Equipment residual life prediction method based on full convolutional network

Abstract

The invention discloses an equipment residual life prediction method based on a full convolutional network, and the method comprises the following steps: firstly, attention distribution: employing a trend attention mechanism to distribute an equipment monitoring signal attention value, and obtaining a new weighted input sample; secondly, deep feature extraction: extracting deep degradation features of the weighted input sample obtained in the step S1 by using a network structure of three full convolutional layers; and finally, calculating a residual life predicted value: compressing the depth degradation characteristic dimension by using a global average pooling layer, and calculating the residual life value of the equipment through a full connection layer and a correction linear unit. According to the equipment residual life prediction method based on the full convolutional network, the trend attention mechanism is provided by considering the signal trend characteristics, the optimal attention distribution of the equipment monitoring signals can be realized at a relatively low calculation cost, the attention on important signals is enhanced, the precision of the equipment residual life prediction is improved, and the prediction accuracy of the equipment residual life is improved. And the use safety and reliability of the equipment can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of mechanical equipment, and in particular to a method for predicting the remaining life of equipment based on a fully convolutional network. Background Art

[0002] Equipment performance directly affects the economy and safety of industrial production. Accurately predicting the remaining life of equipment is of great significance for preventing major accidents, improving maintenance efficiency and reducing operating costs. Through built-in sensors and IoT technology, data on key equipment indicators such as temperature, pressure, vibration, etc. can be continuously collected, providing a rich data source for equipment remaining life prediction. Different sensor signals have different importance to the degradation state of the equipment. The attention mechanism can highlight key features and filter out useless information, thereby improving the accuracy of the prediction.

[0003] In the prior art, the remaining life prediction technology fails to consider the signal trend characteristics in designing the attention mechanism, making it impossible to achieve the optimal attention allocation of the equipment monitoring signal, which in turn leads to low remaining life prediction accuracy and lack of reference. Summary of the invention

[0004] The purpose of the present invention is to provide a method for predicting the remaining life of equipment based on a fully convolutional network, which can achieve optimal attention allocation of equipment monitoring signals at a relatively low computational cost, strengthen attention to important signals, improve the accuracy of predicting the remaining life of equipment, and help improve the safety and reliability of equipment use.

[0005] To achieve the above object, the present invention provides a method for predicting the remaining life of equipment based on a fully convolutional network, comprising the following steps:

[0006] S1. Attention allocation: Use the trend attention mechanism to allocate the attention value of the device monitoring signal to obtain a new weighted input sample;

[0007] S2, deep feature extraction: use the network structure of three fully convolutional layers to extract the deep degradation features of the weighted input samples obtained in S1;

[0008] S3. Calculation of the remaining life prediction value: Use the global average pooling layer to compress the deep degradation feature dimension, and then calculate the remaining life value of the equipment through the fully connected layer and the rectified linear unit.

[0009] Preferably, in S1, the attention value of the device monitoring signal is allocated using the trend attention mechanism, specifically:

[0010] S11, using the average pooling layer to perform a trend squeezing operation, extracting the trend characteristics of the remaining life monitoring signal, and representing each signal sample;

[0011] S12, using an excitation operation including two fully connected layers and a sigmoid activation function to calculate the importance weights of all signal samples;

[0012] S13, multiplying the input sample containing multiple sample signals by the importance weight of the signal sample through a channel multiplication function to obtain a new weighted input sample.

[0013] Preferably, in S11, the feature tensor obtained by the trend squeezing operation is Where C is the number of monitored signals, N is the length of the vector representing the trend, and the expression is:

[0014]

[0015] Where floor(x) is the floor function, x(c,i) is the i-th element of the c-th input signal, and H is the length of the time window;

[0016] Preferably, in S12, the importance weight of the signal sample is expressed as:

[0017] b=ξ(W2σ(W1g));

[0018] in, and They are the corrected linear unit, the sigmoid activation function, the parameters of the first fully connected layer, and the parameters of the second fully connected layer. r is used to adjust the complexity of TaNet.

[0019] Preferably, in S13, the trend attention mechanism inputs Convert to feature tensor The output of the trend attention mechanism is calculated as follows:

[0020] u=Mul(b,x);

[0021] Among them, Mul(b,x) is the channel multiplication function.

[0022] Preferably, in S2, the feature tensor As the input of the network structure with three fully convolutional layers, the three fully convolutional layers are given by:

[0023]

[0024] in, BN and Re lu are convolution operators, batch normalization and rectified linear units, respectively. and They are the convolution kernel parameters of the first full convolution layer, the bias of the first full convolution layer, the convolution kernel parameters of the second full convolution layer, the bias of the second full convolution layer, the convolution kernel parameters of the third full convolution layer, and the bias of the third full convolution layer.

[0025] Preferably, the numbers of three filters in the full convolution layer are set to 64, 128 and 64 respectively; the sizes of three one-dimensional convolution kernels in the full convolution layer are set to 16, 10 and 6 respectively.

[0026] Preferably, in S3, the remaining life prediction value is calculated as follows:

[0027] Receive the deep degradation features extracted from the last fully convolutional layer and calculate the remaining lifetime, including the global average pooling layer, the fully connected layer and the rectified linear unit;

[0028] The global average pooling layer compresses the depth-degraded feature dimension and obtains the i-th feature element p(i), which is expressed as:

[0029]

[0030] in, represents the jth element c3 of the i-th channel;

[0031] The remaining life of the equipment predicted by the fully connected layer and the rectified linear unit is expressed as:

[0032] RUL=σ(W r σ(W d p+b d )+b r );

[0033] in, are the weights of the first and second fully connected layers, b d and b r is the bias of the first fully connected layer and the bias of the second fully connected layer.

[0034] Preferably, the number of units in the fully connected layer after the global average pooling layer is set to 64.

[0035] Therefore, the present invention adopts the above-mentioned method for predicting the remaining life of equipment based on a fully convolutional network, and the beneficial effects are as follows:

[0036] By considering the signal trend characteristics, the present invention proposes a trend attention mechanism, which can achieve optimal attention allocation of equipment monitoring signals at a lower computational cost, strengthen attention to important signals, improve the accuracy of equipment remaining life prediction, and help improve the safety and reliability of equipment use.

[0037] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a flowchart of a first embodiment of a method for predicting the remaining life of equipment based on a fully convolutional network according to the present invention;

[0039] Figure 2 It is a schematic diagram of the cumulative prediction error of the remaining life of the FD1 device in Embodiment 2 of a method for predicting the remaining life of the device based on a fully convolutional network of the present invention;

[0040] Figure 3 It is a schematic diagram of the cumulative prediction error of the remaining life of the FD2 equipment in Example 2 of the equipment remaining life prediction method based on a fully convolutional network of the present invention. DETAILED DESCRIPTION

[0041] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0042] Unless otherwise defined, technical or scientific terms used in the present invention shall have the common meanings understood by one having ordinary skills in the field to which the present invention belongs.

[0043] Embodiment 1

[0044] like Figure 1 As shown, a method for predicting the remaining life of equipment based on a fully convolutional network includes the following steps:

[0045] S1. Use the trend attention mechanism to allocate the attention value of the device monitoring signal. Convert to feature tensor Among them, C is the number of monitoring signals and H is the length of the time window, as follows:

[0046] S11. Use the average pooling layer to perform trend squeezing operation to extract the trend characteristics of the remaining life monitoring signal to effectively represent each signal sample.

[0047] Feature tensor obtained by trend squeezing operation The expression is:

[0048]

[0049] Where floor(x) is a floor function that returns the largest integer not greater than the input x. x(c,i) refers to the i-th element of the c-th input signal. N is the length of the vector representing the trend. The segmented average obtained by the trend squeezing operation can effectively characterize the trend information of the signal sample.

[0050] S12, an excitation operation including two fully connected layers and a sigmoid activation function is used to calculate the importance weights of all signal samples, expressed as:

[0051] b=ξ(W2σ(W1g));

[0052] in, and Refers to the parameters of the corrected linear unit, the sigmoid activation function, the first fully connected layer, and the second fully connected layer. r is used to adjust the complexity of TaNet.

[0053] S13, multiply the input sample containing multiple sample signals by the importance weight of the signal sample through the channel multiplication function to obtain a new weighted input sample. That is, the trend attention mechanism is used to multiply the input Convert to feature tensor The output of the trend attention mechanism is calculated as follows:

[0054] u=Mul(b,x);

[0055] Among them, Mul(b,x) is the channel multiplication function.

[0056] The new weighted input samples are input into the network structure of S2, in this way the representation of signal importance in the network is achieved.

[0057] S2. Deep feature extraction: The feature tensor As the input of the network structure with three fully convolutional layers, the three fully convolutional layers are given by:

[0058]

[0059] in, BN and Relu are convolution operator, batch normalization and rectified linear unit respectively. and They are the convolution kernel parameters of the first full convolution layer, the bias of the first full convolution layer, the convolution kernel parameters of the second full convolution layer, the bias of the second full convolution layer, the convolution kernel parameters of the third full convolution layer, and the bias of the third full convolution layer.

[0060] Batch normalization can control the effective range of the values ​​of each layer and prevent overfitting. Therefore, in this embodiment, no dropout and maxpooling operations are added to prevent overfitting. The number of three filters in the convolution layer is set to 64, 128, and 64, respectively. The sizes of the three one-dimensional convolution kernels are set to 16, 10, and 6, respectively. The number of units in the fully connected layer after global average pooling is set to 64.

[0061] S3, remaining life prediction value calculation: receive the deep degradation features extracted from the last full convolutional layer, use the global average pooling layer to compress the deep degradation feature dimensions, and then calculate the remaining life value of the device through the fully connected layer and the rectified linear unit.

[0062] The global average pooling layer obtains the i-th feature element p(i), and the predicted remaining life of the device can be mathematically described as:

[0063]

[0064] RUL=σ(W r σ(W d p+b d )+b r );

[0065] in, The j-th element c3 of the i-th channel is obtained through the full convolutional layer. are the weights of the first and second fully connected layers. b d and b r is the bias of the first fully connected layer and the bias of the second fully connected layer.

[0066] Embodiment 2

[0067] The prediction method of the first embodiment is adopted, based on the spyder tool, using the NASA engine data set, and dividing the data set into a training set and a test set.

[0068] Experimental Testing

[0069] The prediction method of Example 2, the method without considering the degradation trend characteristics, and the method without the attention mechanism are used to analyze the cumulative prediction error of the remaining life of the equipment, and the error comparison of several sets of data sets is obtained. The cumulative prediction errors of the engines corresponding to the data sets FD1 and FD2 are respectively as follows: Figure 2 and Figure 3 shown.

[0070] As can be seen from the figure, the prediction method of Example 2 is effective in predicting the remaining life of the equipment and has a high prediction accuracy over the entire life cycle.

[0071] Therefore, the present invention adopts the above-mentioned equipment remaining life prediction method based on a fully convolutional network, and proposes a trend attention mechanism by considering the signal trend characteristics. It can achieve the optimal attention allocation of equipment monitoring signals at a lower computational cost, strengthen the attention to important signals, and improve the accuracy of equipment remaining life prediction, which is conducive to improving the safety and reliability of equipment use.

[0072] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for predicting the remaining life of equipment based on a fully convolutional network, characterized in that: The following steps are involved: S1. Attention allocation: Use the trend attention mechanism to allocate the attention value of the device monitoring signal to obtain a new weighted input sample; S2, deep feature extraction: use the network structure of three fully convolutional layers to extract the deep degradation features of the weighted input samples obtained in S1; S3. Calculation of the remaining life prediction value: Use the global average pooling layer to compress the deep degradation feature dimension, and then calculate the remaining life value of the equipment through the fully connected layer and the rectified linear unit.

2. According to claim 1, a method for predicting the remaining life of equipment based on a fully convolutional network is characterized in that: In S1, the trend attention mechanism is used to allocate the attention value of the device monitoring signal, specifically: S11, using the average pooling layer to perform a trend squeezing operation, extracting the trend characteristics of the remaining life monitoring signal, and representing each signal sample; S12, using an excitation operation including two fully connected layers and a sigmoid activation function to calculate the importance weights of all signal samples; S13, multiplying the input sample containing multiple sample signals by the importance weight of the signal sample through a channel multiplication function to obtain a new weighted input sample.

3. According to claim 2, a method for predicting the remaining life of equipment based on a fully convolutional network is characterized in that: In S11, the feature tensor obtained by the trend squeezing operation Where C is the number of monitored signals, N is the length of the vector representing the trend, and the expression is: Where floor(x) is the floor function, x(c,i) is the i-th element of the c-th input signal, and H is the length of the time window.

4. According to claim 3, a method for predicting the remaining life of equipment based on a fully convolutional network is characterized in that: In S12, the importance weight of the signal sample is expressed as: b=ξ(W2σ(W1g)); in, and They are the corrected linear unit, the sigmoid activation function, the parameters of the first fully connected layer, and the parameters of the second fully connected layer. r is used to adjust the complexity of TaNet.

5. According to claim 4, a method for predicting the remaining life of equipment based on a fully convolutional network is characterized in that: In S13, the trend attention mechanism will input Convert to feature tensor The output of the trend attention mechanism is calculated as follows: u=Mul(b,x); Among them, Mul(b,x) is the channel multiplication function.

6. The method for predicting the remaining life of equipment based on a fully convolutional network according to claim 5, characterized in that: In S2, the feature tensor As the input of the network structure with three fully convolutional layers, the three fully convolutional layers are given by: in, BN and Relu are convolution operators, batch normalization and rectified linear units, respectively. and They are the convolution kernel parameters of the first full convolution layer, the bias of the first full convolution layer, the convolution kernel parameters of the second full convolution layer, the bias of the second full convolution layer, the convolution kernel parameters of the third full convolution layer, and the bias of the third full convolution layer.

7. The method for predicting the remaining life of equipment based on a fully convolutional network according to claim 6, characterized in that: The number of filters in the full convolution layer is set to 64, 128, and 64, respectively; the sizes of the three one-dimensional convolution kernels in the full convolution layer are set to 16, 10, and 6, respectively.

8. The method for predicting the remaining life of equipment based on a fully convolutional network according to claim 7, characterized in that: In S3, the remaining life prediction value is calculated as follows: Receive the deep degradation features extracted from the last fully convolutional layer and calculate the remaining lifetime, including the global average pooling layer, the fully connected layer and the rectified linear unit; The global average pooling layer compresses the depth-degraded feature dimension and obtains the i-th feature element p(i), which is expressed as: in, represents the jth element c3 of the i-th channel; The remaining life of the equipment predicted by the fully connected layer and the rectified linear unit is expressed as: RUL=σ(W r σ(W d p+b d )+b r ); in, are the weights of the first and second fully connected layers, b d and b r is the bias of the first fully connected layer and the bias of the second fully connected layer.

9. The method for predicting the remaining life of equipment based on a fully convolutional network according to claim 8, characterized in that: The number of units in the fully connected layer after the global average pooling layer is set to 64.

Images