CN111401126A - A method and device for monitoring combustion conditions based on flame images and deep learning - Google Patents

Abstract

The invention discloses a combustion condition monitoring method and a device based on flame images and deep learning, wherein the method comprises the following steps of firstly, collecting flame images under different combustion conditions by using an image monitoring device; preprocessing the flame image and then dividing the flame image into a training set, a verification set and a test set; establishing convolution sparse self-coding, and performing unsupervised training by using a training set; the trained convolution sparse self-coding is used for extracting deep features of the verification set image; establishing a Soft-max classifier, and performing supervision training by using deep features and labels of the verification set images; the trained Soft-max classifier can accurately classify deep features of the image, so that the combustion working condition recognition is realized; and (4) carrying out performance inspection on the monitoring method by using the test set label image. The monitoring device provided by the invention can stably operate on a high-temperature boiler to realize flame image acquisition, and the monitoring method can accurately identify the combustion working condition of a single flame image, and has a good application prospect in the field of combustion working condition monitoring.

Description

A method and device for monitoring combustion conditions based on flame images and deep learning

technical field

The invention relates to a combustion condition monitoring method and device based on flame images and deep learning, and belongs to the field of combustion detection.

Background technique

The new energy power generation system is limited by the randomness and intermittency of renewable resources, and large-scale grid connection will have a serious negative impact on the stable operation of the power grid. In order to provide enough space for new energy to be connected to the grid, it is necessary to carry out flexible transformation of thermal power generating units to improve the ability of deep peak shaving. Based on this, the power plant boilers may be under frequent and rapid variable load or long-term low-load conditions for a long time, which may lead to difficult ignition of pulverized coal, unstable combustion, and even fire extinguishing of the boiler, making combustion adjustment difficult and combustion efficiency reduced. The safe and stable operation of thermal power generating units brings challenges. Therefore, establishing an accurate and reliable combustion state monitoring system is of great significance for preventing potential dangers and improving the overall performance of boiler operation.

The flame imaging monitoring method, different from the traditional flame detection device, has the advantages of anti-interference and low maintenance, and is a more effective monitoring method. The method usually includes two important steps: feature extraction and state recognition. Feature extraction is to use image processing technology to obtain key features; state recognition is to further analyze the obtained image features to realize combustion state recognition. At present, common image processing techniques are mainly divided into two categories: 1) frequency domain and time domain methods, by analyzing the variance and power spectral density of the gray value change of the image, infer the change trend of the combustion state. This method has the shortcomings of long response time and weak generalization ability; 2) machine learning method, using principal component analysis or partial least squares method to extract image features in an unsupervised manner. However, the obtained features belong to the shallow information of the image and lack robustness, resulting in low recognition accuracy.

The deep learning method is considered to be a major breakthrough in the field of artificial intelligence. Its deep network structure can mine the essential characteristics of the data, which provides the possibility for difficult problems that cannot be solved by traditional data-driven methods. The deep learning network can not only overcome the shortcomings of poor performance of shallow methods, but also avoid the tedious process of feature selection, and can be used to process high-dimensional and large-scale image data. Autoencoder is one of the typical deep learning networks that can extract nonlinear features of images in an unsupervised manner. However, during the training process of self-encoding, gradient disappearance or gradient explosion is easy to occur. There is the possibility of simply copying the information of the input layer, and the key information of the image cannot be really obtained. Therefore, it is still necessary to develop a more intelligent flame image monitoring and identification method to provide guidance for combustion adjustment and optimization, making it more valuable for industrial use.

SUMMARY OF THE INVENTION

In order to solve the above technical problems, the purpose of the present invention is to provide a combustion condition monitoring method and device based on flame images and deep learning.

A combustion condition monitoring method based on flame images and deep learning of the present invention includes the following steps:

Step 1. Use an image monitoring device to collect images of furnace flames and record the combustion conditions. The collected images are preprocessed and divided into training sets, validation sets and test sets. The images in the training set do not have labels, while the validation set and the images in the test set are labeled;

Step 2. Unsupervised training of convolutional sparse auto-encoding with training set images; the combination of sparse penalty term, mean square error and structural similarity as a loss function, expressed as:

L=L _SPT +L _MSE +L _SSIM

In the formula, L _SPT represents the sparse penalty term, L _MSE represents the mean square error, and L _SSIM represents the structural similarity;

Step 3. Use the CSAE network that has completed the training to extract the deep features of the validation set images, and train the Soft-max classifier in combination with the image labels;

Step 4. The trained Soft-max classifier can classify the new image features, realize the combustion condition recognition of a single flame image, and use the test set to test its performance.

An image monitoring device for a combustion condition monitoring method based on flame images and deep learning, the image monitoring device is composed of a cooling jacket, an optical sight glass and an industrial camera, the optical sight glass is connected with the industrial camera, and the cooling The sleeve is sleeved on the outside of the industrial camera and the optical sight glass;

The top of the cooling jacket is designed with a 45° corner, and the cooling jacket is divided into a water interlayer and an air interlayer. On the inner side, the water and cold water interlayer is on the outer side, and a cooling water inlet and a cooling air inlet are arranged at the end of the cooling jacket, and a cooling air outlet is arranged at the top of the cooling jacket, wherein the cooling water inlet is communicated with the water interlayer, and the cooling air inlet and The cooling air outlet is communicated with the air interlayer.

The top of the optical sight glass is equipped with a high temperature resistant lens with a viewing angle of 90°;

The end of the industrial camera is provided with a power interface and a video signal interface.

Further, the feature extraction steps of the convolutional sparse self-encoding network in the step 2 are:

Step 2.1. Send the training set X to the convolutional coding. First, perform feature extraction by k1 convolutional filters (C1(k1@a1×a1+s1)) with a1×a1 window size and step size s1, and then use The ReLU function (y(x)=max(0,x)) performs neuron activation, and finally the feature reduction is performed by the maximum pooling layer (P1(b1×b1+f1)) with a window size of b1×b1 and a step size of f1 dimension, get the feature vector h ₁ ;

Step 2.2, the feature vector h ₁ is sent to the encoder No. 2 and the encoder No. 3 for further processing, the processing process is similar to that of the encoder No. 1, and finally the feature vector h ₃ is obtained;

Step _2.3 . The feature vector h3 is sent to the decoder No. 1. First, the feature dimension is improved by the upsampling layer of the g1×g1 window size (U1(g1×g1)), and then the convolution filter (C4(k4@ a4×a4+s4)), ReLU activation function processing, to obtain feature vector h ₄ ;

Step 2.4, the feature vector h ₄ is sent to the decoder No. 2 and the decoder No. 3 for further processing. The processing process is similar to that of the decoder No. 1, and finally the reconstructed image Z is obtained.

Further, the sparse penalty term L _SPT in the step 2 is expressed as:

In the formula, β represents the sparsity rate; F represents the number of neurons in the hidden layer; p represents the sparse target constant; the average activation of neurons based on the i-axis p _j is expressed as:

In the formula, E represents the number of images in the training set; s _ij (i∈(1, E), j∈(1, F)) represents the activation of the neuron at the jth position; the KL divergence is expressed as:

Further, the loss function L _MSE in the step 2 is expressed as:

In the formula, X _ij and Z _ij represent the grayscale of the (i, j)th position in the input image and the reconstructed image with size A×T, respectively.

Further, the loss function L _SSIM in the step 2 is expressed as:

In the formula, c ₁ =(k ₁ r) ² and c ₂ =(k ₂ r) ² , k ₁ and k ₂ are constants less than 1, r represents the dynamic range of image grayscale; μ _X and μ _Z represent the input The average value of the image and the reconstructed image; σ _X and σ _Z represent the variance of the input image and the reconstructed image; σ _XZ represents the covariance of the input image and the reconstructed image.

Further, the Soft-max classifier in the step 3 is defined as:

In the formula, _yi represents the output of the _ith sample xi; k represents the number of categories of training samples.

By means of the above scheme, the present invention has at least the following advantages:

Compared with the existing combustion condition monitoring technology, a combustion condition monitoring method and device based on flame images and deep learning of the present invention have the following advantages:

1. The image monitoring device provided by the present invention has an efficient cooling effect, which provides a guarantee for the stable operation of the image acquisition system;

2. The monitoring method provided by the present invention can accurately extract the essential features of the flame image without prior knowledge of experts; the new self-encoding loss function effectively solves the shortcomings of difficult training and poor generalization performance; and the Soft-max classifier adopted It is suitable for image feature classification and can realize accurate identification of combustion conditions.

The above description is only an overview of the technical solution of the present invention. In order to understand the technical means of the present invention more clearly, and implement it according to the content of the description, the preferred embodiments of the present invention are described in detail below with the accompanying drawings.

Description of drawings

In order to illustrate the technical solutions of the embodiments of the present invention more clearly, the following briefly introduces the accompanying drawings that need to be used in the embodiments. It should be regarded as a limitation of the scope, and for those of ordinary skill in the art, other related drawings can also be obtained according to these drawings without any creative effort.

FIG. 1 is a schematic flow chart of the principle of a combustion condition monitoring method and device based on flame images and deep learning of the present invention.

FIG. 2 is a schematic diagram of an image monitoring device in a method and device for monitoring combustion conditions based on flame images and deep learning.

3 is a schematic diagram of the structure of a convolutional sparse self-encoding (CSAE) network in a method and device for monitoring combustion conditions based on flame images and deep learning of the present invention.

FIG. 4 is a schematic diagram of flame images of three combustion conditions according to an embodiment of the present invention.

FIG. 5 is a data set structure according to an embodiment of the present invention.

FIG. 6 is a schematic diagram of training loss of a convolutional sparse autoencoder (CSAE) network according to an embodiment of the present invention.

Among them, in Figure 2, 1. Cooling sleeve; 2. Optical sight glass; 3. Industrial camera; 4. Water interlayer; 5. Air interlayer; 6. Cooling water inlet; 7. Cooling air inlet; 8. Cooling air outlet ; 9. Power interface; 10. Video signal interface.

Detailed ways

The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and embodiments. The following examples are intended to illustrate the present invention, but not to limit the scope of the present invention.

Referring to Figure 1 to Figure 6, a preferred embodiment of the present invention is described below in conjunction with the accompanying drawings and the accompanying tables, and the present invention is described in detail:

A combustion condition monitoring device and method based on flame image recognition, the principle flow chart of which is shown in Figure 1, and specifically includes the following steps:

Step 1. Use an image monitoring device (as shown in Figure 2) to collect images of furnace flames, and record combustion conditions. The collected images are divided into training set, validation set and test set after preprocessing, wherein, the images in the training set are unlabeled, while the images in the validation set and test set are labeled;

Step 2. Unsupervised training of convolutional sparse autoencoder (CSAE) using training set images, and its network structure is shown in Figure 3; the sparse penalty term (SPT), mean square error (MSE) and structural similarity (SSIM) Combined as a loss function, expressed as:

L=L _SPT + L _MSE + L _SSIM (1)

In the formula, L _SPT represents the sparse penalty term, L _MSE represents the mean square error, and L _SSIM represents the structural similarity;

Step 3. Use the CSAE network that has completed the training to extract the deep features of the validation set images, and train the Soft-max classifier in combination with the image labels;

Step 4. The trained Soft-max classifier can classify the new image features, realize the combustion condition recognition of a single flame image, and use the test set to test its performance.

The image monitoring device in the step 1:

Consists of cooling jacket, optical sight glass and industrial camera. Different from the traditional structure, the cooling jacket adopts a unique 45° corner design, which effectively overcomes the limitations of the installation position, and has two cooling methods, air cooling and water cooling, to ensure the stable operation of the image acquisition system. The optical sight glass is equipped with a high temperature resistant lens with a viewing angle of 90°, which completely covers the main combustion reaction area.

The feature extraction steps of the convolutional sparse self-encoding (CSAE) network in the step 2 are:

Step 2.1. Send the training set X to the convolutional coding. First, perform feature extraction by k1 convolutional filters (C1(k1@a1×a1+s1)) with a window size of a1×a1 and a stride of s1, and then use The ReLU function (y(x)=max(0,x)) performs neuron activation, and finally the feature reduction is performed by the maximum pooling layer (P1(b1×b1+f1)) with a window size of b1×b1 and a step size of f1 dimension, get the feature vector h ₁ ;

Step 2.2, the feature vector h ₁ is sent to the encoder No. 2 and the encoder No. 3 for further processing, the processing process is similar to that of the encoder No. 1, and finally the feature vector h ₃ is obtained;

Step _2.3 . The feature vector h3 is sent to the decoder No. 1. First, the feature dimension is improved by the upsampling layer of the g1×g1 window size (U1(g1×g1)), and then the convolution filter (C4(k4@ a4×a4+s4)), ReLU activation function processing, to obtain feature vector h ₄ ;

Step 2.4, the feature vector h ₄ is sent to the decoder No. 2 and the decoder No. 3 for further processing. The processing process is similar to that of the decoder No. 1, and finally the reconstructed image Z is obtained; it is worth noting that the activation function of the decoder No. 3 adopts Sigmoid function (y(x)=1/1+e ^-x ) to ensure that the input image and the output image have the same range of values.

The sparse penalty term L _SPT in the step 2 is expressed as:

In the formula, β represents the sparsity rate; F represents the number of neurons in the hidden layer; p represents the sparse target constant; the average activation of neurons based on the i-axis p _j is expressed as:

In the formula, E represents the number of images in the training set; s _ij (i∈(1, E), j∈(1, F)) represents the activation of the neuron at the jth position; the KL divergence is expressed as:

The loss function L _MSE in the step 2 is expressed as:

In the formula, X _ij and Z _ij represent the grayscale of the (i, j)th position in the input image and the reconstructed image with size A×T, respectively.

The loss function L _SSIM in the step 2 is expressed as:

In the formula, c ₁ =(k ₁ r) ² and c ₂ =(k ₂ r) ² , k ₁ and k ₂ are constants less than 1, r represents the dynamic range of image grayscale; μ _X and μ _Z represent the input The average value of the image and the reconstructed image; σ _X and σ _Z represent the variance of the input image and the reconstructed image; σ _XZ represents the covariance of the input image and the reconstructed image.

The Soft-max classifier in step 3 is defined as:

In the formula, _yi represents the output of the _ith sample xi; k represents the number of categories of training samples.

Example 1

The combustion condition monitoring device and method based on flame image recognition of the present invention comprises the following steps:

Step 1) Using an image monitoring device to collect flame images under three combustion conditions, as shown in Figure 4, the image resolution is 384*260*3. FIG. 5 shows the data set structure of this embodiment. The data set contains 1500 images (500 images for each working condition), which are divided into training set and validation set after image preprocessing (image size is compressed to 256*256*3, and the value range is normalized to the range of 0-1). and test set. The training set contains 1080 (360×3) unlabeled images for unsupervised training of the CSAE network; the validation set contains 120 (40×3) labeled images for supervised training of the Soft-max classifier; the test set contains 300 (100×3) label images for CSAE-Soft-max model performance testing.

Step 2) The structure of the convolutional sparse self-encoding network is shown in Figure 3, and the network parameters are summarized in Table 1. The sparse penalty term, mean square error and structural similarity are combined as the loss function of CSAE network, and the training loss under different training times is shown in Figure 6. The results show that when the CSAE network is trained for the 80th time, the loss value has converged, so the number of training times is determined to be 80;

Encoder 1 Encoder 2 Encoder 3 Decoder 1 Decoder 2 Decoder 3 C1(8@3×3+1) C2(4@3×3+1) C3(1@3×3+1) U1(32×32) U2(64×64) U3(128×128) ReLU ReLU ReLU C4(4@3×3+1) C5(4@3×3+1) C6(3@3×3+1) P1(2×2+2) P2(2×2+2) P3(2×2+2) ReLU ReLU Sigmiod

Table 1

Step 3) Utilize the trained CSAE network to extract the deep feature h ₃ of the verification set image, the feature dimension is 64, and combine the corresponding image labels, and perform supervised training on the Soft-max classifier;

Step 4) The trained Soft-max classifier can classify the new image features and realize the combustion condition recognition of a single flame image. Using the test set to test its performance, the recognition accuracy reaches 99.3%.

The experimental results show that the intelligent monitoring model of the convolutional sparse self-encoding network established by the present invention can accurately identify the combustion conditions of a single flame image.

The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. It should be pointed out that for those skilled in the art, some improvements can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the protection scope of the present invention.

Claims (7)

1. a combustion condition monitoring method based on flame image and deep learning, is characterized in that: comprise the following steps: Step 1. Use an image monitoring device to collect images of furnace flames and record the combustion conditions. The collected images are preprocessed and divided into training sets, validation sets and test sets. The images in the training set do not have labels, while the validation set and the images in the test set are labeled; Step 2. Unsupervised training of convolutional sparse auto-encoding with training set images; the combination of sparse penalty term, mean square error and structural similarity as a loss function, expressed as: L=L _SPT +L _MSE +L _SSIM In the formula, L _SPT represents the sparse penalty term, L _MSE represents the mean square error, and L _SSIM represents the structural similarity; Step 3. Use the CSAE network that has completed the training to extract the deep features of the validation set images, and train the Soft-max classifier in combination with the image labels; Step 4. The trained Soft-max classifier can classify the new image features, realize the combustion condition recognition of a single flame image, and use the test set to test its performance. 2. An image monitoring device for the combustion condition monitoring method based on flame image and deep learning according to claim 1, characterized in that: the image monitoring device is composed of a cooling jacket (1), an optical sight glass ( 2) It is composed of an industrial camera (3), the optical sight glass (2) is connected with the industrial camera (3), and the cooling sleeve (1) is sleeved on the outside of the industrial camera (3) and the optical sight glass (2); The top of the cooling jacket (1) is designed with a 45° corner, the cooling jacket (1) is divided into a water interlayer (4) and an air interlayer (5), the water interlayer (4) is a sealed structure, and the gas interlayer (5) is an open structure , and corresponding to the two cooling methods of water cooling and air cooling in turn, in which the air-cooled air interlayer (5) is on the inside, the water-cooled interlayer (4) is on the outside, and a cooling water inlet (6) is provided on the rear end of the cooling jacket (1). ) and a cooling air inlet (7), a cooling air outlet (8) is provided at the top of the cooling jacket (1), wherein the cooling water inlet (6) is communicated with the water interlayer (4), the cooling air inlet (7) and the cooling air The outlet (8) communicates with the gas interlayer (5). The top of the optical sight glass (2) is equipped with a high temperature resistant lens with a viewing angle of 90°; The rear end of the industrial camera (3) is provided with a power interface (9) and a video signal interface (10). 3. a kind of combustion condition monitoring method based on flame image and deep learning according to claim 1, is characterized in that, the feature extraction step of the convolutional sparse self-encoding network in described step 2 is: Step 2.1. Send the training set X to the convolutional coding. First, perform feature extraction by k1 convolutional filters (C1(k1@a1×a1+s1)) with a1×a1 window size and step size s1, and then use The ReLU function (y(x)=max(0,x)) performs neuron activation, and finally the feature reduction is performed by the maximum pooling layer (P1(b1×b1+f1)) with a window size of b1×b1 and a step size of f1 dimension, get the feature vector h ₁ ; Step 2.2, the feature vector h ₁ is sent to the encoder No. 2 and the encoder No. 3 for further processing, the processing process is similar to that of the encoder No. 1, and finally the feature vector h ₃ is obtained; Step _2.3 . The feature vector h3 is sent to the decoder No. 1. First, the feature dimension is improved by the upsampling layer (U1(g1×g1)) of the g1×g1 window size, and then the convolution filter (C4(k4@ a4×a4+s4)), ReLU activation function processing, obtain feature vector h ₄ ; Step 2.4, the feature vector h ₄ is sent to the decoder No. 2 and the decoder No. 3 for further processing. The processing process is similar to that of the decoder No. 1, and finally the reconstructed image z is obtained. 4. a kind of combustion condition monitoring method based on flame image and deep learning according to claim 1, is characterized in that, the sparse penalty term L _SPT in described step 2 is expressed as:

In the formula, β represents the sparsity rate; F represents the number of neurons in the hidden layer; p represents the sparse target constant; the average activation of neurons based on the i-axis p _j is expressed as:

In the formula, E represents the number of images in the training set; s _ij (i∈(1, E), j∈(1, F)) represents the activation of the neuron at the jth position; the KL divergence is expressed as:

5. a kind of combustion condition monitoring method based on flame image and deep learning according to claim 1, is characterized in that, the loss function L _MSE in described step 2 is expressed as:

In the formula, X _ij and z _ij represent the grayscale of the (i, j)th position in the input image and the reconstructed image with size A×T, respectively. 6. a kind of combustion condition monitoring method based on flame image and deep learning according to claim 1, is characterized in that, the loss function L _SSIM in described step 2 is expressed as:

In the formula, c ₁ =(k ₁ r) ² and c ₂ =(k ₂ r) ² , k ₁ and k ₂ are constants less than 1, r represents the dynamic range of image grayscale; μ _x and μ _z represent the input The average value of the image and the reconstructed image; σ _x and σ _z represent the variance of the input image and the reconstructed image; σ _xz represent the covariance of the input image and the reconstructed image. 7. a kind of combustion condition monitoring method based on flame image and deep learning according to claim 1, is characterized in that, the Soft-max classifier in described step 3 is defined as:

In the formula, _yi represents the output of the _ith sample xi; k represents the number of categories of training samples.

Images