensional time series signal into a two-dimensional image through a Gram angle and a difference field matrix according to an embodiment of the present invention;

[0056] Figure 3 A schematic diagram of converting a one-dimensional time series signal into a two-dimensional image through a Gram angle and a field matrix according to an embodiment of the present invention;

[0057] Figure 4 This is a graph showing the change in the loss function of the deep learning network prediction effect according to an embodiment of the present invention;

[0058] Figure 5 This is a graph showing the change in accuracy of the prediction effect of the deep learning network according to an embodiment of the present invention;

[0059] Figure 6 Schematic diagram of an industrial equipment fault prediction device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0060] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0061] Example

[0062] like Figure 1 As shown, a method for predicting industrial equipment faults based on deep learning is provided in this embodiment. The method includes the following steps: collecting one-dimensional time series signals of industrial equipment, preprocessing the one-dimensional time series signals; converting the preprocessed one-dimensional time series signals into two-dimensional images; and predicting faults of industrial equipment based on the two-dimensional images using a deep learning network.

[0063] Specifically in this embodiment, one-dimensional time series signals of industrial equipment are collected and preprocessed. In this embodiment, the one-dimensional time series signals include key indicators such as vibration and temperature. In this embodiment, preprocessing includes existing filtering and noise reduction, removal of invalid values, and filling of missing values. By preprocessing the one-dimensional time series signal, the legitimacy and integrity of the one-dimensional time series signal data are guaranteed. It should be noted that in this embodiment, the .mat format data file of the one-dimensional time series signal also needs to be parsed into a format that can be used for subsequent processing.

[0064] Therefore, in this embodiment, one-dimensional time series signal data collected from an industrial environment is used, and after preprocessing and normalization operations, the stability and availability of the data are ensured. Through standardization, the one-dimensional time series signal data is adjusted to the same scale range, providing a basis for subsequent signal conversion and feature extraction.

[0065] Furthermore, the preprocessed one-dimensional time series signal is converted into a two-dimensional image. Specifically, a one-dimensional time series signal is set, and the one-dimensional time series signal is expressed as:

[0066] x={x1,x2,...,x n}

[0067] Wherein, x represents a one-dimensional time series signal, and n represents a natural number.

[0068] The one-dimensional time series signal is normalized and the calculation formula is:

[0069]

[0070] Wherein, x′ represents the normalized one-dimensional time series signal, max(x) represents the maximum value of the one-dimensional time series signal, and min(x) represents the minimum value of the one-dimensional time series signal.

[0071] The normalized one-dimensional time series signal is mapped to a time series angle. It should be noted that, in this embodiment, the normalized one-dimensional time series signal is mapped to a time series angle, which uses trigonometric functions (arccos) to enhance the characteristic expression of the one-dimensional time series signal, so that the trend and pattern of the one-dimensional time series signal can be visualized in a two-dimensional space.

[0072] Specifically, by mapping the normalized one-dimensional time series signal to a time series angle, the value of the one-dimensional time series signal can be mapped to a polar coordinate system, and then the value in the dynamic weighted Gram angle matrix (DW-GAF) is calculated by trigonometric function (arccos).

[0073] In this embodiment, the normalized one-dimensional time series signal is mapped to a time series angle, and the mapping formula is:

[0074] φ=arccos(x′)

[0075] Among them, φ represents the timing angle, x represents the one-dimensional timing signal, arccos represents the trigonometric function, and x′ represents the normalized one-dimensional timing signal.

[0076] The two-dimensional image is obtained by calculating the timing angle through the dynamic weighted Gram matrix (DW-GAF). The dynamically weighted Gram angle spectrum includes the Gram angle and difference field matrix (GADF) and the Gram angle and field matrix (GASF). It should be noted that, in the present embodiment, the two-dimensional image can be calculated by the Gram angle and difference field matrix (GADF) or the Gram angle and field matrix (GASF). The Gram angle and difference field matrix (GADF) reflects the difference in timing angles and is used to represent the dynamic pattern of changes in one-dimensional timing signals. The Gram angle and field matrix (GASF) reflects the accumulation of timing angles and is used to represent the overall trend of one-dimensional timing signals.

[0077] Specifically, Figure 2 As shown, the Gram angle and difference field matrix (GADF) are used to calculate the time series angle to generate a two-dimensional image. The calculation formula is:

[0078] GAF i，j =DTW i，j ×sin(φ i +φ j )

[0079] Among them, GAF i,j Represents a two-dimensional image, i represents the row index of the Gram angle and the difference field matrix, which is the index of the i-th time point of the time series angle, j represents the column index of the Gram angle and the difference field matrix, which is the index of the j-th time point of the time series angle, DTW i,j represents the cumulative distance matrix, φ i Indicates the angle value at the i-th moment of the timing angle, φ j Indicates the angle value at the jth moment of the timing angle.

[0080] like Figure 3 As shown, the Gram angle and field matrix (GASF) is used to calculate the time series angle to generate a two-dimensional image. The calculation formula is:

[0081] GAF i，j =DTW i，j ×cos(φ i +φ j )

[0082] Among them, GAF i,j Represents a two-dimensional image, i represents the row index of the Gram angle and field matrix, that is, the index of the i-th time point of the time series angle, j represents the column index of the Gram angle and field matrix, that is, the index of the j-th time point of the time series angle, DTW i,jrepresents the cumulative distance matrix, φ i Indicates the angle value at the i-th moment of the timing angle, φ j Indicates the angle value at the jth moment of the timing angle.

[0083] It should be noted that, in this embodiment, the cumulative distance matrix DTW i,j It represents the dynamic alignment cost between the i-th and j-th points in a one-dimensional time series signal. Its calculation formula is based on dynamic programming. The specific steps are as follows:

[0084] Assume two one-dimensional time series signal sequences X = {x1, x2, ..., x m} and Y = {y1, y2, ..., y n}, calculate the Euclidean distance between each one-dimensional time series signal, construct the distance matrix D, and calculate the cumulative distance matrix DTW through dynamic programming. In this embodiment, the distance between each one-dimensional time series signal can be obtained by Euclidean distance calculation, and the calculation formula is:

[0085] D i，j =(x i -y j ) 2

[0086] Among them, x i Represents the value of the i-th time point in the sequence X, y i represents the value of the jth time point in sequence Y, D i,j Represents the distance between the i-th point in sequence X and the j-th point in sequence Y.

[0087] Cumulative distance matrix DTW i,j It represents the minimum cumulative alignment cost from the starting point (1,1) to the position (i,j), and its recursive calculation formula is:

[0088] DTW i，j =D i，j +min(DTW i-1，j , DTW i，j-1 , DTW i-1，j-1 )

[0089] Among them, DTW i-1,j Represents the cost of reaching the current position from above, DTW i,j-1 Represents the cost of reaching the current position from the left, DTW i-1,j-1 Represents the cost of reaching the current position from the diagonal line, DTW 1,1 =D 1,1 Indicates that the starting point of the cumulative distance matrix is ​​the starting point of the distance matrix, and the boundary conditions at other locations are set to:

[0090] DTW i，1 =Di，1 +DTW i-1，1 , DTW 1，j =D 1，j +DTW 1，j-1

[0091] Ensure that the elements on the boundary can be accumulated normally. Finally, DTW m,n Represents the minimum cumulative alignment cost from the start to the end of the sequence.

[0092] Therefore, in this embodiment, by converting the one-dimensional time series signal into a two-dimensional image, the spatiotemporal characteristics of the one-dimensional time series signal are effectively captured, which not only retains the original information of the one-dimensional time series signal, but also extracts the dynamic changes and overall trends in the one-dimensional time series signal, significantly enhancing the comprehensiveness and accuracy of feature extraction.

[0093] Furthermore, the deep learning network predicts faults of industrial equipment based on the two-dimensional image. In this embodiment, the generated two-dimensional image GAF is input into the deep learning network for feature extraction to complete fault prediction.

[0094] Specifically, the deep learning network includes a Swin Transformer model and a two-dimensional convolutional network based on a global attention mechanism. By setting up the deep learning network, it can effectively prevent overfitting on a small-scale dataset while maintaining efficient feature extraction capabilities.

[0095] The Swin Transformer model extracts local features of a two-dimensional image through a window-based multi-head self-attention (W-MSA). It should be noted that in this embodiment, the window-based multi-head self-attention (W-MSA) of the Swin Transformer model reduces the computational complexity by limiting the attention calculation to a local window, and is suitable for high-resolution input. Specifically, the input feature is set to F GAF ∈R H×W×C , where H and W represent the spatial size of the feature map, and C represents the number of channels. The window attention mechanism first divides the input features into multiple local windows, each of which is M×M in size, that is, the entire feature map will be divided into multiple non-overlapping local windows:

[0096]

[0097] Split(·,M) means dividing the feature map into windows of size M×M, where each window is represented as Among them, i represents the i-th window.

[0098] For each local window Calculate the multi-head self-attention (W-MSA) within the window. Assume that the query matrix, key matrix and value matrix in the window are Then the attention calculation is expressed as:

[0099]

[0100] Among them, Q represents the query matrix, K represents the key matrix, V represents the value matrix, T represents the transpose operation, and d k represents the dimension of the key vector and softmax represents the softmax activation function.

[0101] The attention calculation is only performed on the M×M range within the local window, rather than on the entire feature map. Therefore, for each local window of the feature map, the attention only involves the M 2 pixels without involving other windows, thus significantly reducing the computational complexity. Compared with the traditional global attention mechanism, the complexity of the window attention mechanism is O((M 2 )×C), rather than O((H×W) 2 ×C), which can greatly reduce the computational overhead when processing large-scale feature maps. Therefore, the window attention mechanism of the Swin Transformer model can effectively reduce the computational complexity while maintaining sensitivity to local details.

[0102] In this embodiment, the feature extraction layer of the Swin Transformer model includes multiple Transformer blocks, each Transformer block includes a W-MSA layer and an MLP layer, and the attention weight is calculated by the formula:

[0103]

[0104] Among them, Q represents the query matrix, K represents the key matrix, V represents the value matrix, T represents the transpose operation, dk represents the dimension of the key vector, and softmax represents the softmax activation function.

[0105] Furthermore, the SwinTransformer model extracts local features of the two-dimensional image through the window attention mechanism, and the calculation formula is:

[0106] F local =SwinTransformer(Linear(GAF i，j )+PE)

[0107] Among them, F local Represents local features, PE represents position encoding, and Linear represents a linear layer, which is used to perform linear transformation on the input feature map.

[0108] It should be noted that, in this embodiment, the depth of the feature extraction layer and the number of attention heads of the Swin Transformer model can be adjusted according to the size of the data set and the complexity of the model, and the initial settings are:

[0109] depths=(1,), num_heads=(4,)

[0110] In this embodiment, the two-dimensional convolutional network based on the global attention mechanism extracts the global features of the two-dimensional image through the GAM-Attention module. It should be noted that in this embodiment, the use of the GAM-Attention module can enhance the ability to capture global features, and the calculation formula is:

[0111] F ccnv =CNN2d(GAF i，j )

[0112] Attention GAM =σ(W2·ReLU(W1·F conv +b1)+b2)

[0113] F global =Attention GAM (F conv )

[0114] Among them, CNN2d represents a two-dimensional convolutional neural network, F conv Represents the features extracted by the convolutional layer, Attention GAM represents the global attention mechanism module, σ represents the activation function, W1 and W2 represent weight matrices, b1 and b2 represent bias terms, ReLU represents the ReLU activation function, and Fglobal represents the global feature.

[0115] The spatial and channel attention mechanisms are combined through the GAM-Attention module to enhance the feature representation capability.

[0116] In this embodiment, the setting of the convolutional network refers to the existing VGG16 network structure, and the number of convolutional layers and the number of feature channels are respectively set to:

[0117] conv_arch=((2,64), (2,128), (3,256), (3,512), (3,512))

[0118] Therefore, in this embodiment, the window attention mechanism of the SwinTransformer model is used to extract local features, and the GAM-Attention module is used to enhance the ability to capture global features. Through the combination of this dual attention mechanism, the deep learning network can take into account both local details and global structures when processing complex fault modes, thereby improving the detection accuracy of one-dimensional timing signal faults.

[0119] Furthermore, the local features and the global features are concatenated to obtain fused features, which are expressed as:

[0120] F fused =Concat(F local , F global )

[0121] Among them, F fused represents the fusion feature, and Concat(·) represents the concatenation operation in the channel dimension.

[0122] Specifically, if the feature tensor sizes output by the SwinTransformer model and the two-dimensional convolutional network based on the global attention mechanism are both [batch_size, 100, 56, 56], the size of the fused feature tensor is [batch_size, 200, 56, 56].

[0123] Through the feature fusion layer, the features extracted by local attention and global attention are effectively combined, so that the deep learning network can pay attention to the details and grasp the overall features, reducing the missed detection and false detection caused by feature loss during the detection process. This fusion method ensures the sensitivity of the deep learning network in detecting small fault features and improves the accuracy and reliability of detection.

[0124] In this embodiment, the fused features are simplified into a fixed-length feature vector through adaptive average pooling to facilitate subsequent classification processing.

[0125] Furthermore, the fusion features are reduced in dimension by an adaptive average pooling layer, and then classified by a fully connected layer to output a fault prediction result. In this embodiment, the fault prediction result is expressed as:

[0126]

[0127] in, represents the fault prediction result, GAP(·) represents the global average pooling layer, and FC(·) represents the fully connected layer.

[0128] It should be noted that in this embodiment, adaptive average pooling is used to convert the fused feature tensor into a fixed-size feature vector for subsequent classification layer processing. The fully connected layer maps the features to industrial equipment fault category labels through linear transformation and activation functions. Common activation functions include ReLU activation function or softmax activation function to achieve multi-classification tasks.

[0129] In this embodiment, Figure 4 and Figure 5 As shown in the figure, the constructed deep learning network needs to be trained. During the training process, the cross-entropy loss function (Cross-Entropy Loss) is used as the loss indicator. The model parameters are optimized through the back propagation algorithm to gradually improve the fault detection accuracy of the model. During the training process, the appropriate learning rate, batch size and training rounds are set to balance the convergence speed of the model and the training stability.

[0130] In this embodiment, the back propagation algorithm and Adam optimizer are used to optimize the deep learning network, and the learning rate is dynamically adjusted according to the convergence of the deep learning network. By adjusting the batch size and training epochs, it is ensured that the deep learning network can achieve a high fault detection accuracy within a limited training time.

[0131] like Figure 6As shown, this embodiment also provides an industrial equipment fault prediction device based on deep learning, including a signal acquisition module, an edge computing module, a signal transmission module, a signal receiving module and a server. The signal acquisition module is used to collect one-dimensional time series signals of industrial equipment. The edge computing module is used to preprocess the one-dimensional time series signal. The signal transmission module transmits the preprocessed one-dimensional time series signal to the signal receiving module through a transmission medium. The signal receiving module transmits the preprocessed one-dimensional time series signal to the server. The server converts the preprocessed one-dimensional time series signal into a two-dimensional image, and the deep learning network predicts the fault of the industrial equipment based on the two-dimensional image. The industrial equipment fault prediction device based on deep learning provided in this embodiment can implement any method of the industrial equipment fault prediction method based on deep learning, and the specific working process of the industrial equipment fault prediction device based on deep learning can refer to the corresponding process in the embodiment of the industrial equipment fault prediction method based on deep learning. The method and device provided in this embodiment can be implemented in other ways. For example, the device embodiments described above are only illustrative; for example, the division of a module is only a logical function division, and there may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual connection or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, or it can be an electrical, mechanical or other form of connection.

[0132] Therefore, in this embodiment, by constructing a deep learning network, better results can be achieved for industrial equipment fault prediction by using only lightweight detectors. Using detectors with simpler structures in the process of industrial equipment fault prediction can greatly reduce the resources required for calculation, improve the detection speed of the deep learning network and reduce the video memory occupancy required in the detection process, avoid unnecessary waste, and effectively utilize limited computing resources, thereby realizing the practical application of deep learning networks in industrial equipment fault prediction, which has strong practicality and wide applicability.

[0133] The above schematically describes the invention and its implementation methods, which is not restrictive. Without departing from the spirit or basic features of the invention, the invention can be implemented in other specific forms. What is shown in the accompanying drawings is only one of the implementation methods of the invention. The actual structure is not limited thereto, and any figure mark in the claims should not limit the claims involved. Therefore, if a person of ordinary skill in the art is inspired by it, without departing from the purpose of the invention, a structural method and an embodiment similar to the technical solution are designed without creativity, which should all belong to the protection scope of the present invention. In addition, the word "including" does not exclude other elements or steps, and the word "one" before the element does not exclude the inclusion of "multiple" elements. The multiple elements stated in the product claim can also be implemented by one element through software or hardware. The words first, second, etc. are used to indicate the name, and do not indicate any specific order.

Claims

1. A method for predicting industrial equipment failure based on deep learning, comprising the following steps: Collect one-dimensional time series signals of industrial equipment; Preprocessing one-dimensional time series signals; Convert the preprocessed one-dimensional time series signal into a two-dimensional image; Deep learning networks predict faults in industrial equipment based on 2D images.

2. The method for predicting industrial equipment failure based on deep learning according to claim 1, characterized in that: The specific steps of converting the preprocessed one-dimensional time series signal into a two-dimensional image are as follows: A one-dimensional time series signal is set, and normalization processing is performed on the one-dimensional time series signal; Map the normalized one-dimensional time series signal to a time series angle; The two-dimensional image is obtained by calculating the time series angle through the dynamic weighted Gram matrix.

3. The method for predicting industrial equipment failure based on deep learning according to claim 2, characterized in that: The normalized one-dimensional time series signal is mapped to a time series angle, and the mapping formula is: φ=arccos(x′) Among them, φ represents the timing angle, x represents the one-dimensional timing signal, arccos represents the trigonometric function, and x′ represents the normalized one-dimensional timing signal.

4. The method for predicting industrial equipment failure based on deep learning according to claim 3, characterized in that: The dynamic weighted Gram matrix includes a Gram angle and difference field matrix and a Gram angle and field matrix; the two-dimensional image is obtained by calculating the Gram angle and difference field matrix or the Gram angle and field matrix.

5. The method for predicting industrial equipment failure based on deep learning according to claim 4, characterized in that: The two-dimensional image is generated by calculating the time series angle through the Gram angle and the difference field matrix. The calculation formula is: GAF i，j =DTW i，j ×sin(φ i +φ j ) Among them, GAF i,j Represents a two-dimensional image, i represents the row index of the Gram angle and difference field matrix, j represents the column index of the Gram angle and difference field matrix, DTW i,j represents the cumulative distance matrix, φ i Indicates the angle value at the i-th moment of the timing angle, φ j Indicates the angle value at the jth moment of the timing angle; The two-dimensional image is generated by calculating the time series angle through the Gram angle and field matrix. The calculation formula is: GAF i，j =DTW i，j ×cos(φ i +φ j ) Among them, GAF i,j Represents a two-dimensional image, i represents the row index of the Gram angle and field matrix, j represents the column index of the Gram angle and field matrix, DTW i,j represents the cumulative distance matrix, φ i Indicates the angle value at the i-th moment of the timing angle, φ j Indicates the angle value at the jth moment of the timing angle.

6. The method for predicting industrial equipment failure based on deep learning according to claim 5, characterized in that: The deep learning network includes a Swin Transformer model and a two-dimensional convolutional network based on a global attention mechanism; The Swin Transformer model extracts local features of a two-dimensional image through a window attention mechanism, and the calculation formula is: F local =SwinTransformer(Linear(GAF i，j )+PE) Among them, F local Represents local features, PE represents position encoding, and Linear represents linear layer.

7. The method for predicting industrial equipment failure based on deep learning according to claim 6, characterized in that: The two-dimensional convolutional network based on the global attention mechanism extracts the global features of the two-dimensional image through the GAM-Attention module, and the calculation formula is: F conv =CNN2d(GAF i，j ) Attention GAM =σ(W2·ReLU(W1·F conv +b1)+b2) F global =Attention GAM (F conv ) Among them, CNN2d represents a two-dimensional convolutional neural network, F conv Represents the features extracted by the convolutional layer, Attention GAM represents the global attention mechanism module, σ represents the activation function, W1 and W2 represent weight matrices, b1 and b2 represent bias terms, ReLU represents the ReLU activation function, and F global Represents global features.

8. The method for predicting industrial equipment failure based on deep learning according to claim 7, characterized in that: The local features and the global features are concatenated to obtain the fused features; the fused features are expressed as: F fused =Concat(F local ,F global ) Among them, F fused represents the fusion feature, and Concat(·) represents the concatenation operation in the channel dimension.

9. The method for predicting industrial equipment failure based on deep learning according to claim 8, characterized in that: After the fusion features are reduced in dimension through the adaptive average pooling layer, they are classified through the fully connected layer and the fault prediction results are output; the fault prediction results are expressed as: in, represents the fault prediction result, GAP(·) represents the global average pooling layer, and FC(·) represents the fully connected layer.

10. An industrial equipment fault prediction device based on deep learning, characterized in that: include: Signal acquisition module, edge computing module, signal transmission module, signal receiving module and server; The signal acquisition module acquires one-dimensional time series signals of industrial equipment; The edge computing module pre-processes the one-dimensional time series signal; The signal transmitting module transmits the preprocessed one-dimensional time series signal to the signal receiving module through a transmission medium; The signal receiving module transmits the preprocessed one-dimensional time series signal to the server; The server converts the preprocessed one-dimensional time series signal into a two-dimensional image, and the deep learning network predicts faults of the industrial equipment based on the two-dimensional image.

Images