Sealing gasket leakage rate prediction method based on LSTM-CAM model

Abstract

The invention discloses a sealing gasket leakage rate prediction method based on an LSTM-CAM model, and belongs to the technical field of gasket sealing and machine learning crossing. The method comprises the following steps: S1, acquiring gasket load and leakage rate data of gaskets with different degradation degrees in a test process; dividing a training set, a verification set and a test set according to a time sequence by taking the leakage rate as label data; s2, constructing an LSTM-CAM model, and training, testing and verifying the LSTM-CAM model through the training set, the test set and the verification set to obtain a prediction model; and S3, performing leakage rate prediction on an input to-be-predicted sample by using the prediction model to obtain a leakage rate prediction result. The method is suitable for predicting the performance degradation of the sealing gasket in the chlor-alkali chemical device in the concentrated hydrochloric acid corrosion environment, and high-precision prediction is achieved through the processes of time sequence feature extraction, nonlinear activation, feature integration and result prediction.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of gasket sealing and machine learning, specifically relating to a method for predicting the leakage rate of sealing gaskets based on an LSTM-CAM model. Background Technology

[0002] In chlor-alkali chemical production, gaskets are exposed to corrosive media such as acids and alkalis for extended periods, and their performance deterioration can easily lead to seal failure and leakage risks. Traditional gasket aging assessment methods mostly rely on empirical models (such as exponential models), which lack sufficient prediction accuracy under small sample conditions. Support Vector Regression (SVR) models, on the other hand, are prone to getting trapped in local optima because their predictive performance is highly dependent on kernel function and hyperparameter settings. Therefore, there is an urgent need for a novel predictive model that can efficiently learn nonlinear mapping relationships under small sample conditions and possesses strong generalization ability, in order to achieve accurate modeling and prediction of the performance degradation trend of sealing materials. Summary of the Invention

[0003] To address the gasket sealing failure caused by performance degradation in harsh media environments such as acid and alkali corrosion in chlor-alkali plants, this invention proposes a Long Short-Term Memory (LSTM) network model combining Channel Attention Mechanism (CAM) based on test data of gasket leakage rate. This model constructs a mapping relationship between degradation time and gasket sealing performance degradation, establishing a gasket leakage rate performance degradation prediction model, thus laying the foundation for degradation prediction technology of typical sealing elements in chlor-alkali plants.

[0004] To achieve the above objectives, the present invention provides the following solution: a method for predicting the leakage rate of sealing gaskets based on an LSTM-CAM model, comprising the following steps:

[0005] S1. Collect gasket load and leakage rate data of gaskets with different degrees of degradation during the test; and divide the data into training set, validation set and test set according to time series, using leakage rate as the label data;

[0006] S2. Construct an LSTM-CAM model, and train, test, and validate the LSTM-CAM model using the training set, the test set, and the validation set to obtain a prediction model;

[0007] S3. Use the prediction model to predict the leakage rate of the input sample to be predicted, and obtain the leakage rate prediction result.

[0008] More preferably, the input to the prediction model is aging time and gasket load, and the output is the leakage rate prediction result.

[0009] More preferably, the prediction model includes: an input layer, an LSTM layer, an attention mechanism layer, a Sigmoid layer, a fully connected layer, and an output layer;

[0010] The input layer is used to input 2D data;

[0011] The LSTM layer uses lstmlayer, contains 10 hidden units, and has an output dimension of 10;

[0012] The attention mechanism layer is used to generate attention weights between 0 and 1;

[0013] The Sigmoid layer is used to map the output of the attention mechanism layer to the range (0, 1);

[0014] The fully connected layer is used to transform the output of the Sigmoid layer into the final prediction result and output a feature vector of fixed size;

[0015] The output layer is used to generate continuous predicted values ​​from the feature vector using a linear transformation, thereby obtaining the leakage rate prediction result.

[0016] More preferably, the LSTM layer includes three gating mechanisms: a forget gate, an input gate, and an output gate;

[0017] The forget gate is used to output a value between 0 and 1 based on the current input and the hidden state of the previous time step:

[0018] f t =σ(W f ·[h t-1 ,x t ]+b f );

[0019] In the formula, σ is the sigmoid activation function, and W f h represents the weight of the forget gate. t-1 x is the hidden state of the previous time step. t For the current input, b f It is the bias term of the forget gate;

[0020] The input gate consists of two parts: a sigmoid gate, used to determine information updates; and a tanh gate, used to generate candidate memory contents.

[0021] i t =σ(W i ·[h t-1 ,x t ]+b i );

[0022]

[0023] In the formula, i t The output of the input gate, W represents the current memory state. i and W c It is the weight of the input gate and the candidate memory content, b i and b c It is W i and W c The bias term;

[0024] Update memory units: Based on the input gate's decision and the forget gate's output, update the current memory state:

[0025]

[0026] In the formula, C t C represents the current memory state. t-1 It is the memory state of the previous time step, f t and i t It is the output of the forget gate and the input gate. Indicates candidate memory states;

[0027] The output gate is used to determine the output based on the current memory state and the current input:

[0028] O t =σ(W o ·[h t-1 ,x t ]+b o );

[0029] The final output is based on the product of the updated memory units after passing through the tanh activation function:

[0030] h t =O t ·tanh(C t );

[0031] In the formula, O t It is the output of the output gate, h t It is the hidden state at the current time step, W o It is the weight of the output gate, b o It is the bias term of the output gate, tanh(C t ) is the value after tanh activation of the current memory cell state.

[0032] More preferably, the attention mechanism layer includes three steps: compression, activation, and multiplication;

[0033] The compression is used to aggregate global information from the input feature map through global average pooling, compressing spatial information into a channel-level representation.

[0034] The incentive is used to generate channel attention weights through the fully connected layer, assigning a weight value to each channel;

[0035] The multiplication is used to multiply the obtained channel attention weights with the input feature map channel by channel, thereby adjusting the feature response of each channel.

[0036] More preferably, the compression step includes:

[0037]

[0038] In the formula, U C (i,j) represents the 2D matrix of the c-th channel, where i = 1, 2…H, j = 1, 2…W, Z C This is the feature output corresponding to the compression operation; F sq (U c ) represents the computational channel feature map U C The average pixel value at all spatial locations (i, j).

[0039] More preferably, the excitation step includes:

[0040] S C =F ex (Z c ,O)=σ(O2δ(O1,Z c ));

[0041] In the formula, δ represents the ReLU activation function, O1 and O2 are the weights of the two fully connected layers, and S C This is the output after activation; Fex(Z) c O) represents the channel statistics Z obtained from the compression operation. C Transformed into channel-adaptive weights.

[0042] More preferably, the multiplication steps include:

[0043]

[0044] In the formula, This represents the enhanced features after processing by the channel attention mechanism; F scale (U C F C ) represents the channel attention weight S C Channel-by-channel scaling process applied to the original feature map Uc.

[0045] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0046] This invention significantly improves the accuracy and reliability of leak rate prediction for rubber gaskets under long-term aging conditions through LSTM-CAM technology. The time-series-based modeling method effectively captures the dynamic characteristics of leak rate changes with degradation time and load. Compared to traditional regression models and standard machine learning methods, the LSTM-CAM prediction model of this invention exhibits significant advantages: the model is particularly suitable for modeling small-sample time-series data; it can effectively handle the nonlinear temporal characteristics in leak rate prediction, improving the accuracy of long-term aging prediction and providing reliable data support for life assessment and maintenance decisions of sealing components in the chlor-alkali industry. Attached Figure Description

[0047] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 This is a flowchart illustrating a gasket leakage rate prediction method based on an LSTM-CAM model provided in an embodiment of the present invention.

[0049] Figure 2 This is an overall architecture diagram of the prediction model of a gasket leakage rate prediction method based on the LSTM-CAM model provided in an embodiment of the present invention;

[0050] Figure 3 This is a schematic diagram of the channel attention mechanism in an embodiment of the present invention. Detailed Implementation

[0051] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0053] Example 1:

[0054] like Figure 1 As shown, this embodiment provides a method for predicting the leakage rate of sealing gaskets based on an LSTM-CAM model, including the following steps:

[0055] S1. Collect gasket load and leakage rate data of gaskets with different degrees of degradation during the test; and divide the data into training set, validation set and test set according to time series, using leakage rate as the label data.

[0056] Specifically, sensors were used to collect gasket load and leakage rate data for gaskets with different levels of degradation during the testing process. The leakage rate was recorded as tagged data, outliers were removed, and missing values ​​were filled. The dataset was divided into training, validation, and test sets according to time sequence. This chronological division helps prevent future information leakage. The proportions were: training set (70%), validation set (15%), and test set (15%).

[0057] S2. Construct an LSTM-CAM model, and train, test, and validate the LSTM-CAM model using training, testing, and validation sets to obtain the prediction model.

[0058] In this embodiment, the LSTM-CAM model is built as follows: The input layer receives the raw time-series data and passes it to subsequent layers of the network. The LSTM layer processes the time-series data and outputs a 10-dimensional feature vector. The attention mechanism layer generates attention weights through global average pooling and a fully connected layer to adjust the model's attention to different parts of the data. The Sigmoid layer calculates and normalizes the attention weights, mapping the output of the previous layer to the (0,1) interval, providing support for subsequent weighted summation. The fully connected layer converts the weighted output into a fixed-size feature vector to prepare for the regression task. Finally, the output layer generates continuous predicted values ​​through linear transformation to complete the regression prediction task.

[0059] During training, the training loss curve reflects the trend of the model's loss value with the number of iterations, serving as an important tool for evaluating whether the training process is normal and whether the model has converged. The loss gradually decreases with the number of iterations, indicating that the LSTM-CAM model is learning the nonlinear mapping relationship between the input and output. The root mean square error (RMSE) is used to measure the error between the model's predicted output and the experimental data. In the parameter settings, the Adam optimizer is used to improve training efficiency, with 1500 iterations and a learning rate of 0.01. At the 600th iteration, the RMSE gradually converges and tends to 0.02, indicating that the model quickly learns the relationship between the input and output. Further evaluation of the LSTM-CAM model's prediction accuracy and generalization ability is conducted. Early stopping is used to detect the validation set loss.

[0060] Model Evaluation: The industrial evaluation index is accuracy 1 - MAPE, where the formula for Mean Percentage Error (MAPE) is:

[0061]

[0062] In the formula, n represents the number of samples in the experiment; y i Indicates the actual value; This represents the predicted value.

[0063] S3. Use a prediction model to predict the leakage rate of the input sample to be predicted, and obtain the leakage rate prediction result.

[0064] A further implementation involves taking aging time and gasket load as inputs to the prediction model, converting the data into a time-series signal through normalization, and outputting the leakage rate prediction result.

[0065] like Figure 2 As shown, the prediction model includes: an input layer, an LSTM layer, an attention mechanism layer, a Sigmoid layer, a fully connected layer, and an output layer. The input layer is used to input 2D data; the LSTM layer uses multiple LSTM-Cell units, containing 10 hidden units, and has an output dimension of 10; the attention mechanism layer is used to generate attention weights between 0 and 1; the Sigmoid layer is used to map the output of the attention mechanism layer to the range (0, 1); the fully connected layer is used to transform the output of the Sigmoid layer into the final prediction result and output a feature vector of fixed size; the output layer is used to use a linear transformation to generate continuous prediction values ​​from the feature vector, thus obtaining the leakage rate prediction result.

[0066] The structural parameters of the prediction model are as follows: The input layer uses a sequence layer to define the input dimension. The input is 2-dimensional, indicating that there are two feature inputs (aging days and stress level), forming [T] i P i The [T] combinations correspond to each time point in the time series. The LSTM layer uses multiple LSTM-Cell units to capture the input [T] i P i ] and corresponding leakage rate L i The layer contains 10 hidden units, has an output dimension of 10, and 520 learnable parameters. The attention mechanism layer CAM introduces attention weights through a fully connected layer with 10 neurons, focusing on the dependencies of L. i The most influential input features [T] i P i The output layer has a dimension of 10. The Sigmoid layer is used to calculate attention weights, mapping the output of the attention mechanism layer to the (0,1) range. The fully connected layer transforms the weighted output of the previous layer into the final prediction result, outputting a fixed-size feature vector with a dimension of 1 and 11 learnable parameters. Finally, a linear transformation is applied through the output layer based on the fully connected layer to output the predicted leakage rate L for 30 days and 90 days, respectively, with an output dimension of 1.

[0067] The LSTM-CAM model combines the powerful temporal modeling capabilities of LSTM with the characteristics of the channel attention mechanism, which can effectively assign weights to different input features, thereby improving prediction accuracy.

[0068] A further implementation lies in the fact that LSTM is a special type of Recurrent Neural Network (RNN) with long-term memory capabilities, effectively capturing the temporal dependencies in sequential data, thus avoiding the vanishing and exploding gradient problems that traditional RNNs are prone to in long sequences. It controls the flow of information through three gating mechanisms (forget gate, input gate, and output gate), controlling the information flow at each time step to determine which information should be remembered, which should be forgotten, and how to generate the output.

[0069] The forget gate is used to output a value between 0 and 1, based on the current input and the hidden state of the previous time step, indicating how much historical information is forgotten. 0 represents complete forgetting, and 1 represents complete retention.

[0070] f t =σ(W f ·[h t-1 ,x t ]+b f );

[0071] In the formula, σ is the sigmoid activation function, and W f h represents the weight of the forget gate. t-1 x is the hidden state of the previous time step. t For the current input, b f It is the bias term of the forget gate.

[0072] The input gate determines the impact of the current input information on the memory unit, consisting of two parts: the sigmoid gate, which determines the information update; and the tanh gate, which generates candidate memory contents.

[0073] i t =σ(W i ·[h t-1 ,x t ]+b i );

[0074]

[0075] In the formula, i t The output of the input gate, W represents the current memory state. i and W c It is the weight of the input gate and the candidate memory content, b i and bc It is W i and W c The bias term.

[0076] Update memory units: Based on the input gate's decision and the forget gate's output, update the current memory state:

[0077]

[0078] In the formula, C t C represents the current memory state. t-1 It is the memory state of the previous time step, f t and i t It is the output of the forget gate and the input gate. Indicates candidate memory states;

[0079] The output gate determines the final output. Based on the current memory state and the current input, it decides which information will ultimately be output.

[0080] O t =σ(W o ·[h t-1 ,x t ]+b o ).

[0081] The final output is based on the product of the updated memory units after passing through the tanh activation function:

[0082] h t =O t ·tanh(C t );

[0083] In the formula, O t It is the output of the output gate, h t It is the hidden state at the current time step, W o It is the weight of the output gate, b o It is the bias term of the output gate, tanh(C t ) is the value after tanh activation of the current memory cell state.

[0084] To help the model more accurately weight the output of LSTM and improve prediction performance, a channel attention mechanism is introduced. The core idea is that different channels contribute differently to the model's predictions, and the model should assign higher weights to more important channels to highlight key features and suppress irrelevant features.

[0085] Further implementation lies in, such as Figure 3 As shown, the input feature map X has a size of H×W×C1, and after passing through F... trConvolutional processing yields an H×W×C2 feature layer. The attention mechanism layer (CAM) consists of three steps: compression, activation, and multiplication. Compression involves aggregating global information from the input feature map through global average pooling. Its purpose is to compress spatial information into a 1×1×C2 channel-level representation. Activation involves generating a 1×1×C2 channel attention weight through a fully connected layer. The goal is to assign a weight value to each channel, mapping feature information to weights between 0 and 1, and indicating the channel's importance in the feature map. Multiplication involves multiplying the obtained channel attention weights with the input feature map channel by channel, thereby adjusting the feature response of each channel and outputting an H×W×C2 feature layer with attention weights.

[0086] Specifically, the compression steps include:

[0087]

[0088] In the formula, U C (i,j) represents the 2D matrix of the c-th channel, where i = 1, 2…H, j = 1, 2…W, Z C This is the feature output corresponding to the compression operation; F sq (U c ) represents the computational channel feature map U C The average pixel value at all spatial locations (i, j).

[0089] The motivational steps include:

[0090] S C =F ex (Z c ,O)=σ(O2δ(O1,Z c ));

[0091] In the formula, δ represents the ReLU activation function, O1 and O2 are the weights of the two fully connected layers, and S C This is the output after activation; Fex(Z) c O) represents the channel statistics Z obtained from the compression operation. C Transformed into channel-adaptive weights.

[0092] The steps of multiplication include:

[0093]

[0094] In the formula, This represents the enhanced features after processing by the channel attention mechanism; F scale (U C F C ) represents the channel attention weight S C Channel-by-channel scaling process applied to the original feature map Uc.

[0095] In this embodiment, CAM aggregates global information for each channel using global average pooling, then generates channel weights through a fully connected layer, and finally generates weight values ​​between 0 and 1 using a sigmoid activation function to represent the importance of that time step.

[0096] Attention Weighes=σ(Fully Connected Output);

[0097] In the formula, Fully Connected Output represents the original value of the output of the fully connected layer.

[0098] A fully connected layer is a structure in a Convolutional Neural Network (CNN) located at the end of the network. Each node in it is interconnected with all nodes in the previous layer, integrating the features extracted from the previous layer and mapping these features to the sample label space. In this embodiment, the output features are weighted to generate the model's output for regression prediction. The calculation method is as follows:

[0099] y t =W output ·h t +b output ;

[0100] In the formula, y t This is the final prediction result, W output The weights of the output layer, h t Combining the hidden state of CAM, b output It is the bias term of the output layer.

[0101] Example 2:

[0102] This embodiment provides an example of a gasket leakage rate prediction method based on an LSTM-CAM model, which is used to further illustrate the method proposed in this invention.

[0103] Data Acquisition and Preprocessing: Air tightness tests were conducted on EPDM rubber gaskets in a chlor-alkali chemical hydrochloric acid environment at degradation times of 0, 13, 21, 30, 45, 60, and 90 days. The test gas medium was helium, and the working pressure was 2 MPa. Six different loads (7, 10, 15, 20, 25, and 30 MPa) were sequentially applied to the gaskets under test in the leakage rate platform system. The leakage rate of the gaskets gradually increased over time. A helium gas spectrometer was used to measure the helium leakage rate during the air tightness test, testing the leakage rate changes of each group of gaskets under different loads.

[0104] Based on the above experimental data, six sets of data (corresponding to EPDM gasket leakage rate data with aging times of 0 days, 7 days, 13 days, 21 days, 45 days, and 60 days) were selected from eight sets of experimental data as training samples. An extended 25-day key aging node was used to enhance the model's ability to capture nonlinear degradation patterns. Experimental data of 30 days and 90 days were used as test samples. Input and output data were constructed, where the input consisted of two variables: aging time and load level. Specifically, taking sample 1 as an example, the training sample X was [0, 7], [0, 10], [0, 15], [0, 20], [0, 25], and [0, 30]. The output was the leakage rate of the EPDM gasket. Data normalization was performed, mapping the input and output data to the [0, 1] interval to convert the input data into a time-series signal. The normalized data was converted into a cell array format to meet the input requirements of the LSTM network. Finally, after data preprocessing, the data was input into the LSTM-CAM network for training and prediction.

[0105] In the prediction section, predictions were made for the 30-day test set and the 90-day out-of-sample data, and the accuracy was calculated. Data points from the training, test, and prediction sets were plotted using a 3D scatter plot, with different colors used to distinguish between actual and predicted values. Predictions were also made for grid data, and surfaces were plotted.

[0106] Experimental results show that, compared with the basic LSTM model, the LSTM-CAM model of this invention significantly improves the consistency between the 90-day predicted leakage rate and the measured value under the same training set conditions. At the same time, the short-term prediction reliability is verified by 30 days of measured data.

[0107] Six training samples were trained using LSTM and LSTM-CAM models respectively, and the optimal weight model was selected for testing and prediction. The leakage rate prediction results of different models are shown in Table 1.

[0108] Table 1

[0109]

[0110] As shown in Table 1, the combination of LSTM and channel attention mechanism improves the accuracy of leakage rate prediction for both 30 and 90 days, indicating that LSTM-CAM can effectively improve the model's prediction accuracy and performance.

[0111] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for predicting the leakage rate of sealing gaskets based on an LSTM-CAM model, characterized in that, Includes the following steps: S1. Collect gasket load and leakage rate data of gaskets with different degrees of degradation during the test; and divide the data into training set, validation set and test set according to time series, using leakage rate as the label data; S2. Construct an LSTM-CAM model, and train, test, and validate the LSTM-CAM model using the training set, the test set, and the validation set to obtain a prediction model; S3. Use the prediction model to predict the leakage rate of the input sample to be predicted, and obtain the leakage rate prediction result.

2. The method for predicting the leakage rate of sealing gaskets based on an LSTM-CAM model according to claim 1, characterized in that, The inputs to the prediction model are aging time and gasket load, and the output is the leakage rate prediction result.

3. The method for predicting the leakage rate of sealing gaskets based on an LSTM-CAM model according to claim 1, characterized in that, The prediction model includes: an input layer, an LSTM layer, an attention mechanism layer, a Sigmoid layer, a fully connected layer, and an output layer; The input layer is used to input 2D data; The LSTM layer uses lstmlayer, contains 10 hidden units, and has an output dimension of 10; The attention mechanism layer is used to generate attention weights between 0 and 1; The Sigmoid layer is used to map the output of the attention mechanism layer to the range (0, 1); The fully connected layer is used to transform the output of the Sigmoid layer into the final prediction result and output a feature vector of fixed size; The output layer is used to generate continuous predicted values ​​from the feature vector using a linear transformation, thereby obtaining the leakage rate prediction result.

4. The method for predicting the leakage rate of sealing gaskets based on an LSTM-CAM model according to claim 3, characterized in that, The LSTM layer includes three gating mechanisms: forget gate, input gate, and output gate; The forget gate is used to output a value between 0 and 1 based on the current input and the hidden state of the previous time step: f t =σ(W f ·[h t-1 ,x t ]+b f ); In the formula, σ is the sigmoid activation function, and W f h represents the weight of the forget gate. t-1 x is the hidden state of the previous time step. t For the current input, b f It is the bias term of the forget gate; The input gate consists of two parts: a sigmoid gate, used to determine information updates; and a tanh gate, used to generate candidate memory contents. i t =σ(W i ·[h t-1 ,x t ]+b i ); In the formula, i t The output of the input gate, W represents the current memory state. i and W c It is the weight of the input gate and the candidate memory content, b i and b c It is W i and W c The bias term; Update memory units: Based on the input gate's decision and the forget gate's output, update the current memory state: In the formula, C t C represents the current memory state. t-1 It is the memory state of the previous time step, f t and i t It is the output of the forget gate and the input gate. Indicates candidate memory states; The output gate is used to determine the output based on the current memory state and the current input: The t =σ(W o ·[h t-1 ,x t ]+b o ); The final output is based on the product of the updated memory units after passing through the tanh activation function: h t =O t ·tanh(C t ); In the formula, O t It is the output of the output gate, h t It is the hidden state at the current time step, W o It is the weight of the output gate, b o It is the bias term of the output gate, tanh(C t ) is the value after tanh activation of the current memory cell state.

5. The method for predicting the leakage rate of sealing gaskets based on an LSTM-CAM model according to claim 4, characterized in that, The attention mechanism layer includes three steps: compression, activation, and multiplication. The compression is used to aggregate global information from the input feature map through global average pooling, compressing spatial information into a channel-level representation. The incentive is used to generate channel attention weights through the fully connected layer, assigning a weight value to each channel; The multiplication is used to multiply the obtained channel attention weights with the input feature map channel by channel, thereby adjusting the feature response of each channel.

6. The method for predicting the leakage rate of a sealing gasket based on an LSTM-CAM model according to claim 5, characterized in that, The compression steps include: In the formula, U C (i,j) represents the 2D matrix of the c-th channel, where i = 1, 2…H, j = 1, 2…W, Z C This is the feature output corresponding to the compression operation; F sq (U c ) represents the computational channel feature map U C The average pixel value at all spatial locations (i, j).

7. The method for predicting the leakage rate of a sealing gasket based on an LSTM-CAM model according to claim 5, characterized in that, The incentive steps include: S C =F ex (Z c ,O)=σ(O2δ(O1,Z c )); In the formula, δ represents the ReLU activation function, O1 and O2 are the weights of the two fully connected layers, and S C This is the output after activation; Fex(Z) c O) represents the channel statistics Z obtained from the compression operation. C Transformed into channel-adaptive weights.

8. The method for predicting the leakage rate of sealing gaskets based on an LSTM-CAM model according to claim 5, characterized in that, The steps of the multiplication include: In the formula, This represents the enhanced features after processing by the channel attention mechanism; F scale (U C F C ) represents the channel attention weight S C Channel-by-channel scaling process applied to the original feature map Uc.

Images