CN113780138B - Self-adaptive robustness VOCs gas leakage detection method, system and storage medium - Google Patents

Abstract

The invention relates to a self-adaptive robustness VOCs gas leakage detection method, a system and a storage medium, which comprises the following steps of acquiring infrared video data and carrying out preprocessing operation; extracting one-dimensional time sequence characteristic data of pixel points with a certain length from infrared video data, and training a one-dimensional convolutional neural network classifier; training parameter alpha of prior gamma distribution by using one-dimensional convolutional neural network classifier and EVT algorithm with output value led into Bayesian framework ₀ And beta ₀ (ii) a And inputting related parameters, adjusting a threshold value through a self-adaptive algorithm, and outputting a prediction result. The method fully utilizes the time-space characteristics of pixel points in VOCs gas regions in infrared video data to carry out pre-screening on the infrared video image by using the convolutional neural network, optimizes and adjusts the screening threshold value through an extreme value theory in a Bayesian framework, approaches the right tail part of a probability density function of a fraction by using index distribution, and uses gamma conjugate prior learned from training data, so that the variability of error rate can be reduced and the overall performance can be improved.

Description

Adaptive robust VOCs gas leak detection method, system and storage medium

technical field

The invention relates to the technical field of VOCs gas leak detection in the field of environmental detection, in particular to an adaptive robust VOCs gas leak detection method, system and storage medium based on extreme value theory.

Background technique

In recent years, with the rapid development of the petrochemical industry, the issue of production safety has become more and more important. The release of volatile organic compounds (VOCs) can lead to human health problems such as cancer, birth defects and reproductive effects. VOCs also lead to the formation of ozone, which is the main source of smog and one of the main causes of respiratory diseases in urban areas and areas near oil refineries and chemical plants. Therefore, the detection and treatment of VOCs has become a current air treatment problem. a focus.

Considering the absorption of infrared light by VOCs gas, the color of the VOCs leakage area on the infrared video data is darker than the surrounding area (white-hot mode). Infrared video data will be affected by factors such as light, temperature, humidity and climate. The complex imaging conditions make the detection of VOCs gas leaks less reliable. In summary, we propose an adaptive and robust VOCs gas leak detection method based on extreme value theory, extracting pixel spatiotemporal information in infrared video data for leak pre-screening, using extreme value theory in the Bayesian framework ( EVT) to optimally adjust the screening threshold, by approximating the right tail of the probability density function of the fraction with an exponential distribution (a special case of the generalized Pareto distribution), and using a gamma conjugate prior learned from the training data, which can be reduced Variability in error rates and improve overall performance. It aims to achieve robust detection of VOCs gas leakage, so as to adapt to VOCs gas leakage detection of various imaging conditions.

SUMMARY OF THE INVENTION

An adaptive robust VOCs gas leak detection method based on extreme value theory proposed by the invention can realize reliable VOCs leak detection under complex imaging conditions.

To achieve the above object, the present invention has adopted the following technical solutions:

Include the following steps,

Step 1: Obtain the data of the VOCs leakage area and the non-leakage area in the infrared video data for preprocessing;

Step 2: Extract one-dimensional time series feature data of a certain length of pixels from the infrared video data, and train a one-dimensional convolutional neural network classifier;

Step 3: Sample the spatiotemporal features of several pixels from the infrared video data multiple times, use a one-dimensional convolutional neural network classifier, import the output values into the EVT algorithm in the Bayesian framework, and train the parameters of the prior gamma distribution α ₀ and β ₀ ;

Step 4: Input the relevant parameters, adjust the threshold through the adaptive algorithm, and output the prediction result;

Wherein, the step 2: extracting a certain length of pixel point one-dimensional time series feature data from the infrared video data, and training a one-dimensional convolutional neural network classifier; specifically includes the following subdivision steps S21 to S23:

Step S21: Extract one pixel from every 8*8 or 16*16 block of the VOCs gas region of the segmented scene video frame with VOCs leakage to form one-dimensional time series leakage data of several pixels with length L (X ^L , 1 ), L is the number of scene frames, where 1 means that the data comes from an area with VOCs leakage, and ^XL = [x′ ₁ x′ ₂ ... x′ _L ] ^T ; at the same time, there is never a segmented scene with no VOCs leakage In the VOCs gas area in the same way, several one-dimensional time series normal data (X ^L , 0) of pixels with the same length are extracted, where 0 means that this data comes from the normal area;

Step S22: First, perform numerical normalization on the extracted one-dimensional time series data ^XL of pixels, so that each element of the one-dimensional time series data XL of pixels satisfies 0≤x′ ^i≤255 , _i =1, 2, . .., L, and then zero-average each one-dimensional time series data element x′ _i ; then divide the two types of data after processing, 80% as training data, and 20% as verification data;

Step S23: Use the processed training data to train a one-dimensional convolutional neural network classifier, the input of the one-dimensional convolutional neural network classifier is the one-dimensional time series data ^XL of pixels, and the output is D( ^XL ), where D( X ^L )∈(0, 1), stop training when the classification accuracy of the classifier on the validation data set reaches more than 98%, thus obtaining a one-dimensional convolutional neural network classification model;

Step 3: Sample the spatiotemporal features of several pixels from the infrared video data multiple times, use a one-dimensional convolutional neural network classifier, import the output value into the EVT algorithm in the Bayesian framework, and train the parameter α of the prior gamma distribution ₀ and β ₀ , specifically including the following subdivision steps S31 to S34:

Step S31: randomly extract a pixel from every 8*8 or 16*16 block of the dark part of the video frame of the segmented scene to be measured, obtain K one-dimensional time series data XL of pixels with length ^L , and send it into a stage-1 dimensional convolutional neural network classifier, and get the output D( ^XL ), where D( ^XL )∈(0,1);

Step S32: Construct the EVT training algorithm data set T by using K pieces of ^XL =[x' ₁ x' ₂ ... x' _L ] ^T as data x, and K pieces of D( ^XL ) as corresponding label sequences y;

Step S33: Select the negative sample g={x _i |y _i =0} from the data set T, find the upper threshold u of the negative sample g according to the right tail probability p _u , take out the right tail t through the upper threshold, and update the Sufficient statistics n and s for data marked as anomalies;

Step S34: Adjust the parameters α ₀ and β ₀ of the prior gamma distribution, expressed as

α ₀ =1+w ₀

where w ₀ is the weight assigned to the sample count of the training set;

Described step 4: input relevant parameters, adjust the threshold through the adaptive algorithm, and output the prediction result, which specifically includes the following subdivision steps S41 to S43:

Step S41: Divide the video frame of the scene to be tested, and use a one-dimensional convolutional neural network classifier to construct a data set; according to the method of step S33, find out the upper limit threshold of the data set, and take out all the samples of the right tail t1 to execute KolmogorovSmirnov test to find and eliminate anomalies, the KolmogorovSmirnov test method is

in

Calculate D _n , then remove the largest sample, and use the remaining samples to calculate Dn; keep looping, and finally select the label value of the sample that makes D _n the smallest and record it as

为阈值，提取出右尾部，使用训练期间的先验估计来计算整个序列的后验，更新为α₁和β₁；Step S42: After removing the abnormality, select the

is the threshold value, extract the right tail, use the prior estimation during training to calculate the posterior of the entire sequence, and update it to α ₁ and β ₁ ;

根据右尾部概率p_u查找窗口的上限阈值u2，提取右尾部，调整α、β和

通过计算得到y_j，表示为Step S43: Take the posterior calculated in S42 as the prior, and set the window W with the sample x _j as the center, expressed as

Find the upper threshold u2 of the window according to the right tail probability p _u , extract the right tail, adjust α, β and

y _j is obtained by calculation, which is expressed as

where p _f is the target error rate, and the label value y _j of all samples is continuously obtained in a loop.

Further, the above step 1: obtaining infrared video data with and without leakage of VOCs and performing data preprocessing, which specifically includes the following subdivided steps S11 to S12:

Step S11: Acquire infrared video data with and without leakage of VOCs;

Step S12: Perform preprocessing operations such as random rotation, frame size normalization, and scene segmentation on the infrared video data.

On the other hand, the present invention also discloses an adaptive and robust VOCs gas leak detection system based on extreme value theory, comprising the following units:

The data acquisition and processing unit is used to acquire the data of the VOCs leakage area and the non-leakage area in the infrared video data for preprocessing;

The one-dimensional network structure training unit is used to extract one-dimensional time series feature data of a certain length of pixels from infrared video data, and train a one-dimensional convolutional neural network classifier;

The parameter determination unit is used to sample the spatiotemporal features of several pixels from the infrared video data multiple times. It uses a one-dimensional convolutional neural network classifier, and the output value is imported into the EVT algorithm in the Bayesian framework to train the parameters of the prior gamma distribution. α ₀ and β ₀ ;

The prediction unit is used to input relevant parameters, adjust the threshold through the adaptive algorithm, and output the prediction result;

The specific processing steps of the one-dimensional network structure training unit are as follows:

Step S21: Extract one pixel from every 8*8 or 16*16 block of the VOCs gas region of the segmented scene video frame with VOCs leakage to form one-dimensional time series leakage data of several pixels with length L (X ^L , 1 ), L is the number of scene frames, where 1 means that the data comes from an area with VOCs leakage, and ^XL = [x′ ₁ x′ ₂ ... x′ _L ] ^T ; at the same time, there is never a segmented scene with no VOCs leakage In the VOCs gas area in the same way, several one-dimensional time series normal data (X ^L , 0) of pixels with the same length are extracted, where 0 means that this data comes from the normal area;

Step S22: First, perform numerical normalization on the extracted one-dimensional time series data ^XL of pixels, so that each element of the one-dimensional time series data XL of pixels satisfies 0≤x′ ^i≤255 , _i =1, 2, . .., L, and then zero-average each one-dimensional time series data element x′ _i ; then divide the two types of data after processing, 80% as training data, and 20% as verification data;

Step S23: Use the processed training data to train a one-dimensional convolutional neural network classifier, the input of the one-dimensional convolutional neural network classifier is the one-dimensional time series data ^XL of pixels, and the output is D( ^XL ), where D( X ^L )∈(0, 1), stop training when the classification accuracy of the classifier on the validation data set reaches more than 98%, thus obtaining a one-dimensional convolutional neural network classification model;

The specific processing steps of the parameter determination unit are as follows:

Step S31: randomly extract a pixel from every 8*8 or 16*16 block of the dark part of the video frame of the segmented scene to be measured, obtain K one-dimensional time series data XL of pixels with length ^L , and send it into a one-dimensional volume Integrate the neural network classifier to get the output D( ^XL ), where D( ^XL )∈(0,1);

Step S32: Construct the EVT training algorithm data set T by using K pieces of ^XL =[x' ₁ x' ₂ ... x' _L ] ^T as data x, and K pieces of D( ^XL ) as corresponding label sequences y;

Step S33: Select the negative sample g={x _i |y _i =0} from the data set T, find the upper threshold u of the negative sample g according to the right tail probability p _u , take out the right tail t through the upper threshold, and update the Sufficient statistics n and s for data marked as anomalies,

Step S34: Adjust the parameters α ₀ and β ₀ of the prior gamma distribution, expressed as

α ₀ =1+w ₀

where w ₀ is the weight assigned to the sample count of the training set;

The specific processing steps of the prediction unit are as follows:

Step S41: Divide the video frame of the scene to be tested, and use a one-dimensional convolutional neural network classifier to construct a data set; according to the method of step S33, find out the upper limit threshold of the data set, and take out all the samples of the right tail t1 to execute KolmogorovSmirnov test to find and eliminate anomalies, the KolmogorovSmirnov test method is

in

Calculate D _n , then remove the largest sample, and use the remaining samples to calculate Dn; keep looping, and finally select the label value of the sample that makes D _n the smallest and record it as

为阈值，提取出右尾部，使用训练期间的先验估计来计算整个序列的后验，更新为α₁和β₁；Step S42: After removing the abnormality, select the

is the threshold value, extract the right tail, use the prior estimation during training to calculate the posterior of the entire sequence, and update it to α ₁ and β ₁ ;

根据右尾部概率p_u查找窗口的上限阈值u2，提取右尾部，调整α、β和

通过计算得到y_j，表示为Step S43: Take the posterior calculated in S42 as the prior, and set the window W with the sample x _j as the center, expressed as

Find the upper threshold u2 of the window according to the right tail probability p _u , extract the right tail, adjust α, β and

y _j is obtained by calculation, which is expressed as

where p _f is the target error rate, and the label value y _j of all samples is continuously obtained in a loop.

In another aspect, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to execute the above-mentioned adaptive robustness VOCs gas based on extreme value theory Steps of a leak detection method.

It can be seen from the above technical solutions that the adaptive and robust VOCs gas leak detection method and system based on the extreme value theory of the present invention overcomes the shortcomings of the existing methods, and makes full use of the spatiotemporal characteristics of the pixels in the VOCs gas region in the infrared video data. Infrared video images are pre-screened by an integrated neural network, and the screening threshold is optimally adjusted by extreme value theory (EVT) within a Bayesian framework, by approximating the probability of the score with an exponential distribution (a special case of the generalized Pareto distribution). The right tail of the density function, and using a gamma-conjugate prior learned from the training data, reduces error rate variability and improves overall performance. So as to realize the robust detection of VOCs leakage.

Description of drawings

1 is a schematic diagram of the overall network model of the method of the present invention;

Figure 2 is a graph of the experimental results of the present invention.

Detailed ways

In order to make the purposes, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments These are some embodiments of the present invention, but not all embodiments.

As shown in FIG. 1 , the adaptive and robust VOCs gas leak detection method based on extreme value theory described in this embodiment includes the following steps:

Step 1: Obtain the data of the VOCs leakage area and the non-leakage area in the infrared video data for preprocessing;

Step 2: Extract one-dimensional time series feature data of a certain length of pixels from the infrared video data, and train a one-dimensional convolutional neural network classifier.

Step 3: Sample the spatiotemporal features of several pixels from the infrared video data multiple times, use a one-dimensional convolutional neural network classifier, import the output values into the EVT algorithm in the Bayesian framework, and train the parameters of the prior gamma distribution α ₀ and β ₀ .

Step 4: Input the relevant parameters, adjust the threshold through the adaptive algorithm, and output the prediction result.

The following specific instructions:

Further, the above step 1: obtain infrared video data with and without leakage of VOCs and perform data preprocessing. Specifically, it includes the following subdivision steps S11 to S12:

S11: Obtain infrared video data with and without leakage of VOCs.

S12: Perform preprocessing operations such as random rotation, frame size normalization, and scene segmentation on the infrared video data.

Further, in the above step 2: extracting one-dimensional time series feature data of pixels of a certain length from the infrared video data, and training a one-dimensional convolutional neural network classifier. Specifically, it includes the following subdivision steps S21 to S23:

S21: Extract one pixel from every 8*8 or 16*16 block of the dark part (VOCs gas region) of the segmented scene video frame with VOCs leakage to form a number of lengths L (the number of scene frames, in the present invention, L= 160) of the pixel point one-dimensional time series leakage data (X ^L , 1), where 1 represents that the data comes from the leakage area of VOCs, and ^XL =[x′ ₁ x′ ₂ ... x′ _L ] ^T ; A number of one-dimensional time series normal data (X ^L , 0) of pixels with the same length are extracted from the dark part of the segmented scene without VOCs leakage in the same way, where 0 means that the data comes from the normal area.

S22: First, perform numerical normalization on the extracted one-dimensional time series data ^XL of pixels, so that each element satisfies 0≤x′ _i ≤255, i=1, 2, . . . , L, and then for each element One-dimensional time series data elements _x'i are zero-averaged. Then, the two types of data after processing are divided, 80% are used as training data, and 20% are used as verification data.

S23: Use the processed training data to train a one-dimensional convolutional neural network classifier, the input of the one-stage classifier is the one-dimensional time series data ^XL of pixels, and the output is to obtain the output D( ^XL ), where D( ^XL ) ∈ (0, 1), stop training when the classifier achieves more than 98% classification accuracy on the validation dataset. Thus, a one-stage classification model such as Table1 is obtained;

Table 1

One-stage network structure

Further, the above-mentioned step 3: sampling the spatiotemporal features of several pixels from the infrared video data multiple times, using a one-dimensional convolutional neural network classifier, and importing the output value into the EVT algorithm in the Bayesian framework, training the prior gamma distribution. Parameters α ₀ and β ₀ . The EVT training algorithm process is shown in Table 2;

Specifically, it includes the following subdivision steps S31 to S34:

S31: Randomly extract a pixel from every 8*8 or 16*16 block of the dark part of the video frame of the segmented scene to be tested, obtain K one-dimensional time series data XL of pixels with length ^L (the number of scene frames), and send it to Enter the one-stage one-dimensional convolutional neural network to obtain the output D( ^XL ), where D( ^XL )∈(0, 1).

S32: Construct the EVT training algorithm data set T by taking K pieces of ^XL =[x' ₁ x' ₂ ... x' _L ] ^T as data x, and K pieces of D( ^XL ) as corresponding label sequences y.

S33: Select the negative sample g={x _i |y _i =0} from the data set T, and substitute it into the formula { _gi >u}=gp _u , which is expressed as finding negative samples according to the right tail probability p _u Upper threshold u for sample g, where parameter p _u is the probability of the right tail. The right tail t is taken out by the upper threshold, and the sufficient statistics n and s of the data not marked as abnormal are updated, which can be expressed as

_nj+1 = _nj +t

s _j+1 =s _j +∑t

Among them, n ₀ and s ₀ are initialized to 0.

S34: Adjust the parameters α ₀ and β ₀ of the prior gamma distribution, which can be expressed as

α ₀ =1+w ₀

where w ₀ is the weight assigned to the sample count of the training set.

Table 2

EVT training algorithm details

Further, the above-mentioned step 4: input relevant parameters, adjust the threshold through the adaptive algorithm, and output the prediction result, wherein the details of the EVT adaptive threshold algorithm are shown in Table 3;

Specifically, it includes the following subdivision steps S41 to S43:

S41: According to step S33, find the upper threshold, take out all the samples in the tail and perform a series of Kolmogorov Smirnov (KS) tests to find and eliminate abnormalities. The video frame of the scene to be tested is divided, and a one-dimensional convolutional neural network classifier is used to construct a data set; according to the method of step S33, the upper threshold of the data set is found, and all samples in the right tail t1 are taken out to perform the KS test, KS The test method is

in

最后选择使D_n，i最小的i的值记为

S42：在去除异常后，选择以

为阈值，根据步骤S33，提取出右尾部，使用训练期间的先验估计来计算整个序列的后验，更新为α₁和β₁。 _Dn is calculated or referred to as _Dn,1 . Then, the largest sample is removed, and _Dn,2 is calculated using the remaining samples. Iterate continuously until you get

Finally, select the value of i that minimizes D _{n, i} and is recorded as

S42: After removing the exception, select to

As the threshold, according to step S33, the right tail is extracted, and the posterior of the entire sequence is calculated using the prior estimation during training, and updated to α ₁ and β ₁ .

根据{w_i＞u}＝wp_u查找上限阈值u2，根据S33，提取右尾部，调整α、β和

通过计算得到y_j，表示为S43: Use the posterior calculated in S42 as the prior. Sets the sample-centered window, which is represented as

Find the upper threshold u2 according to { _wi > u}=wp _u , according to S33, extract the right tail, adjust α, β and

y _j is obtained by calculation, which is expressed as

where p _f is the target error rate. Continuous iteration can obtain an adjusted score y _j , ie, an adaptive threshold is achieved.

Table 3

EVT Adaptive Thresholding Algorithm Details

Figure 2 shows a frame with VOCs leakage in the infrared video. This frame is not recognized as a VOCs leakage frame in the one-dimensional convolutional neural network. Through the EVT threshold adaptive algorithm, the right tail is extracted to calculate y _j , where red The box is the leaked area. It can be seen from this that the method of the present invention can effectively detect VOCs gas leakage, and in different scenarios, the model with EVT algorithm will always have higher detection performance than single-dimensional CNN.

To sum up, the method of the present invention based on the extreme value theory for adaptive and robust VOCs gas leak detection, using a one-dimensional convolutional neural network can achieve rapid pre-detection of VOCs gas leaks while reducing the amount of calculation, so that the algorithm can adapt to On devices with limited performance; adopting the EVT threshold adaptive method can improve the robustness of VOCs leakage detection in infrared video data, and reduce the impact of false identification caused by factors such as illumination, temperature, and climate.

On the other hand, the present invention also discloses an adaptive and robust VOCs gas leak detection system based on extreme value theory, comprising the following units:

The data acquisition and processing unit is used to acquire the data of the VOCs leakage area and the non-leakage area in the infrared video data for preprocessing;

The one-dimensional network structure training unit is used to extract one-dimensional time series feature data of a certain length of pixels from infrared video data, and train a one-dimensional convolutional neural network classifier;

The parameter determination unit is used to sample the spatiotemporal features of several pixels from the infrared video data multiple times. It uses a one-dimensional convolutional neural network classifier, and the output value is imported into the EVT algorithm in the Bayesian framework to train the parameters of the prior gamma distribution. α ₀ and β ₀ ;

The prediction unit is used to input relevant parameters, adjust the threshold through the adaptive algorithm, and output the prediction result.

Further, the specific processing steps of the one-dimensional network structure training unit are as follows:

S21: Extract one pixel from every 8*8 or 16*16 block of the VOCs gas region of the segmented scene video frame with VOCs leakage to form one-dimensional time series leakage data of several pixels with length L (X ^L , 1) , L is the number of scene frames, where 1 means that the data comes from the area with VOCs leakage, and ^XL = [x′ ₁ x′ ₂ ... x′ _L ] ^T ; at the same time, it also never has VOCs leakage in the segmentation scene The VOCs gas area extracts several one-dimensional time series normal data (X ^L , 0) of pixels with the same length in the same way, where 0 represents that the data comes from the normal area;

S22: First, perform numerical normalization on the extracted one-dimensional time series data ^XL of pixels, so that each element of the one-dimensional time series data XL of pixels satisfies 0≤x′ ^i≤255 , _i =1, 2, .. , L, and then zero-average each one-dimensional time series data element x′ _i ; then divide the two types of data after processing, 80% as training data, and 20% as verification data;

S23: Use the processed training data to train a one-dimensional convolutional neural network classifier. The input of the one-stage classifier is the one-dimensional time series data ^XL of pixels, and the output is D( ^XL ), where D( ^XL )∈( 0, 1), when the classification accuracy rate of the classifier on the validation data set reaches more than 98%, the training is stopped, so as to obtain a one-stage classification model.

Further, the specific processing steps of the parameter determination unit are as follows:

S31: Randomly extract one pixel from every 8*8 or 16*16 block in the dark part of the video frame of the segmented scene to be tested, obtain K one-dimensional time series data XL of pixels with length ^L , and send it into one-stage one-dimensional Convolutional Neural Network, get the output D( ^XL ), where D( ^XL )∈(0,1);

S32: Construct the EVT training algorithm dataset T by using K ^XL = [x' ₁ x' ₂ ... x' _L ] ^T as data x, and K D(X ^L ) as corresponding label sequence y;

S33: Select the negative sample g={x _i |y _i =0} from the data set T, and substitute it into the formula { _gi >u}=gp _u to find the upper threshold u, where the parameter p _u is the right tail Probability, taking out the right tail t through the upper threshold, updating the sufficient statistics n and s of the data not marked as abnormal, expressed as

_nj+1 = _nj +t

s _j+1 =s _j +∑t

Among them, n ₀ and s ₀ are initialized to 0;

S34: Adjust the parameters α ₀ and β ₀ of the prior gamma distribution, expressed as

α ₀ =1+w ₀

where w ₀ is the weight assigned to the sample count of the training set.

Further, the specific processing steps of the prediction unit are as follows:

S41: According to step S33, find the upper limit threshold, take out all the samples at the tail, and then divide the video frame of the scene to be tested, and use a one-dimensional convolutional neural network to construct a data set; According to the method of step S33, find the upper limit threshold of the data set , and take out all samples of the right tail t1 to perform a series of KolmogorovSmirnov (KS) tests to find and eliminate anomalies, KS is

in

最后选择使D_n，i最小的i的值记为

Calculate D _n or call it D _{n, 1} , then remove the largest sample and use the remaining samples to calculate D _{n, 2} , and iterate until you get

Finally, select the value of i that minimizes D _{n, i} and is recorded as

为阈值，提取出右尾部，使用训练期间的先验估计来计算整个序列的后验，更新为α₁和β₁；S42: After removing the exception, select to

is the threshold value, extract the right tail, use the prior estimation during training to calculate the posterior of the entire sequence, and update it to α ₁ and β ₁ ;

根据右尾部概率p_u查找窗口的上限阈值u2，提取右尾部，调整α、β和

通过计算得到y_j，表示为S43: Take the posterior calculated in S42 as the prior, and set the window W with the sample x _j as the center, expressed as

Find the upper threshold u2 of the window according to the right tail probability p _u , extract the right tail, adjust α, β and

y _j is obtained by calculation, which is expressed as

where p _f is the target error rate, and iteratively obtains the adjusted score y _j , that is, the adaptive threshold is realized.

The present invention also discloses a computer-readable storage medium, which stores a computer program, and when the computer program is executed by a processor, causes the processor to execute the steps of the above method.

It is understandable that the system provided by the embodiment of the present invention corresponds to the method provided by the embodiment of the present invention, and reference may be made to the corresponding part of the above-mentioned method for explanation, examples and beneficial effects of related content.

As will be appreciated by those skilled in the art, the embodiments of the present application may be provided as a method, a system, or a computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) having computer-usable program code embodied therein.

The present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the present application. It will be understood that each process and/or block in the flowchart illustrations and/or block diagrams, and combinations of processes and/or blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to the processor of a general purpose computer, special purpose computer, embedded processor or other programmable data processing device to produce a machine such that the instructions executed by the processor of the computer or other programmable data processing device produce Means for implementing the functions specified in a flow or flow of a flowchart and/or a block or blocks of a block diagram.

These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory result in an article of manufacture comprising instruction means, the instructions The apparatus implements the functions specified in the flow or flow of the flowcharts and/or the block or blocks of the block diagrams.

These computer program instructions can also be loaded on a computer or other programmable data processing device to cause a series of operational steps to be performed on the computer or other programmable device to produce a computer-implemented process such that The instructions provide steps for implementing the functions specified in the flow or blocks of the flowcharts and/or the block or blocks of the block diagrams.

The above embodiments are only used to illustrate the technical solutions of the present invention, but not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: The recorded technical solutions are modified, or some technical features thereof are equivalently replaced; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims (4)

1. A self-adaptive robustness VOCs gas leakage detection method based on an extreme value theory is characterized by comprising the following steps: comprises the following steps of (a) carrying out,

step 1: acquiring data of VOCs leakage areas and non-leakage areas in infrared video data to carry out preprocessing operation;

step 2: extracting one-dimensional time sequence characteristic data of pixel points with a certain length from infrared video data, and training a one-dimensional convolutional neural network classifier;

and step 3: sampling the space-time characteristics of a plurality of pixel points from the infrared video data for multiple times, using a one-dimensional convolution neural network classifier, leading the output value into an EVT algorithm in a Bayesian framework, and training the parameter alpha of prior gamma distribution ₀ And beta ₀ ；

And 4, step 4: inputting relevant parameters, adjusting a threshold value through a self-adaptive algorithm, and outputting a prediction result;

wherein the step 2: extracting one-dimensional time sequence characteristic data of pixel points with a certain length from infrared video data, and training a one-dimensional convolutional neural network classifier; the method specifically comprises the following subdivision steps S21-S23:

step S21: extracting one pixel from each 8X 8 or 16X 16 block of VOCs gas area of segmented scene video frame with VOCs leakage to form a plurality of pixel point one-dimensional time sequence leakage data (X) with length L ^L 1), L is the number of scene frames, where 1 represents that the data comes from the area where VOCs are leaking, and X is ^L ＝[x′ ₁ x′ ₂ ...x′ _L ] ^T (ii) a Meanwhile, a plurality of pixel point one-dimensional time sequence normal data (X) with the same length are extracted from VOCs gas regions in segmentation scenes without VOCs leakage in the same way ^L 0), where 0 represents that the data is from a normal region;

step S22: firstly, one-dimensional time sequence data X of pixel points obtained by extraction is subjected to ^L Carrying out numerical value normalization to enable one-dimensional time sequence data X of pixel points ^L Each element satisfies 0 ≦ x' _i 1, 2, L followed by x 'for each one-dimensional time series data element' _i Carrying out zero equalization; then, the two types of data after the processing are respectively segmented, wherein 80% of the data are used as training data, and 20% of the data are used as verification data;

step S23: training a one-dimensional convolutional neural network classifier by using the processed training data, wherein the input of the one-dimensional convolutional neural network classifier is pixel point one-dimensional time sequence data X ^L The output is D (X) ^L ) Wherein D (X) ^L ) E (0, 1), stopping training when the classification accuracy of the classifier on the verification data set reaches more than 98%, thereby obtaining a one-dimensional convolution neural network classification model;

the step 3: sampling the space-time characteristics of a plurality of pixel points from infrared video data for a plurality of times, using a one-dimensional convolution neural network classifier, leading the output value into an EVT algorithm in a Bayesian framework, and training the parameter alpha of prior gamma distribution ₀ And beta ₀ Specifically, the method comprises the following subdivision steps S31 to S34:

step S31: randomly extracting one pixel from each 8 × 8 or 16 × 16 block of the dark part of the segmented scene video frame to be detected to obtain K pixel point one-dimensional time sequence data X with length L ^L Sending the data to a one-stage one-dimensional convolution neural network classifier to obtain an output D (X) ^L ) Wherein D (X) ^L )∈(0，1)；

Step S32: mixing K X ^L ＝[x′ ₁ x′ ₂ ...x′ _L ] ^T As data X, K D (X) ^L ) Is corresponding toConstructing an EVT training algorithm data set T by using a label sequence y;

step S33: picking out negative samples g ═ x from the dataset T _i |y _i 0, according to the right tail probability p _u Searching an upper limit threshold u of the negative sample g, taking out the right tail t through the upper limit threshold, and updating sufficient statistics n and s of data which are not marked as abnormal;

step S34: adjusting a parameter alpha of a prior gamma distribution ₀ And beta ₀ Is shown as

α ₀ ＝1+w ₀

Wherein, w ₀ Is the weight assigned to the sample count of the training set;

the step 4: inputting relevant parameters, adjusting a threshold value through an adaptive algorithm, and outputting a prediction result, wherein the method specifically comprises the following subdivision steps S41-S43:

step S41: constructing a data set by using a one-dimensional convolution neural network classifier on a to-be-detected segmented scene video frame; according to the mode of step S33, an upper threshold value of the data set is found, all samples of the right tail t1 are taken out to carry out a KolmogorovSmirnov test to find and eliminate the abnormity, and the KolmogorovSmirnov test method is that

Wherein

Is calculated to obtain D _n Then, the largest sample is removed and Dn is calculated using the remaining samples; continuously circulating, and finally selecting D _n The label value of the smallest sample is noted

Step S42: after removing the anomaly, selecting to

For the threshold, the right tail is extracted, the a priori estimate during training is used to calculate the a posteriori of the whole sequence, updated to α ₁ And beta ₁ ；

Step S43: the posterior obtained by calculation in S42 is used as a prior, and a sample x is set _j A window W as a center, denoted by

According to the right tail probability p _u Searching the upper limit threshold u2 of the window, extracting the right tail part, and adjusting alpha, beta and

by calculating to obtain y _j Is shown as

Wherein p is _f Continuously and circularly obtaining the label values y of all samples for the target error rate _j 。

2. The extremum theory-based adaptive robust VOCs gas leak detection method of claim 1, wherein: the step 1: acquiring infrared video data with and without leakage of VOCs and preprocessing the data, wherein the method specifically comprises the following subdivision steps S11-S12:

step S11: acquiring infrared video data with VOCs leakage and no leakage;

step S12: and carrying out preprocessing operations of random rotation, frame size normalization and scene segmentation on the infrared video data.

3. The utility model provides a self-adaptation robustness VOCs gas leak detection system based on extreme value theory which characterized in that: comprises the following units which are connected with each other,

the data acquisition and processing unit is used for acquiring data of VOCs leakage areas and non-leakage areas in the infrared video data to carry out preprocessing operation;

the one-dimensional network structure training unit is used for extracting one-dimensional time sequence characteristic data of pixel points with a certain length from the infrared video data and training a one-dimensional convolutional neural network classifier;

a parameter determination unit for sampling the space-time characteristics of a plurality of pixel points from the infrared video data for a plurality of times, using a one-dimensional convolution neural network classifier, leading the output value into an EVT algorithm in a Bayesian framework, and training the parameter alpha of the prior gamma distribution ₀ And beta ₀ ；

The prediction unit is used for inputting relevant parameters, adjusting a threshold value through a self-adaptive algorithm and outputting a prediction result;

the one-dimensional network structure training unit comprises the following specific processing steps:

step S21: extracting one pixel from each 8X 8 or 16X 16 block of VOCs gas area of segmented scene video frame with VOCs leakage to form a plurality of pixel point one-dimensional time sequence leakage data (X) with length L ^L 1), L is the number of scene frames, where 1 represents that the data comes from the area where VOCs are leaking, and X is ^L ＝[x′ ₁ x′ ₂ ...x′ _L ] ^T (ii) a Meanwhile, a plurality of pixel point one-dimensional time sequence normal data (X) with the same length are extracted from VOCs gas regions in segmentation scenes without VOCs leakage in the same way ^L 0), where 0 represents that the data is from a normal region;

step S22: firstly, extracting the obtained one-dimensional time sequence data X of the pixel points ^L Carrying out numerical value normalization to enable one-dimensional time sequence data X of pixel points ^L Each element satisfies 0 ≦ x' _i 1, 2, L followed by x 'for each one-dimensional time series data element' _i Carrying out zero equalization; then, the two types of data after the processing are respectively segmented, wherein 80% of the data are used as training data, and 20% of the data are used as verification data;

step S23: training a one-dimensional convolutional neural network classifier by using the processed training data, wherein the input of the one-dimensional convolutional neural network classifier is pixel point one-dimensional time sequence data X ^L The output is D (X) ^L ) Wherein D (X) ^L ) E (0, 1), stopping training when the classification accuracy of the classifier on the verification data set reaches more than 98%, thereby obtaining a one-dimensional convolution neural network classification model;

the parameter determination unit comprises the following specific processing steps:

step S31: randomly extracting one pixel from each 8X 8 or 16X 16 block of the dark part of the segmented scene video frame to be detected to obtain K pixel point one-dimensional time sequence data X with the length L ^L Sending the data into a one-dimensional convolutional neural network classifier to obtain an output D (X) ^L ) Wherein D (X) ^L )∈(0，1)；

Step S32: mixing K X ^L ＝[x′ ₁ x′ ₂ ...x′ _L ] ^T As data X, K D (X) ^L ) Constructing an EVT training algorithm data set T for the corresponding label sequence y;

step S33: picking out negative samples g ═ x from the dataset T _i |y _i 0, according to the right tail probability p _u Searching an upper threshold u of the negative sample g, taking out a right tail t through the upper threshold, updating sufficient statistics n and s of data which are not marked as abnormal,

step S34: adjusting a parameter alpha of a prior gamma distribution ₀ And beta ₀ Is shown as

α ₀ ＝1+w ₀

Wherein, w ₀ Is the weight assigned to the sample count of the training set;

the prediction unit comprises the following specific processing steps:

step S41: constructing a data set by using a one-dimensional convolution neural network classifier on a to-be-detected segmented scene video frame; according to the mode of the step S33, an upper threshold value of the data set is found, all samples of the right tail t1 are taken out, and a KolmogorovSmirnov test is carried out to find and eliminate the abnormity, wherein the KolmogorovSmirnov test method is that

Wherein

Is calculated to obtain D _n Then, the largest sample is removed and Dn is calculated using the remaining samples; continuously circulating, and finally selecting D _n The label value of the smallest sample is noted

Step S42: after removing the anomaly, selecting to

For the threshold, the right tail is extracted, the a priori estimate during training is used to calculate the a posteriori of the whole sequence, updated to α ₁ And beta ₁ ；

Step S43: the posterior obtained by calculation in S42 is used as a prior, and a sample x is set _j A window W as a center, denoted by

According to the right tail probability p _u Searching the upper limit threshold u2 of the window, extracting the right tail part, and adjusting alpha, beta and

by calculating to obtain y _j Is shown as

Wherein p is _f Continuously and circularly obtaining the label values y of all samples for the target error rate _j 。

4. A computer readable storage medium storing a computer program which, when executed by a processor, causes the processor to perform the steps of the extremum theory based adaptive robust VOCs gas leak detection method of claim 1 or 2.

Images