Seabed double-path output pressure prediction method and system

Abstract

The invention belongs to the technical field of submarine double-path output pressure measurement, and particularly relates to a submarine double-path output pressure prediction method and system, which solve the problem of precise detection of submarine pressure, improve the precision of submarine pressure detection, reduce the test cost and improve the pressure test efficiency by using the advantages of deep learning. A proposed jump attention mechanism is improved on the basis of a traditional multi-head self-attention mechanism, a parameterization function is designed to approximate the output of the multi-head self-attention mechanism, and the parameterization function is used for replacing the multi-head self-attention mechanism in part of encoder layers, so that the model complexity is greatly reduced, and the algorithm complexity is reduced. And the consumption of computing resources is reduced.

Description

Technical Field

[0001] The present application belongs to the technical field of submarine dual-path output pressure measurement, and specifically relates to a submarine dual-path output pressure prediction method and system. Background Art

[0002] Marine environmental detection is a fundamental work to achieve sustainable development of the ocean, protect the marine environment, promote the marine economy and ensure national security. Seabed pressure measurement is an important field in marine scientific research and engineering applications. It involves marine geological structure, resource exploration, environmental monitoring and other aspects. With the development of marine technology, the accuracy and reliability of seabed pressure measurement are increasingly required. At present, seabed pressure sensors are usually required to have high measurement accuracy, good stability and the ability to work in harsh environments for a long time. In order to meet these requirements, pressure sensor data acquisition and calibration methods and technologies are particularly important. In the actual application of marine pressure sensors in marine detection, they are often affected by the complex environment inside the sensor and outside the ocean. These factors will cause the performance of the pressure sensor to fluctuate, thereby increasing the difficulty of data acquisition and calibration. Summary of the invention

[0003] Based on the above problems, the present invention proposes a data acquisition and calibration method for a submarine dual-output pressure sensor based on a Transformer model, which uses the advantages of deep learning to solve the problem of accurate submarine pressure detection, improve the accuracy of submarine pressure detection, reduce the cost of testing, and improve the efficiency of pressure testing. Its technical solution is:

[0004] A method for predicting seabed dual-path output pressure comprises the following steps:

[0005] S1. Obtain the original data of the dual-channel output of the seabed, including voltage, pressure and temperature, and use the voltage value V and temperature T as the training set, and the pressure as the training label;

[0006] S2. Extract features of the original data using an adaptive feature extractor to obtain a feature set F;

[0007] S3. Input the feature set F into the improved Transformer model for training;

[0008] S4. Use the calibrated Transformer model to predict the dual-path output pressure at the seafloor.

[0009] Preferably, the adaptive feature extractor comprises a polynomial feature module and an adaptive feature extraction module;

[0010] The polynomial feature module expands the feature set by constructing polynomial features of the input and captures the nonlinear relationship between the inputs. The module generates polynomial combinations of these features, including their powers and interactions, by inputting the original data V and T. The polynomial feature set F can be expressed as:

[0011] F={1,V,T,V 2 ,T 2 ,VT}

[0012] The adaptive feature extraction module uses a small neural network to automatically learn the complex relationship between input features and sets an adaptive feature extraction function f θ , where θ represents the network parameters, the function consists of the following multi-layer neural network:

[0013] f θ (F) = σ(W2[σ(W1F)+b2])

[0014] Where W1 and W2 are weight matrices, b2 is the bias term, and σ is the activation function.

[0015] Preferably, in step S3, the improved Transformer model removes the position encoding in the traditional Transformer model, adds a jump attention mechanism, and uses an extreme learning machine to replace the decoder in the traditional Transformer model for output.

[0016] Preferably, in step S3, the skip attention mechanism alternately selects whether to skip the multi-head attention mechanism in each encoder. For the encoder layer that does not skip self-attention, the multi-head self-attention mechanism is used to calculate the output features as the input of the feedforward neural network MLP; for the encoder layer that skips the self-attention calculation, the features output by the parameterized function Φ are used as the input of the feedforward neural network MLP:

[0017] Multi-head self-attention mechanism calculates output feature X l as follows:

[0018]

[0019] X l ←MLP(X l )+X l

[0020] X l-1 Represents the output of the l-1th encoder in Transformer;

[0021] The features output by the parameterized function Φ are as follows:

[0022]

[0023] in yes , FC1 and FC2 represent two linear layers, DwC represents depth-wise separable convolution, and the ECA module performs global average pooling in the channel dimension of the input and uses point-wise convolution to enhance the cross-channel dependency of the input.

[0024] Preferably, in step S3, the extreme learning machine includes an input, a hidden layer and an output; its input is X l or Marked with X L , Where L represents the number of encoders in the model;

[0025] The extreme learning machine is expressed by the following formula:

[0026]

[0027] Where m is the number of hidden nodes, β j is the connection hidden node h j and the weight vector of the output node, g(·) represents the activation function, w j is the connection hidden node h j and the weight vector of the input node, b j Represents the hidden node h j The bias value, w j and b j is randomly generated.

[0028] Preferably, β j By solving the following linear equation:

[0029]

[0030] Y is the output value of the extreme learning machine, which is also the pressure value finally predicted by the improved Transformer model;

[0031] The above formula can also be expressed in matrix form, namely:

[0032] Hβ=Y

[0033] in:

[0034]

[0035] The goal of the extreme learning machine is to solve β to minimize the gap between Hβ and Y. The approximate solution of β can be obtained by the following formula:

[0036] β≈H + Y

[0037] Among them, H +is the Moore–Penrose generalized inverse of H.

[0038] Preferably, the mean square error MSE and the coefficient of determination R are used 2 To evaluate the training results of the proposed model, MSE and R 2 is defined as follows:

[0039]

[0040] Where N represents the number of samples, Y i represents the pressure value predicted by the model for the i-th sample, P i represents the true pressure value in the i-th sample, Represents the mean of the pressure values ​​in the sample.

[0041] Preferably, the trained model is tested using a test set, and the model is calibrated based on the test results; if the deviation between the output and the set pressure is large, the model training process is repeated until the output value of the calibrated pressure sensor is kept within 0.5% of the output value of the reference pressure sensor; if the deviation between the output and the set pressure is within a certain range and reaches an accuracy index of 0.5%, the trained model predicts the pressure.

[0042] A seabed dual-channel output pressure prediction system, comprising a data acquisition module, a data processing module and an output module;

[0043] The data acquisition module uses sensors to obtain raw data and transmits it to the computer through a communication acquisition board; the data processing module uses an adaptive feature extractor to extract features from the raw data and inputs the feature set F into the improved Transformer model for training;

[0044] The output module is used to visually output the predicted value of the dual-path output pressure of the seabed.

[0045] Preferably, the improved Transformer model removes position encoding and mask attention, and replaces layer normalization with batch normalization.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] 1. For ocean exploration, the variability and unpredictability of the ocean environment have a significant impact on experimental errors. Due to the complexity of the ocean environment, the collected raw data may require complex post-processing and correction to improve data quality, which increases the workload and complexity of data processing. The data processing efficiency and effect of traditional classical algorithms such as linear interpolation and polynomial fitting or simple deep learning models are relatively limited. The Transformer model is a deep learning model based on the self-attention mechanism that can efficiently process large-scale data sets, which is very useful for calibration processes that require a large amount of complex data for training to improve accuracy, and improves work test efficiency.

[0048] 2) An adaptive feature extractor is designed to expand the voltage and temperature features in the input data to help the Transformer model better understand the relationship between the inputs and improve the accuracy of calibration;

[0049] 3) Since Transformer is mainly designed for sequence data, and the input data in our pressure sensor calibration task is not sequence data and has permutation invariance, we simplified the traditional Transformer model, removed unnecessary position encoding and mask attention, and replaced layer normalization with batch normalization;

[0050] 4) The proposed jump attention mechanism is an improvement on the traditional multi-head self-attention mechanism. A parameterized function is designed to approximate the output of the multi-head self-attention mechanism, and the parameterized function is used to replace the multi-head self-attention mechanism in some encoder layers, which greatly reduces the model complexity and the consumption of computing resources, ensuring that the Transformer model can run successfully in the conditioning chip;

[0051] 5) Remove the decoder module in Transformer and use the extreme learning machine instead of the linear layer to process the output of the encoder to obtain the final prediction result, which can further reduce the consumption of computing resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 Schematic diagram of the pressure sensor calibration process.

[0053] Figure 2 To improve the overall structure of the Transformer model.

[0054] Figure 3 Schematic diagram of the adaptive feature extractor.

[0055] Figure 4 Schematic diagram of extreme learning machine.

[0056] Figure 5 This is the traditional Transformer network structure diagram.

[0057] Figure 6 Overall structure diagram of the pressure sensor data acquisition and communication system. DETAILED DESCRIPTION

[0058] The following is a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings in the examples of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0059] Figure 1 , Figure 6 As shown, a submarine dual-channel output pressure prediction system includes a data acquisition module, a data processing module and an output module;

[0060] The data acquisition module uses sensors to obtain raw data and transmits it to the computer through a communication acquisition board;

[0061] The data processing module: uses an adaptive feature extractor to extract features from the original data, and inputs the feature set F into the improved Transformer model for training;

[0062] The output module is used to visually output the predicted value of the dual-path output pressure of the seabed.

[0063] Pressure Source: Select a pressure source that can provide accurate and stable pressure, such as a pressure pump, pressure generator or pressure controller, to produce the set pressure required in the calibration environment.

[0064] Pressure sensor calibration should be performed in a dust-free, vibration-free, and temperature-constant environment to avoid external factors affecting the calibration results.

[0065] Data collection and preprocessing:

[0066] The pressure sensor data acquisition and communication system is composed of a pressure sensor, a conditioning chip, a communication acquisition board and a computer. The pressure sensor and the conditioning chip together form a pressure sensor module. The pressure sensor converts pressure into a voltage signal and outputs it to the conditioning chip; the conditioning chip integrates temperature sensing and a trained Transformer model, and the voltage signal is passed through the Transformer model to obtain the measured pressure value; the pressure sensor module interacts with the computer through the communication acquisition board, and the communication acquisition board sends the temperature digital signal and pressure digital signal in the pressure sensor module to the computer. The collected data is used to train the Transformer model, and finally all the parameters of the trained Transformer model are saved to the conditioning chip through the data acquisition board.

[0067] A set of 11 standard pressure values ​​is formulated, which correspond to 0%, 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%, and 100% of the range of the pressure sensor to be calibrated. Three standard temperatures are formulated, which are the minimum, maximum, and intermediate values ​​of the working environment temperature of the pressure sensor. The 11 standard pressures and 3 standard temperatures are arranged and combined to obtain a total of 33 sets of pressure sensor calibration environments.

[0068] In order to provide a stable temperature environment, the calibration process needs to be carried out in a constant temperature room. If there is no constant temperature room, air conditioners and heaters can be used to control the temperature of the calibration area.

[0069] The pressure sensor and pressure generator both use the American standard pipe thread (NPT) interface, and stainless steel hard pipes are used to connect the pressure sensor and pressure generator. When connecting the pipes and interfaces, O-rings are used for sealing to prevent pressure leakage.

[0070] Before starting the pressure generator, use soapy water or other leak detection fluid to apply to the connection between the pressure generator and the pressure sensor. If bubbles are generated at the connection, there is a leak on the surface connection and the connection between the pressure generator and the pressure sensor needs to be readjusted.

[0071] After confirming that all connections are secure and leak-free, start the pressure generator and gradually increase the pressure to the desired calibration pressure.

[0072] The pressure sensor detects pressure changes, generates a differential voltage signal and transmits it to the conditioning chip through the differential port. The digital-to-analog converter inside the conditioning chip converts the received analog signal into a voltage value digital signal.

[0073] The communication acquisition board reads the voltage value digital signal in the conditioning chip through the IIC interface and transmits it to the computer through the RS485 interface. The computer records the pressure value P applied by the pressure generator in the current calibration environment, the set temperature T, and the digital voltage signal value V after being processed by the conditioning chip. Finally, 33 sets of data are obtained, as shown in Table 1 below, where T and V are used as the input of the Transformer, and the corresponding P is used as the label. The Transformer model outputs a predicted pressure value Y based on the input V and T. The training goal of the model is to minimize the gap between Y and P.

[0074] Table 1 Data record table

[0075]

[0076]

[0077] The data collected by the data acquisition system are used to make data sets for training and testing of the Transformer model. Among the 33 sets of collected data, 27 sets are randomly selected as training data sets and 6 sets are used as test data sets.

[0078] Figure 1-2 As shown, a method for predicting seabed dual-path output pressure includes the following steps:

[0079] S1. Obtain the original data of the dual-channel output of the seabed, including voltage, pressure and temperature, and use the voltage value V and temperature T as the training set, and the pressure as the training label;

[0080] S2. Extract features of the original data using an adaptive feature extractor to obtain a feature set F;

[0081] S3. Input the feature set F into the improved Transformer model for training;

[0082] S4. Use the calibrated Transformer model to predict the dual-path output pressure at the seafloor.

[0083] The adaptive feature extractor is Figure 3 As shown, it includes two modules: polynomial feature extraction and adaptive feature extraction.

[0084] The polynomial feature module expands the feature set by constructing polynomial features of the input and captures the nonlinear relationship between the inputs. For the input features V and T of the model, the module generates a polynomial combination of these features, including their powers and interactions. The polynomial feature set F can be expressed as:

[0085] F={1,V,T,V 2 ,T 2 ,VT}

[0086] The adaptive feature module uses a small neural network to automatically learn the complex relationship between input features. This neural network consists of several fully connected layers, each followed by a nonlinear activation function. Through training data, the network can learn to map the original features to the new feature space. Define an adaptive feature extraction function f θ , where θ represents the network parameters, the function consists of the following multi-layer neural network:

[0087] f θ (F) = σ(W2[σ(W1F)+b2]) where W1 and W2 are weight matrices, b2 is the bias term, and σ is the activation function, such as RELU.

[0088] The adaptive feature extractor combines the polynomial feature and adaptive feature modules. First, the polynomial feature module is used to generate an extended feature set F, and then these features are input into the adaptive feature module for further learning and transformation. In this way, the model can not only capture the nonlinear relationship of features, but also automatically learn the potential complex patterns in the data.

[0089] Jump Attention Mechanism:

[0090] The core component of Transformer is the self-attention mechanism, but the traditional self-attention mechanism is computationally expensive. Let X = {x1, x2, …, x n} = f θ (F),X∈R n×d Represents the features extracted by the adaptive feature extractor, which is used as the input of the encoder in the Transformer. In the multi-head self-attention block, X is first embedded into three learnable embedding representations {Q, K, V}∈R n×d In which Q is the query matrix, K is the key matrix, and V is the value matrix. The specific calculation formula is as follows:

[0091] Q=XW Q ∈R n×d

[0092] K=XW K ∈R n×d

[0093] V=XW V ∈R n×d

[0094]

[0095] Where W Q ∈R d×d , W K ∈R d×d , WV ∈R d×d Represents three learnable weight matrices.

[0096] In the multi-head attention mechanism, Q, K, and V are divided into h heads, namely:

[0097]

[0098] Where Q i , K i 、V i The shape is For each head, use Q i , K i Calculate the attention matrix A i :

[0099]

[0100] Using the attention matrix A i Sum value matrix V i Compute the output value for each head:

[0101]

[0102] Finally, the outputs of all heads are combined and transformed by the linear transformation W O Remap the dimension of the feature to n×d:

[0103]

[0104] Then the output of the multi-head attention block is represented as X MHA Input into the feedforward neural network MLP, the complete calculation process is as follows:

[0105]

[0106] X l ←MLP(X l )+X l

[0107] Where X l-1 represents the output of the l-1th encoder in the Transformer. Specifically, X 0 =X; Represents the output of the multi-head self-attention mechanism in the l-th encoder.

[0108] The self-attention calculations between different layers in the Transformer are highly correlated. Therefore, the self-attention calculations can be skipped in some layers and the calculation results of the previous layer can be directly reused. In order to ensure that the layers that skip the self-attention calculations can still capture the relationship between the input features, we design a parameterized function Φ: To approximate the output of multi-head self-attention, that is:

[0109]

[0110] in yes , FC1 and FC2 represent two linear layers, DwC represents depthwise separable convolution, and ECA is an efficient channel attention module. The ECA module performs global average pooling in the channel dimension of the input and uses point-by-point convolution to enhance the cross-channel dependency of the input.

[0111] The skip attention mechanism alternately chooses whether to skip the multi-head attention mechanism in each encoder. For the encoder layer that does not skip self-attention, the multi-head self-attention mechanism is used to calculate the output features as the input of the feedforward neural network MLP; for the encoder layer that skips the self-attention calculation, the features output by the parameterized function Φ are used as the input of the feedforward neural network MLP, that is:

[0112]

[0113] X l ←MLP(X l )+X l

[0114] Figure 4 As shown, the extreme learning machine:

[0115] In order to further reduce the computing resource consumption of the Transformer model, we replaced the linear layer used to output the final result with an extreme learning machine. Unlike traditional neural network training methods, the core idea of ​​the extreme learning machine is to randomly determine the weights and biases of hidden neurons and determine the weights of the output layer through analytical methods (such as least squares). This method greatly simplifies the training process and has a very fast training speed.

[0116] The network structure of the extreme learning machine is as follows Figure 4 As shown, the features processed by the jump attention As the input of the extreme learning machine, where L represents the number of encoders in the model, the extreme learning machine can be expressed by the following formula:

[0117]

[0118] Where m is the number of hidden nodes, β j is the connection hidden node h j and the weight vector of the output node, g(·) represents the activation function, w j is the connection hidden node h j and the weight vector of the input node, b jRepresents the hidden node h j The bias value, w j and b j is randomly generated, and β j It can be obtained by solving the following linear equation:

[0119]

[0120] Y is the output value of the extreme learning machine, which is also the pressure value finally predicted by the Transformer model;

[0121] The above formula can also be expressed in matrix form, namely:

[0122] Hβ=Y

[0123] in:

[0124]

[0125] The goal of the extreme learning machine is to solve β to minimize the gap between Hβ and Y. The approximate solution of β can be obtained by solving the following formula:

[0126] β≈H + Y

[0127] Among them, H + is the Moore–Penrose generalized inverse of H.

[0128] Model test calibration:

[0129] The trained model is tested using the test set, and the model is tuned based on the test results. If the deviation between the output and the set pressure is large, the model training process is repeated until the output value of the calibrated pressure sensor is kept within 0.5% of the output value of the reference pressure sensor; if the deviation between the output and the set pressure is within a certain range and reaches the accuracy index of 0.5%, the trained model is saved by the computer to the conditioning chip through the communication acquisition board.

[0130] Using mean square error MSE and coefficient of determination R 2 To evaluate the training results of the proposed model, MSE and R 2 is defined as follows:

[0131]

[0132] Where N represents the number of samples, Y i represents the pressure value predicted by the model for the i-th sample, P i represents the true pressure value in the i-th sample, Represents the mean of the pressure values ​​in the sample.

[0133] The performance of the improved Transformer model is compared with that of the multilayer perceptron and the traditional Transformer model. The results are shown in the following table:

[0134] Table 2 Performance comparison

[0135] Multilayer Perceptron Traditional Transformer model Improved Transformer Model MSE 2.72 0.31 0.30 <![CDATA[R 2 ]]> 0.91 0.99 0.99 Parameter quantity 877K 31M 12M

[0136] Using the improved Transformer model for pressure sensor calibration can effectively deal with complex nonlinear environmental factors and improve the measurement accuracy of the sensor. Compared with traditional calibration methods, the Transformer model has higher flexibility and adaptability. By designing a reasonable data processing process and system architecture, real-time sensor calibration can be achieved to ensure the stability and accuracy of the measurement system. The following are the key innovations and technical points to be protected in a method for data acquisition and calibration of a submarine dual-output pressure sensor.

[0137] A method for collecting and calibrating submarine dual-output pressure sensors innovatively introduces the Transformer model to calibrate the data of submarine pressure sensors, combining the advantages of deep learning in processing complex nonlinear data. Traditional pressure sensor calibration methods usually rely on simple linear models or experience-based algorithms, which are difficult to accurately capture nonlinear errors in complex environments. The Transformer model, through its self-attention mechanism, can effectively model the complex features and nonlinear relationships in sensor data, thereby significantly improving calibration accuracy.

[0138] This innovative method of using the Transformer model for pressure sensor calibration has greatly improved the automation level and accuracy of sensor calibration, and can better adapt to the changing pressure conditions in the complex seabed environment, ensuring high reliability and high accuracy of data acquisition, giving it broad application prospects and competitiveness in the fields of ocean exploration and seabed monitoring.

[0139] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for predicting seabed dual-path output pressure, characterized in that: The following steps are involved: S1. Obtain the original data of the dual-channel output of the seabed, including voltage, pressure and temperature, and use the voltage value V and temperature T as the training set, and the pressure as the training label; S2. Extract features of the original data using an adaptive feature extractor to obtain a feature set F; S3. Input the feature set F into the improved Transformer model for training; S4. Use the calibrated Transformer model to predict the dual-path output pressure at the seafloor.

2. The method for predicting the submarine dual-path output pressure according to claim 1, characterized in that: The adaptive feature extractor comprises a polynomial feature module and an adaptive feature extraction module; The polynomial feature module expands the feature set by constructing polynomial features of the input and captures the nonlinear relationship between the inputs. The module generates polynomial combinations of these features, including their powers and interactions, by inputting the original data V and T. The polynomial feature set F can be expressed as: F={1,V,T,V 2 ,T 2 ,VT} The adaptive feature extraction module uses a small neural network to automatically learn the complex relationship between input features and sets an adaptive feature extraction function f θ , where θ represents the network parameters, the function is composed of the following multi-layer neural network composition: f θ (F)=σ(W2[σ(w1F)+b2]) Where W1 and W2 are weight matrices, b2 is the bias term, and σ is the activation function.

3. The method for predicting the seabed dual-path output pressure according to claim 1, characterized in that: In step S3, the improved Transformer model removes the position encoding in the traditional Transformer model, adds a jump attention mechanism, and uses an extreme learning machine to replace the decoder in the traditional Transformer model for output.

4. The method for predicting the seabed dual-path output pressure according to claim 3, characterized in that: In step S3, the skip attention mechanism alternately selects whether to skip the multi-head attention mechanism in each encoder. For the encoder layer that does not skip self-attention, the multi-head self-attention mechanism is used to calculate the output features as the input of the feedforward neural network MLP; for the encoder layer that skips the self-attention calculation, the features output by the parameterized function Φ are used as the input of the feedforward neural network MLP: Multi-head self-attention mechanism calculates output feature X l as follows: X l ←MLP(X l )+X l X l-1 Represents the output of the l-1th encoder in Transformer; The features output by the parameterized function Φ are as follows: in yes , FC1 and FC2 represent two linear layers, DwC represents depth-wise separable convolution, and the ECA module performs global average pooling in the channel dimension of the input and uses point-wise convolution to enhance the cross-channel dependency of the input.

5. The method for predicting the seabed dual-path output pressure according to claim 1, characterized in that: In step S3, the extreme learning machine includes input, hidden layer and output; its input is X l or Marked with X L , Where L represents the number of encoders in the model; The extreme learning machine is expressed by the following formula: Where m is the number of hidden nodes, β j is the connection hidden node h j and the weight vector of the output node, g(·) represents the activation function, w j is the connection hidden node h j and the weight vector of the input node, b j Represents the hidden node h j The bias value, w j and b j is randomly generated.

6. The method for predicting the seabed dual-path output pressure according to claim 5, characterized in that: β j By solving the following linear equation: Y is the output value of the extreme learning machine, which is also the pressure value finally predicted by the improved Transformer model; The above formula can also be expressed in matrix form, namely: Hβ=Y in: The goal of the extreme learning machine is to solve β to minimize the gap between Hβ and Y. The approximate solution of β can be obtained by the following formula: β≈H + AND Among them, H + is the Moore–Penrose generalized inverse of H.

7. The method for predicting the seabed dual-path output pressure according to claim 5, characterized in that: Using mean square error MSE and coefficient of determination R 2 To evaluate the training results of the proposed model, MSE and R 2 is defined as follows: Where N represents the number of samples, Y i represents the pressure value predicted by the model for the i-th sample, P i represents the true pressure value in the i-th sample, Represents the mean of the pressure values ​​in the sample.

8. The method for predicting the seabed dual-path output pressure according to claim 7, characterized in that: The trained model is tested using the test set, and the model is calibrated based on the test results; if the deviation between the output and the set pressure is large, the model training process is repeated until the output value of the calibrated pressure sensor is kept within 0.5% of the output value of the reference pressure sensor; if the deviation between the output and the set pressure is within a certain range and reaches an accuracy index of 0.5%, the trained model predicts the pressure.

9. A submarine dual-path output pressure prediction system, using the submarine dual-path output pressure prediction method according to any one of claims 1 to 8, characterized in that: It includes a data acquisition module, a data processing module and an output module; The data acquisition module uses sensors to obtain raw data and transmits it to the computer through a communication acquisition board; The data processing module: uses an adaptive feature extractor to extract features from the original data, and inputs the feature set F into the improved Transformer model for training; The output module is used to visually output the predicted value of the dual-path output pressure on the seabed.

10. The submarine dual-output pressure prediction system according to claim 9, characterized in that: The improved Transformer model removes position encoding and mask attention, and replaces layer normalization with batch normalization.

Images