A method for monitoring gas leakage based on spectral vector data processing

Abstract

A method for monitoring gas leakage based on spectral vector data processing, which adopts technical solutions such as collecting spectral vector data, inputting into a generative adversarial network for training, differentiating true sample data and false sample data to achieve data augmentation, inputting into a feature extraction model for training, adjusting the weights and biases of the neural network using different strategies at each stage, inputting into a feature dimensionality reduction model for training, reconstructing features from the low-dimensional representation, inputting into a classifier model for training, converting to qubits, and processing sample data using the trained feature extraction model, data dimensionality reduction model, and classifier model. It overcomes the technical problems of poor model generalization ability caused by insufficient training samples, relying on global gradient information to update weights, and overfitting, and produces technical effects such as improving the quantity and quality of generated data of spectral data samples, enhancing the flexibility of the algorithm at different data stages, minimizing the difference between input and reconstructed output, enhancing the performance of the classifier, and accurately predicting the leakage states of different types of gases.

Description

Technical Field

[0001] The present invention belongs to the technical field of environmental protection early warning and data processing, and specifically relates to a spectral data analysis technology for monitoring gas leakage. Background Art

[0002] Traditional gas leakage detection methods rely on chemical or physical sensing technologies, which are limited in outdoor open spaces. These technologies are restricted by the sensitivity, stability, or response time of sensors, making it difficult to provide early warnings in a timely manner. Traditional methods also show certain limitations when dealing with interference recognition and small-scale leakage recognition in complex environments.

[0003] In summary, the prior art has the following deficiencies:

[0004] There is a lack of effective data augmentation techniques and insufficient training sample sizes, resulting in weak model generalization ability and affecting the accuracy and reliability of the early warning system;

[0005] Feature extraction and dimensionality reduction techniques rely on traditional algorithms, which are not flexible enough, are easily affected by data changes, and cause overfitting problems, reducing the effectiveness and accuracy of feature processing;

[0006] Traditional classifiers are unable to handle complex data interactions effectively, especially in the accurate classification of spectral data. They cannot effectively utilize the deep relationships between data, limiting the improvement of classification performance. Summary of the Invention

[0007] To solve the problems existing in the prior art, the present invention combines various algorithms of data processing technology and adopts the following technical solutions:

[0008] Collect spectral vector data from 3 - 12um wavelengths using a spectral sensing device, and label sample data with categories of no gas leakage, slight gas leakage, normal gas leakage, and severe gas leakage.

[0009] Input the sample data into a generative adversarial network trained with time-frequency feature-assisted synthesis. Use the generator to generate fake sample data close to the real ones to deceive the discriminator, and use the discriminator to try to distinguish between real and fake sample data to achieve data augmentation.

[0010] Input the augmented data into a feature extraction model for training. Use the circular dynamic optimization algorithm to optimize the neural network parameters, adjust the optimization strategy at different training stages, simulate each stage of the cell cycle, and adopt different strategies to adjust the weights and biases of the neural network at each stage.

[0011] The data after feature extraction is input into the feature dimensionality reduction model for training, and the duality feature dimensionality reduction of the encoder and decoder of the autoencoder neural network algorithm based on dual loss is used. The encoder compresses the input feature vector into a low-dimensional representation, and the decoder attempts to reconstruct the input feature from the low-dimensional representation.

[0012] The dimensionality-reduced data is input into the classifier model for training, and the support vector machine algorithm based on quantum information correlation is used to convert the dimensionality-reduced feature vector into qubits.

[0013] The trained feature extraction model, data dimensionality reduction model, and classifier model are used to process the sample data, and the spectral vector data is labeled as and the feature extraction model is used to convert it into a feature vector and the data dimensionality reduction model is used to convert it into a dimensionality-reduced vector and the prediction result is obtained by inputting it into the classifier model , that is, the gas is not leaked, the gas is slightly leaked, the gas is normally leaked, the gas is severely leaked, realizing the monitoring and early warning of gas leakage.

[0014] According to the above technical solution, compared with the prior art, the present invention has the following improved technical effects:

[0015] The acquisition, annotation, and preprocessing of training data are time-consuming and laborious. The spectral vector data is stored in a structured JSON format. The traditional generative adversarial network uses cross-entropy loss to train the generator and discriminator. The generative adversarial network based on time-frequency feature-assisted synthesis, the generator is responsible for generating as realistic sample data as possible, being similar to the real data at the pixel level and close to the real data in terms of time-frequency features. The discriminator distinguishes whether the sample is generated by the generator, overcomes the technical problems that insufficient training samples easily lead to poor model generalization ability and affect the accuracy of the model, and has the technical effects of optimizing the loss function, increasing the quantity and quality of the generated data of spectral data samples, improving the model generalization ability and the accuracy of monitoring and early warning.

[0016] Inspired by the periodic metabolic process of biological cells, the traditional gradient descent method relies on global gradient information to update weights and optimizes the reactions and adaptabilities of organisms through different mechanisms at different stages. The circular dynamic optimization algorithm enhances the flexibility of the algorithm at different data stages and the adaptability to subtle changes, adapts to the changes of training data, reduces the risk of overfitting, effectively selects the local optimal solution, and improves the accuracy of low-dimensional feature representation.

[0017] The traditional encoder uses the reconstruction error as the optimization target. The autoencoder neural network algorithm based on dual loss minimizes the difference between the input and the reconstructed output, and reveals more structural information in the encoded feature representation through the sparse self-expression layer, deeply excavating the inherent structural characteristics of the data.

[0018] Based on the traditional vector machine, the present invention considers the mutual relationship and non-locality between features based on the quantum coding strategy of quantum information correlation, enhances the performance of the classifier, more effectively expresses the complex interaction between data, and enables the model to more accurately predict different types of gas leakage states. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is the training flow chart of the feature extraction model, Figure 2 is the training flow chart of the feature dimensionality reduction model. DETAILED DESCRIPTION OF THE INVENTION

[0020] The technical solution of the present invention will be specifically described below with reference to the accompanying drawings.

[0021] After collecting the spectral vector data, start data augmentation:

[0022] Initialize the network parameters of the generator and discriminator of the generative adversarial network, and set the initial weights and biases;

[0023] Use Uniform(-1,1) to represent the random numbers extracted from the uniform distribution in the interval [-1,1], represents the input dimension of the network layer, represents the output dimension of the network layer, and use the Glorot() strategy to represent the initialization algorithm , use to calculate the initial weights of the generator, to calculate the initial weights of the discriminator;

[0024] Use to represent a part of the set of randomly selected true sample data, represents random noise, represents the network parameters of the generator. Randomly select some samples from the spectral data vector set as the training starting point. The generator, based on the current network parameters, inputs random noise to generate false sample data. Use to represent, and calculate the time-frequency features of the false sample data. Use to represent;

[0025] Use to represent the loss function of the discriminator, represents the learning rate, represents the gradient of the loss function with respect to the discriminator parameters. The discriminator receives the false sample data and the true sample data bx. According to the time-frequency features and the original data of the two, let , and update its network parameters , use to represent, in order to distinguish between false sample data and true sample data;

[0026] Let \(M\) denote the dimension of the time-frequency features, denote the \(j\)-th time-frequency feature of the fake sample data, denote the \(j\)-th time-frequency feature of the real sample data, and use to calculate the time-frequency feature loss;

[0027] Use to denote the loss function of the generator, denote the time-frequency features of the real sample data, denote the weight of the time-frequency feature loss, denote the fixed learning rate, denote the adaptive learning rate at the \(t\)-th iteration, denote the first-order moment estimate at the \(t\)-th iteration, denote the second-order moment estimate at the \(t\)-th iteration, denote the gradient of the loss function with respect to the generator parameters. Use the time-frequency feature loss to guide the network training, and let , update the loss function of the generator, and use to calculate and update the network parameters of the generator ; use to denote;

[0028] Use to denote the gradient at the current step, and denote the decay rate. Use to calculate the estimate of the first-order moment, and calculate the estimate of the second-order moment;

[0029] Repeat the above process iteratively until the preset stop condition is satisfied. Preferably, the number of iterations takes a value of 1000.

[0030] Input the augmented data into the feature extraction model for training. The process is as Figure 1 shown:

[0031] Initialize the neural network model. Use a 3-layer fully connected neural network and the ReL_RU activation function for feature extraction. The number of hidden layers of the neural network is 2. The first hidden layer includes 100 neurons, and the second hidden layer includes 50 neurons;

[0032] Use \(\sim\) to denote being subject to a specific distribution. Use to denote a normal distribution with a mean of 0 and a variance of , and use to denote the initial learning rate. Initialize the weights of the neural network using to denote, and initialize the biases of the neural network using to denote;

[0033] Use Denote the weight parameter, Denote the bias parameter, Denote the acceleration coefficient, Denote the loss function for neural network training, Denote the gradient of the loss function with respect to the weight parameter, Denote the gradient of the loss function with respect to the bias parameter, Denote the weight parameter at the t-th iteration, Denote the bias parameter at the (t + 1)-th iteration, ∂ denotes the partial derivative, Denote the growth-phase learning rate. In the growth phase, update the neural network parameters using the accelerated gradient descent algorithm, using and to calculate;

[0034] Use to denote the preset maximum acceleration coefficient, Denote the minimum function, use to calculate the acceleration coefficient and avoid over-adjustment;

[0035] Use to denote the importance coefficient of the i-th weight parameter nr denotes the number of weight parameters. In the synthesis phase, synthesize multiple weight parameters into new parameters to improve the generalization ability of the network, using to calculate;

[0036] Use to calculate the decay rate parameter, based on the magnitude of the gradient of the growth-phase weight parameter Use to dynamically calculate the importance coefficient;

[0037] Use to denote the performance threshold. In the evaluation phase, evaluate the current network performance in the form of a decision function and decide whether to adjust the cycle optimization strategy, using to denote. If is 1, enter the growth phase; otherwise, enter the prophase of division;

[0038] Use to denote the adjustment increment of the weight parameter, to denote the adjustment increment of the bias parameter, to denote the amplitude parameter of fine-tuning, to denote the frequency parameter of fine-tuning. In the prophase of division, simulate the preparatory division stage of cells, using and to fine-tune the weights;

[0039] Use to denote the scaling factor of the perturbation amplitude, using and to calculate;

[0040] Use to represent element-wise multiplication, represent the gradient threshold, and constrain the weights in the splitting phase with the weight pruning function in the way of, and calculate with ;

[0041] Use to represent the attenuation coefficient, represent the number of completed cycles. In the dormant period, the network gradually reaches a stable state, adjusts the learning rate to decrease, and calculate with ;

[0042] Repeat the above process iteratively until the preset stop condition is met. The preferred number of iterations is 1000.

[0043] Input the data after feature extraction into the feature dimensionality reduction model for training, and the process is as Figure 2 shown:

[0044] Use to represent the encoder parameters, represent the decoder parameters, and set the weight mode and bias mode of the encoder and decoder to random initialization;

[0045] Use to represent the input feature vector, z represents the encoded low-dimensional vector, represent the reconstructed output vector. The input feature vector is transformed into a low-dimensional vector by the encoder, and represented by , and then reconstructed into an output vector of the original dimension by the decoder, and represented by ;

[0046] Use n to represent the number of samples, and represent the regularization parameter, S represents the self-expression matrix, represent a single sample, represent the reconstructed sample, represent the encoded sample, represent the dual loss, represent the reconstruction error, represent the sparse self-expression error, and calculate the dual loss with . The reconstruction error measures the error between the original input and the reconstructed output, and calculate with . The sparse self-expression error measures the error between the encoded low-dimensional samples and the sparse linear combination, and calculate with ;

[0047] Use Z to represent the matrix of all encoded samples of, represent the hyperparameter that controls the matrix density, The Euclidean distance is denoted as, the Softmax( ) function is denoted as Softmax( ), the self-expression matrix S is constructed, and and is denoted as;

[0048] Using to denote the learning rate of the dual loss, to denote the dual loss with respect to the gradient of, to denote the dual loss with respect to the gradient of, calculate the dual loss using the chain rule with respect to the encoder parameters and the decoder parameters the gradient of, update the parameters, and use and to denote;

[0049] Repeat the above process iteratively until the preset stop condition is satisfied. Preferably, the number of iterations is 1000.

[0050] Input the dimension-reduced data into the classifier model for training:

[0051] Using to denote the initialized quantum state, to denote the first initial rotation angle parameter, to denote the second initial rotation angle parameter, to denote the third initial rotation angle parameter, to denote the initial state of n qubits, to denote the quantum gate for each qubit, including rotation and entanglement operations, initialize the quantum state encoding module, and use to denote;

[0052] Using to denote the rotation operation around the z-axis, to denote the rotation operation around the y-axis, θ denotes the first rotation angle parameter, ϕ denotes the second rotation angle parameter, to denote the third rotation angle parameter, and construct the quantum gate using the basic quantum operations ;

[0053] Using to denote the quantum state encoded by the eigenvector qx, design the quantum encoding unit according to the input feature qx, and use to denote, convert the dimension-reduced eigenvector into a quantum state, and each eigenvector corresponds to a quantum state, and use to denote;

[0054] Using Denote the k-th component of the feature vector qx, using Calculate the quantum coding unit;

[0055] Use the correlation metric of quantum states as the kernel function of the support vector machine to construct a quantum-enhanced classification model, and calculate the overlap between two quantum states using the quantum classifier kernel function;

[0056] Use Denote the quantum kernel function, Denote the quantum state of the i-th data point, Denote the quantum state of the j-th data point, using Calculate the squared modulus of the inner product of two quantum states, reflecting the similarity between two feature vectors in the quantum state space;

[0057] Train the support vector machine using the correlation metric between quantum states, and calculate the classification boundary through quantum simulation or a quantum computer;

[0058] Use Denote the optimized Lagrange multiplier, α denotes the Lagrange multiplier vector, Q denotes the matrix constructed by the quantum kernel function K, N denotes the number of training samples, Denote the label of the i-th sample, Denote the label of the j-th sample, C denotes the regularization parameter of the support vector machine, using Calculate the element in the i-th row and j-th column of the matrix constructed by the quantum kernel function K, and let the classification function Subject to Constraints;

[0059] Use b to denote the bias term of the decision boundary, use sgn( ) to denote the sign function, output the classification prediction, using Denote the decision function of the support vector machine algorithm based on quantum information correlation.

[0060] The above Take the value of 0.03, Take the value of 0.01, Take the value of 0.3, And Take the value of 2, Take the value of 0.01, Select the cross-entropy loss, Take the value of 0.01, Take the value of 0.5, Take the value of 5, Take the value of 3.14, Take the value of 0.95, Take the value of 0.01, Take the value of 0.95, Take the value of 0.01, Take the value of π / 4, Taking the value of π / 2, When taking the value of π / 3 and C taking the value of 10, the effect is better.

[0061] Using the trained feature extraction model, data dimensionality reduction model and classifier model to process the sample data to achieve the monitoring and early warning of gas leakage.

[0062] The above are the embodiments of the present invention and do not limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention are all included in the protection scope of the present invention.

Claims

1. A method for monitoring gas leakage based on spectral vector data processing, characterized in that Including: Collect spectral vector data from 3 - 12um wavelength using a spectral sensing device, and label sample data with categories of no gas leakage, slight gas leakage, normal gas leakage, and severe gas leakage; Input the sample data into a generative adversarial network trained with time - frequency feature - assisted synthesis. Use the generator to generate fake sample data close to the real ones to deceive the discriminator, and use the discriminator to try to distinguish between real and fake sample data to achieve data augmentation; Input the augmented data into a feature extraction model for training. Use the cyclic dynamic optimization algorithm to optimize the neural network parameters, adjust the optimization strategy at different training stages, simulate each stage of the cell cycle, and adopt different strategies to adjust the weights and biases of the neural network at each stage; Input the data after feature extraction into a feature dimensionality reduction model for training. Use the duality feature of the encoder and decoder of the auto - encoding neural network algorithm based on dual loss for feature dimensionality reduction. The encoder compresses the input feature vector into a low - dimensional representation, and the decoder tries to reconstruct the input feature from the low - dimensional representation; Input the data after dimensionality reduction into a classifier model for training. Use the support vector machine algorithm based on quantum information correlation to convert the dimensionality - reduced feature vector into quantum bits; Process the sample data with the trained feature extraction model, data dimensionality reduction model, and classifier model, and label the spectral vector data as , and use the feature extraction model to convert it into a feature vector , use the data dimensionality reduction model to convert it into a dimensionality-reduced vector , input it into the classifier model to obtain the prediction result , that is, the gas is not leaking, the gas is slightly leaking, the gas is normally leaking, the gas is severely leaking, realizing the monitoring and early warning of gas leakage; The data augmentation includes: Initialize the network parameters of the generator and discriminator of the generative adversarial network, and set the initial weights and biases; Use Uniform(-1, 1) to represent the random numbers drawn from a uniform distribution in the interval [-1, 1]. Represents the input dimension of the network layer. Represents the output dimension of the network layer, and the Glorot() strategy is used to represent the initialization algorithm. , using Calculate the initial weights of the generator. Calculate the initial weights of the discriminator. Use To represent a subset of the randomly selected true sample data, To represent random noise, To represent the network parameters of the generator. Randomly select some samples from the spectral data vector set as the training starting point. The generator, based on the current network parameters, inputs random noise to generate fake sample data. Use To represent the calculation of the time-frequency characteristics of the fake sample data. Use To represent; Use to represent the loss function of the discriminator, to represent the learning rate, to represent the gradient of the loss function with respect to the discriminator parameters. The discriminator receives fake sample data and real sample data bx. According to their time-frequency features and the original data, let , and update its network parameters ; use to represent to distinguish between fake sample data and real sample data; Let \(M\) denote the dimension of the time-frequency features, denote the \(j\)-th time-frequency feature of the false sample data, denote the \(j\)-th time-frequency feature of the true sample data, and use to calculate the time-frequency feature loss; Use represent the loss function of the generator, represent the time-frequency features of the real sample data, represent the weight of the time-frequency feature loss, represent the fixed learning rate, represent the adaptive learning rate at the t-th iteration, represent the first-order moment estimate at the t-th iteration, represent the second-order moment estimate at the t-th iteration, represent the gradient of the loss function with respect to the generator parameters. Use the time-frequency feature loss to guide the network training. Let , update the loss function of the generator. Use to calculate and update the network parameters of the generator ; Use to represent; Use to represent the gradient of the current step, and to represent the decay rate, use to calculate the estimate of the first moment, to calculate the estimate of the second moment; Repeat the above process iteratively until the preset stop condition is met; The input of the augmented data into the feature extraction model for training includes: Initialize the neural network model, use a 3 - layer fully - connected neural network, and use the ReL_RU activation function for feature extraction. The number of hidden layers of the neural network is 2. The first hidden layer includes 100 neurons, and the second hidden layer includes 50 neurons; Let ∼ denote being subject to a specific distribution, and use to represent a normal distribution with a mean of 0 and a variance of . Use to represent the initial learning rate, initialize the weights of the neural network, and use to represent the initialization of the biases of the neural network; use to represent. Use to represent the weight parameter, to represent the bias parameter, to represent the acceleration coefficient, to represent the loss function for neural network training, to represent the gradient of the loss function with respect to the weight parameter, to represent the gradient of the loss function with respect to the bias parameter, to represent the weight parameter at the t-th iteration, to represent the bias parameter at the (t + 1)-th iteration, ∂ represents the partial derivative, to represent the growth phase learning rate. During the growth phase, the neural network parameters are updated using the accelerated gradient descent algorithm, and are calculated using and ; Use to represent the preset maximum acceleration coefficient, to represent the minimum value function, and use to calculate the acceleration coefficient and avoid excessive adjustment; Use to represent the importance coefficient of the i-th weight parameter , nr represents the number of weight parameters. During the synthesis period, multiple weight parameters are synthesized into a new parameter to improve the generalization ability of the network, and it is calculated using ; Use to calculate the decay rate parameter, based on the gradient magnitude of the growth phase weight parameter and use to dynamically calculate the importance coefficient; Use to represent the performance threshold. During the evaluation period, the current network performance is evaluated in the form of a decision function to determine whether to adjust the cycle optimization strategy. Use to represent that if is 1, it enters the growth period; otherwise, it enters the prophase of division. Use to represent the adjustment increment of the weight parameter, to represent the adjustment increment of the bias parameter, to represent the amplitude parameter of the fine-tuning, to represent the frequency parameter of the fine-tuning. During the pre-splitting stage, to simulate the preparatory splitting stage of cells, use and to fine-tune the weights; Use to represent the scaling factor of the perturbation amplitude, and use and for calculation; Use to represent element multiplication, represent the gradient threshold, and constrain the weights in the form of a weight pruning function during the splitting period, using to calculate; Calculate; Use to represent the attenuation coefficient, to represent the number of completed cycles. During the dormant period, the network gradually reaches a stable state, adjusts the learning rate to decrease, and uses to calculate; Repeat the above process iteratively until the preset stop condition is met; The input of the data after feature extraction into the feature dimensionality reduction model for training includes: Use to represent the encoder parameters, to represent the decoder parameters, and set the weight and bias methods of the encoder and decoder to random initialization; Use to represent the input feature vector, z to represent the encoded low-dimensional vector, and to represent the reconstructed output vector. The input feature vector is transformed into a low-dimensional vector by the encoder, which is represented by , and then reconstructed into an output vector of the original dimension by the decoder, which is represented by ; Let \(n\) denote the number of samples, and denote the regularization parameter, \(S\) denote the self - expression matrix, \(\mathbf{x}_i\) denote a single sample, \(\hat{\mathbf{x}}_i\) denote the reconstructed sample, \(\mathbf{z}_i\) denote the encoded sample, \(L_d\) denote the dual loss, \(L_r\) denote the reconstruction error, \(L_s\) denote the sparse self - expression error, and use to calculate the dual loss, the reconstruction error measures the error between the original input and the reconstructed output, and use to calculate, the sparse self - expression error measures the error between the encoded low - dimensional sample and the sparse linear combination, and use to calculate; Let \(Z\) denote all the encoded samples matrix, a hyperparameter representing the density of the control matrix, denote the Euclidean distance, Softmax( ) denote the Softmax function, construct the self-expression matrix \(S\), and use and to represent; Use to represent the learning rate of the dual loss, represent the dual loss with respect to the gradient of, represent the dual loss with respect to the gradient of, calculate the dual loss using the chain rule with respect to the encoder parameters and the decoder parameters the gradient of, update the parameters, use and to represent; Repeat the above process iteratively until the preset stop condition is met; The conversion of the dimensionality - reduced feature vector into quantum bits includes: Use to represent the initialized quantum state, to represent the first initial rotation angle parameter, to represent the second initial rotation angle parameter, to represent the third initial rotation angle parameter, to represent the initial state of n qubits, to represent the quantum gates for each qubit, including rotation and entanglement operations, and the initialized quantum state encoding module, use to represent; Use to represent a rotation operation about the z-axis, to represent a rotation operation about the y-axis, θ represents the first rotation angle parameter, ϕ represents the second rotation angle parameter, to represent the third rotation angle parameter, and use basic quantum operations to combine to construct a quantum gate; Use to represent the quantum state after encoding the feature vector qx. Design a quantum encoding unit according to the input feature qx, and use to represent converting the dimensionality-reduced feature vector into a quantum state, where each feature vector corresponds to a quantum state, and use to represent; Use to represent the k-th component of the eigenvector qx, and use to calculate the quantum encoding unit; Use the correlation measure of quantum states as the kernel function of the support vector machine to construct a quantum - enhanced classification model, and calculate the overlap between two quantum states using the quantum classifier kernel function; Use to represent the quantum kernel function, to represent the quantum state of the i-th data point, to represent the quantum state of the j-th data point, and use to calculate the modulus square of the inner product of two quantum states, reflecting the similarity between two eigenvectors in the quantum state space; Train the support vector machine using the correlation measure between quantum states, and calculate the classification boundary through quantum simulation or a quantum computer; Use represents the optimized Lagrange multiplier, α represents the Lagrange multiplier vector, Q represents the matrix constructed by the quantum kernel function K, N represents the number of training samples, represents the label of the i-th sample, represents the label of the j-th sample, C represents the regularization parameter of the support vector machine, and use to calculate the element in the i-th row and j-th column of the matrix constructed by the quantum kernel function K, and let the classification function be subject to constraints; Let \(b\) denote the bias term of the decision boundary, and let \(\text{sgn}( )\) denote the sign function and output the classification prediction. denote the decision function of the support vector machine algorithm based on quantum information correlation.

2. The method for monitoring gas leakage based on spectral vector data processing according to claim 1, wherein The value is 0.03, value is 0.01, value is 0.3, and the value is 2, and the number of iterations is 1000.

3. The method for monitoring gas leakage based on spectral vector data processing according to claim 1, characterized in that, The value is 0.01, select cross-entropy loss, value is 0.01, value is 0.5, value is 5, value is 3.14, value is 0.95, value is 0.01, value is 0.95, and the number of iterations is 1000.

4. The method for monitoring gas leakage based on spectral vector data processing according to claim 1, wherein The value is 0.01 and the number of iterations is 1000.

5. The method for monitoring gas leakage based on spectral vector data processing according to claim 1, characterized in that The said takes the value of π / 4, takes the value of π / 2, takes the value of π / 3, and C takes the value of 10.

Images