Method for monitoring gas leakage based on spectral vector data processing

Abstract

The invention discloses a method for monitoring gas leakage based on spectral vector data processing. The method comprises the following steps: collecting spectral vector data, inputting generative adversarial network training, distinguishing true sample data and false sample data to realize data expansion, inputting feature extraction model training, adopting different strategies to adjust the weight and bias of a neural network in each stage, inputting feature dimension reduction model training, representing reconstructed features from low dimensions, and extracting the reconstructed features. According to the technical scheme of inputting a classifier model for training, converting into quantum bits and processing sample data by using the trained feature extraction model, the data dimension reduction model and the classifier model, the technical problems of poor model generalization ability, weight updating depending on global gradient information and overfitting caused by insufficient training samples are solved; the technical effects of improving the quantity and quality of the data generated by the spectral data sample, enhancing the flexibility of the algorithm in different data stages, minimizing the difference between input and reconstruction output, enhancing the performance of the classifier and accurately predicting the leakage states of different types of gases are achieved.

Description

Technical Field

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

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

[0003] In summary, the prior art has the following deficiencies: Lack of effective data expansion technology and insufficient training samples lead to weak model generalization ability, affecting the accuracy and reliability of the early warning system; Feature extraction and dimensionality reduction techniques rely on traditional algorithms, which are not flexible enough and are easily affected by data changes, resulting in overfitting problems and reducing the effectiveness and accuracy of feature processing; Traditional classifiers are unable to handle complex data interactions, especially in the accurate classification of spectral data. They are unable to effectively utilize the deep relationships between data, which limits the improvement of classification performance. Summary of the invention

[0004] In order to solve the problems existing in the prior art, the present invention combines multiple algorithms of data processing technology and adopts the following technical solutions: Use spectral sensing equipment to collect spectral vector data from 3-12um wavelengths, and mark the sample data into categories such as no gas leakage, slight gas leakage, normal gas leakage, and severe gas leakage.

[0005] The sample data is input into the generative adversarial network training based on time-frequency feature assisted synthesis. The generator is used to generate fake sample data that is close to the real one to deceive the discriminator. The discriminator is used to try to distinguish between real sample data and fake sample data to achieve data expansion.

[0006] The expanded data is input into the feature extraction model training, and the neural network parameters are optimized using a circular dynamic optimization algorithm. The optimization strategy is adjusted at different training stages to simulate the various stages of the cell cycle. Different strategies are used in each stage to adjust the weights and biases of the neural network.

[0007] The data after feature extraction is input into the feature dimension reduction model training, and the duality feature dimension 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 to a low-dimensional representation, and the decoder attempts to reconstruct the input features from the low-dimensional representation.

[0008] The reduced-dimensional data is input into the classifier model for training, and the reduced-dimensional feature vectors are converted into quantum bits using a support vector machine algorithm based on quantum information association.

[0009] The sample data is processed using the trained feature extraction model, data dimension reduction model and classifier model, and the spectral vector data is marked as , using the feature extraction model Convert to feature vector , using the data dimensionality reduction model Convert to a reduced dimension vector , input the classifier model to get the prediction result , that is, no gas leakage, slight gas leakage, normal gas leakage, and serious gas leakage, to achieve gas leakage monitoring and early warning.

[0010] According to the above technical solution, compared with the prior art, the present invention has the following improved technical effects: The acquisition, labeling and preprocessing of training data are time-consuming and labor-intensive. 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. Based on the generative adversarial network assisted by time-frequency feature synthesis, the generator is responsible for generating sample data that is as realistic as possible, 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. It overcomes the technical problem that insufficient training samples can easily lead to poor model generalization ability and affect the accuracy of the model. It has the technical effect of optimizing the loss function, increasing the quantity and quality of data generated by spectral data samples, and improving the generalization ability of the model and the accuracy of monitoring and early warning.

[0011] The traditional gradient descent method is inspired by the periodic metabolic process of biological cells, relies on global gradient information to update weights, and optimizes the response and adaptability of organisms through different mechanisms at different stages. The circular dynamic optimization algorithm enhances the flexibility of the algorithm at different data stages and its ability to adapt to subtle changes, adapts to changes in training data, reduces the risk of overfitting, effectively selects local optimal solutions, and improves the accuracy of low-dimensional feature representation.

[0012] Traditional encoders use reconstruction error as the optimization target. The dual loss-based autoencoder neural network algorithm minimizes the difference between the input and the reconstructed output, and uses the sparse self-expressive layer to force more structural information to be revealed in the encoded feature representation, thus deeply mining the intrinsic structural characteristics of the data.

[0013] Based on the traditional vector machine, the present invention considers the correlation and non-locality between features based on the quantum coding strategy of quantum information association, enhances the performance of the classifier, more effectively expresses the complex interactions between data, and enables the model to more accurately predict different types of gas leakage states. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is the feature extraction model training flow chart. Figure 2 It is a flow chart of feature dimension reduction model training. DETAILED DESCRIPTION

[0015] The technical solution of the present invention is described in detail below with reference to the accompanying drawings.

[0016] After collecting the spectral vector data, start data augmentation: 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 a random number 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 uses the Glorot() strategy to represent the initialization algorithm ,use Calculate the initial weights of the generator, Calculate the initial weights of the discriminator; use Represents a set of randomly selected quantities of true sample data, represents random noise, Represents the network parameters of the generator. Some samples are randomly selected from the spectral data vector set as the training starting point. The generator generates fake sample data based on the current network parameters and inputs random noise. Indicates that the time-frequency characteristics of the fake sample data are calculated and used express; use represents the loss function of the discriminator, represents the learning rate, Represents the gradient of the loss function with respect to the discriminator parameters, and the discriminator receives fake sample data and true sample data bx, according to the time-frequency characteristics and original data of the two, let , update its network parameters ,use Representation, to distinguish false sample data from true sample data; M represents the dimension of time-frequency features. represents the jth time-frequency feature of the fake sample data, Represents the jth time-frequency feature of the true sample data, and is used Calculate time-frequency feature loss; use represents the loss function of the generator, Represents the time-frequency characteristics of the real sample data, represents the weight of time-frequency feature loss, represents a fixed learning rate, represents the adaptive learning rate of the tth iteration, represents the first-order moment estimate of the t-th iteration, represents the second-order moment estimate of the t-th iteration, Represents the gradient of the loss function with respect to the generator parameters, and uses the time-frequency feature loss to guide network training. , update the loss function of the generator, and use Calculate and update the network parameters of the generator ,use express; use represents the gradient of the current step, and The decay rate is expressed as Compute estimates of the first moments, Compute estimates of second-order moments; The above process is iterated repeatedly until the preset stop condition is met, and the preferred number of iterations is 1000.

[0017] The expanded data is input into the feature extraction model for training. The process is as follows Figure 1 As shown: Initialize the neural network model, use a 3-layer fully connected neural network, 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; We use ∼ to indicate that it obeys a specific distribution, and It means the mean is 0 and the variance is The normal distribution of Represents the initial learning rate, initializes the weights of the neural network, and uses Indicates that the bias of the initialization neural network is express; use represents the weight parameter, represents the bias parameter, represents the acceleration factor, represents the loss function of neural network training, represents the gradient of the loss function with respect to the weight parameter, represents the gradient of the loss function with respect to the bias parameter, represents the weight parameter of the tth iteration, represents the bias parameter of the t+1th iteration, ∂ represents the partial derivative, represents the learning rate in the growth period. In the growth period, the accelerated gradient descent algorithm is used to update the neural network parameters. and calculate; use Indicates the preset maximum acceleration factor, Represents the minimum function, using Calculate the acceleration factor and avoid over-adjustment; use Represents the i-th weight parameter The importance coefficient of nr represents the number of weight parameters. During the synthesis period, multiple weight parameters are synthesized into new parameters to improve the generalization ability of the network. calculate; use Calculate the decay rate parameter based on the growth period weight parameter The gradient size is Dynamic calculation of importance coefficients; use represents the performance threshold, and the decision function The current network performance is evaluated in the form of Indicates that if If it is 1, it enters the growth phase, otherwise it enters the pre-mitotic phase; use Indicates the adjustment increment of the weight parameter, Indicates the adjustment increment of the bias parameter, represents the amplitude parameter for fine-tuning, represents the fine-tuned frequency parameter, simulating the cell's preparation for division in the prophase. and Fine-tune weights; use Denotes the scaling factor of the perturbation amplitude, and is expressed as and calculate; use represents element-wise multiplication, Represents the gradient threshold, which is used to prune the function with weights during the split period Constrain the weights by calculate; use represents the attenuation coefficient, Indicates the number of completed cycles. During the dormant period, the network gradually reaches a stable state, and the learning rate is adjusted to decrease. calculate; The above process is iterated repeatedly until the preset stop condition is met, and the preferred number of iterations is 1000.

[0018] The data after feature extraction is input into the feature dimension reduction model training. The process is as follows Figure 2 As shown: use Indicates encoder parameters, Represents the decoder parameters, setting the weights and biases of the encoder and decoder to random initialization; use represents the input feature vector, z represents the encoded low-dimensional vector, Represents the reconstructed output vector. The input feature vector is converted into a low-dimensional vector through the encoder. Represented, and then reconstructed into the output vector of the original dimension through the decoder, express; Let n represent the number of samples, and represents the regularization parameter, S represents the self-expression matrix, Represents a single sample, Represents the reconstructed sample, Represents the encoded sample, represents the dual loss, represents the reconstruction error, Denotes the sparse self-expression error, and uses Calculate the dual loss, the reconstruction error measures the error between the original input and the reconstructed output, and Calculate the error between the low-dimensional sample and the sparse linear combination after encoding by sparse self-expression error measurement, and use calculate; Let Z represent all coded samples The matrix of represents the hyperparameter controlling the density of the matrix, represents the Euclidean distance, Softmax( ) represents the Softmax function, constructs the self-expression matrix S, and uses and express; use represents the learning rate of the dual loss, Represents the dual loss about The gradient, Represents the dual loss about The gradient of , using the chain rule to calculate the dual loss About encoder parameters and decoder parameters The gradient of , update the parameters, use and express; The above process is iterated repeatedly until the preset stop condition is met, and the preferred number of iterations is 1000.

[0019] Input the reduced dimension data into the classifier model for training: use represents the initialized quantum state, represents the first initial rotation angle parameter, represents the second initial rotation angle parameter, Represents the third initial rotation angle parameter, represents the initial state of n quantum bits, Represent the quantum gates of each quantum bit, including rotation and entanglement operations, initialize the quantum state encoding module, and use express; use Represents the rotation operation around the z-axis, represents the rotation operation around the y-axis, θ represents the first rotation angle parameter, ϕ represents the second rotation angle parameter, Represents the third rotation angle parameter, using a combination of basic quantum operations Constructing quantum gates; use Represents the quantum state after the feature vector qx is encoded. According to the input feature qx, the quantum encoding unit is designed and used Indicates that the reduced-dimensional feature vectors are converted into quantum states, each feature vector corresponds to a quantum state, and is expressed as express; use Denote the kth component of the eigenvector qx, using Computational quantum coding unit; Using the correlation measure of quantum states as the kernel function of the support vector machine, a quantum-enhanced classification model is constructed, and the overlap between two quantum states is calculated using the quantum classifier kernel function; use represents the quantum kernel function, represents the quantum state of the ith data point, Represents the quantum state of the jth data point, using Calculate the square of the inner product of two quantum states, reflecting the similarity of the two eigenvectors in the quantum state space; Train support vector machines using correlation measures between quantum states and compute classification boundaries using quantum simulations or quantum computers; 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 i-th sample label, represents the jth sample label, C represents the regularization parameter of the support vector machine, and Calculate the elements of the i-th row and j-th column of the matrix constructed by the quantum kernel function K, and let the classification function by Constraints; Let b denote the bias term of the decision boundary, sgn() denote the symbolic function and output classification prediction, and Represents the decision function of the support vector machine algorithm based on quantum information association.

[0020] Above The value is 0.03. The value is 0.01. The value is 0.3. and Value 2, The value is 0.01. Select cross entropy loss, The value is 0.01. The value is 0.5. Value 5, The value is 3.14. The value is 0.95. The value is 0.01. The value is 0.95. The value is 0.01. The value is π / 4. The value is π / 2. When the value is π / 3 and the value of C is 10, the effect is better.

[0021] The trained feature extraction model, data dimension reduction model and classifier model are used to process sample data to achieve gas leakage monitoring and early warning.

[0022] The above are embodiments of the present invention and do not limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention are 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: include: Use spectral sensing equipment to collect spectral vector data from 3-12um wavelengths, and label the sample data into categories such as no gas leakage, slight gas leakage, normal gas leakage, and severe gas leakage; The sample data is input into the generative adversarial network training based on time-frequency feature-assisted synthesis. The generator is used to generate fake sample data close to the real one to deceive the discriminator. The discriminator is used to try to distinguish between real sample data and fake sample data to achieve data expansion. The expanded data is input into the feature extraction model training, and the neural network parameters are optimized using the circular dynamic optimization algorithm. The optimization strategy is adjusted at different training stages to simulate the various stages of the cell cycle. Different strategies are used in each stage to adjust the weights and biases of the neural network. The extracted data is input into the feature dimension reduction model training, and the duality feature dimension reduction is performed by the encoder and decoder of the autoencoder neural network algorithm based on dual loss. The encoder compresses the input feature vector to a low-dimensional representation, and the decoder attempts to reconstruct the input feature from the low-dimensional representation. The reduced-dimensional data is input into the classifier model for training, and the reduced-dimensional feature vector is converted into quantum bits using a support vector machine algorithm based on quantum information association; The sample data is processed using the trained feature extraction model, data dimension reduction model and classifier model, and the spectral vector data is marked as , using the feature extraction model Convert to feature vector , using the data dimensionality reduction model Convert to a reduced dimension vector , input the classifier model to get the prediction result , that is, no gas leakage, slight gas leakage, normal gas leakage, and serious gas leakage, to achieve gas leakage monitoring and early warning.

2. The method for monitoring gas leakage based on spectral vector data processing according to claim 1 is characterized in that: The data expansion 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 a random number 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 uses the Glorot() strategy to represent the initialization algorithm ,use Calculate the initial weights of the generator, Calculate the initial weights of the discriminator; use Represents a set of randomly selected quantities of true sample data, represents random noise, Represents the network parameters of the generator. Some samples are randomly selected from the spectral data vector set as the training starting point. The generator generates fake sample data based on the current network parameters and inputs random noise. Indicates that the time-frequency characteristics of the fake sample data are calculated and used express; use represents the loss function of the discriminator, represents the learning rate, Represents the gradient of the loss function with respect to the discriminator parameters, and the discriminator receives fake sample data and true sample data bx, according to the time-frequency characteristics and original data of the two, let , update its network parameters ,use Representation, to distinguish false sample data from true sample data; M represents the dimension of time-frequency features. represents the jth time-frequency feature of the fake sample data, Represents the jth time-frequency feature of the true sample data, and is used Calculate time-frequency feature loss; use represents the loss function of the generator, Represents the time-frequency characteristics of the real sample data, represents the weight of time-frequency feature loss, represents a fixed learning rate, represents the adaptive learning rate of the tth iteration, represents the first-order moment estimate of the t-th iteration, represents the second-order moment estimate of the t-th iteration, Represents the gradient of the loss function with respect to the generator parameters, and uses the time-frequency feature loss to guide network training. , update the loss function of the generator, and use Calculate and update the network parameters of the generator ,use express; use represents the gradient of the current step, and The decay rate is expressed as Compute estimates of the first moments, Compute estimates of second-order moments; The above process is iterated repeatedly until the preset stopping condition is met.

3. The method for monitoring gas leakage based on spectral vector data processing according to claim 2 is characterized in that: The step of inputting the expanded data into the feature extraction model for training comprises: Initialize the neural network model, use a 3-layer fully connected neural network, 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; We use ∼ to indicate that it obeys a specific distribution, and It means the mean is 0 and the variance is The normal distribution of Represents the initial learning rate, initializes the weights of the neural network, and uses Indicates that the bias of the initialization neural network is used express; use represents the weight parameter, represents the bias parameter, represents the acceleration factor, represents the loss function of neural network training, represents the gradient of the loss function with respect to the weight parameter, represents the gradient of the loss function with respect to the bias parameter, represents the weight parameter of the tth iteration, represents the bias parameter of the t+1th iteration, ∂ represents the partial derivative, represents the learning rate in the growth period. In the growth period, the accelerated gradient descent algorithm is used to update the neural network parameters. and calculate; use Indicates the preset maximum acceleration factor, Represents the minimum function, using Calculate the acceleration factor and avoid over-adjustment; use Represents the i-th weight parameter The importance coefficient of nr represents the number of weight parameters. During the synthesis period, multiple weight parameters are synthesized into new parameters to improve the generalization ability of the network. calculate; use Calculate the decay rate parameter based on the growth period weight parameter The gradient size is Dynamic calculation of importance coefficients; use represents the performance threshold, and the decision function The current network performance is evaluated in the form of Indicates that if If it is 1, it enters the growth phase, otherwise it enters the pre-mitotic phase; use Indicates the adjustment increment of the weight parameter, Indicates the adjustment increment of the bias parameter, represents the amplitude parameter for fine-tuning, represents the fine-tuned frequency parameter, simulating the cell's preparation for division in the prophase. and Fine-tune weights; use represents the scaling factor of the perturbation amplitude, and is expressed as and calculate; use represents element-wise multiplication, Represents the gradient threshold, which is used to prune the function with weights during the split period Constrain the weights by calculate; use represents the attenuation coefficient, Indicates the number of completed cycles. During the dormant period, the network gradually reaches a stable state, and the learning rate is adjusted to decrease. calculate; The above process is iterated repeatedly until the preset stopping condition is met.

4. The method for monitoring gas leakage based on spectral vector data processing according to claim 3 is characterized in that: The step of inputting the feature-extracted data into the feature dimension reduction model training comprises: use Indicates encoder parameters, Represents the decoder parameters, setting the weights and biases of the encoder and decoder to random initialization; use represents the input feature vector, z represents the encoded low-dimensional vector, Represents the reconstructed output vector. The input feature vector is converted into a low-dimensional vector through the encoder. Represented, and then reconstructed into the output vector of the original dimension through the decoder, express; Let n represent the number of samples, and represents the regularization parameter, S represents the self-expression matrix, Represents a single sample, Represents the reconstructed sample, Represents the encoded sample, represents the dual loss, represents the reconstruction error, Denotes the sparse self-expression error, and uses Calculate the dual loss, the reconstruction error measures the error between the original input and the reconstructed output, and Calculate the error between the low-dimensional sample and the sparse linear combination after encoding by sparse self-expression error measurement, and use calculate; Let Z represent all coded samples The matrix of represents the hyperparameter controlling the density of the matrix, represents the Euclidean distance, Softmax( ) represents the Softmax function, constructs the self-expression matrix S, and uses and express; use represents the learning rate of the dual loss, Represents the dual loss about The gradient, Represents the dual loss about The gradient of , using the chain rule to calculate the dual loss About encoder parameters and decoder parameters The gradient of , update the parameters, use and express; The above process is iterated repeatedly until the preset stopping condition is met.

5. The method for monitoring gas leakage based on spectral vector data processing according to claim 4 is characterized in that: The step of converting the dimension-reduced feature vector into a quantum bit comprises: use represents the initialized quantum state, represents the first initial rotation angle parameter, represents the second initial rotation angle parameter, Represents the third initial rotation angle parameter, represents the initial state of n quantum bits, Represent the quantum gates of each quantum bit, including rotation and entanglement operations, initialize the quantum state encoding module, and use express; use Represents the rotation operation around the z-axis, represents the rotation operation around the y-axis, θ represents the first rotation angle parameter, ϕ represents the second rotation angle parameter, Represents the third rotation angle parameter, using a combination of basic quantum operations Constructing quantum gates; use Represents the quantum state after the feature vector qx is encoded. According to the input feature qx, the quantum encoding unit is designed and used Indicates that the reduced-dimensional feature vectors are converted into quantum states, each feature vector corresponds to a quantum state, and is expressed as express; use Denote the kth component of the eigenvector qx, using Computational quantum coding unit; Using the correlation measure of quantum states as the kernel function of the support vector machine, a quantum-enhanced classification model is constructed, and the overlap between two quantum states is calculated using the quantum classifier kernel function; use represents the quantum kernel function, represents the quantum state of the ith data point, Represents the quantum state of the jth data point, using Calculate the square of the inner product of two quantum states, reflecting the similarity of the two eigenvectors in the quantum state space; Train support vector machines using correlation measures between quantum states and compute classification boundaries using quantum simulations or quantum computers; 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 i-th sample label, represents the jth sample label, C represents the regularization parameter of the support vector machine, and Calculate the elements of the i-th row and j-th column of the matrix constructed by the quantum kernel function K, and let the classification function by Constraints; Let b denote the bias term of the decision boundary, sgn() denote the symbolic function and output classification prediction, and Represents the decision function of the support vector machine algorithm based on quantum information association.

6. The method for monitoring gas leakage based on spectral vector data processing according to claim 2 is characterized in that: Said The value is 0.

03. The value is 0.

01. The value is 0.

3. and The value is 2 and the number of iterations is 1000.

7. The method for monitoring gas leakage based on spectral vector data processing according to claim 3 is characterized in that: Said The value is 0.

01. Select cross entropy loss, The value is 0.

01. The value is 0.

5. Value 5, The value is 3.

14. The value is 0.

95. The value is 0.

01. The value is 0.95 and the number of iterations is 1000.

8. The method for monitoring gas leakage based on spectral vector data processing according to claim 4 is characterized in that: Said The value is 0.01 and the number of iterations is 1000.

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

Images