The application of machine learning techniques to detect combustion modes in a pulverised coal boiler

. The development of machine learning algorithms based on semi-industrial thermal benches will approach the development of an automated system capable of detecting and tweaking energy-efficient and environmentally friendly combustion modes in large power plants and increasing their efficiency without significant changes in the design of boiler equipment. Determination of combustion modes and optimisation of the combustion process based on neural network analysis of visualisation patterns of the coal flame in the boiler. Determining the combustion mode in the furnace space and superimposing (automatically adjusting) the parameters based on sensor readings to bring it to the optimum mode and maintain stable combustion is a complex task. Currently, the selection of necessary parameters is done by operator-assisted automatic process control systems, but this process is based on known design parameters and is not always efficient or environmentally friendly in practice. This problem can be solved by determining the combustion mode in the furnace space using modern machine learning methods and automatic parameter optimisation in a continuous mode article.


Introduction
Development of machine learning algorithms based on semi-industrial thermal benches will make it possible to approach the creation of an automated system capable of detecting and adjusting energy efficient and environmentally friendly combustion modes at large power plants and increase their efficiency without significant changes in the design of boiler equipment [1][2][3][4].For example, in [5] an experiment on adjustment of coal combustion in a 660 MW boiler was conducted.The authors showed that the developed model of classification without a teacher can successfully classify the combustion modes.The authors applied a model based on convolution autocoder, principal component method and Markov model to control the combustion mode based on flame images that are collected from the combustion control system of the furnace.In [6] the results of investigations of geometrical and light characteristics of pulverized coal flame using the developed flame control system are presented.The results obtained under different combustion conditions show that the developed system is able to characterize the flame both qualitatively and quantitatively.Correlations between measured flame parameters and corresponding furnace operation parameters, including furnace load, primary air mass flow rate, and particle size are identified and analyzed.In [7], a neural network-based prediction of the performance (capacity) of a 600 MW coal-fired power plant is discussed.The development of such models is particularly important for improving the efficiency of power plants.In the literature, increased attention has been paid to the modelling and reduction of NOX, SOX emissions for coal-fired boilers using advanced machine learning techniques [8][9].The authors carried out a comparative analysis of different machine learning models: linear regression, support vector method, full-link and recurrent neural networks.Based on the simulation results, it was shown that it was possible to reduce NOX emissions by up to 15-20%, which was further verified by the authors experimentally.Timely and accurate prediction of basic quality characteristics of stripped coal plays an important role in condition monitoring and production management, but coal quality characteristics are usually difficult to measure directly online in industrial practice.In [10], the authors proposed to solve the problem using virtual sensors based on deep learning.The proposed model was a multilevel autoencoder and a bi-directional recurrent neural network (Bidirectional Long Term Memory network, Bi-LSTM).An important feature was the joint use of marked and unmarked data, which was used to improve the quality of the model with the teacher.
A promising method for regulating and detecting combustion processes in thermal power plants is the use of machine learning techniques.Monitoring tools play an important role in maintaining or optimising combustion processes.Flame visualisation can be used as an additional source of data since the observed characteristics of a flame, such as its size, shape, front position, brightness and glow spectrum, as well as the way these parameters change over time, provide a large amount of information about the combustion process.A number of works in this direction are devoted to attempting to formulate a method or algorithm for E3S Web of Conferences 459, 07012 (2023) https://doi.org/10.1051/e3sconf/202345907012XXXIX Siberian Thermophysical Seminar estimating combustion process parameters based on analysis of individual characteristics of flame images.An example of the use of information on the spectral composition of radiation (flame colour) directly related to the chemical reactions taking place is the method of planar pyrometry and many others [12][13].The main elements of the furnace are: a fuel supply system; a straight-flow burner located horizontally on the furnace axis; a combustion chamber with gauges and sight-glasses; a combustion chamber with tangential secondary air supply; and a smoke exhauster.The combustion chamber works as follows.Coal is poured into the coal bunker, after which it is fed into the air duct by a screw feeder, where it is mixed with the primary air.The supply of primary air into the blender is carried out by blower MSH BL-520-670.The pulverised coal mixture enters the electric arc block, where coal ignition is initiated by passing through plasma arcs.The combustion mixture is fed straight into the combustion chamber where it is mixed with secondary air fed through a tangential swirler.In the combustion chamber coal dust combustion takes place.After passing through the combustion chamber, the unburned fuel enters the afterburning chamber where it mixes with the tertiary air required for complete combustion.The tertiary air is supplied to the afterburning chamber via a tangential swirler.The tertiary air supply is located in the lower part of the combustion chamber.The supply air temperature in all experiments was 22 °C.
Before carrying out a series of experiments, the coal screw feeder and air flow sensors were calibrated.The minimum air flow rate was determined from the conditions of coal fuel transportation.Measurement of air flow rates was carried out during the whole experiment.
The composition of the gas mixture was measured by a TEST 203 gas analyser along the axis of the unit.Principle of operation: measuring CH4, CO, CO2 -optoabsorption, O2 -electrochemical, H2 -polarographic.Temperature measurement is carried out with platinumrhodium thermocouples placed along the axis of the unit.

Combustion of composite fuel
Deep learning methods were used to monitor coal flame combustion in the furnace and to determine combustion anomalies from flame images.For three different alpha and three time points.The images with lower alpha clearly show more soot (darkened area).
The problem was solved using two approaches.In the first one, teacherless learning was used: a neural network autoencoder was developed, which is a combination of convolutional layers, full-coupled layers and upsampling layers.The neural network autoencoder was trained to reconstruct combustion modes from flame images corresponding to high values of the excess air ratio.Then, using the trained autoencoder, abnormal combustion modes with a lower excess air ratio were identified, at which an increase in unburned coal dust was observed.For this purpose, the reconstruction error of the original image was used.Since the autoencoder is trained only on normal images, the reconstruction error of abnormal images will be much higher, which allows the automatic detection of abnormal observationsTo simplify the implementation of the neural network, a Python development was used using the Keras framework and the Tensorflow backend.A simple fully coupled autoencoder had the following layer architecture: Input (256x256) -Flatten (65536) -Dense (1024) -Dense (1024) -Dense (512) -Dense (1024) -Dense (65536) -Reshape (256x256), where Input is input layer, Flatten is rectification operator, Dense is fully coupled layer, Reshape is shape change operator.
The second model is a convolutional autoencoder, which consists only of convolutional and transpose convolutional layers.In the encoder, the input data passes through 12 convolution layers (Conv2D) with kernels 3x3 and filter sizes starting with 8 and increasing to 32.Decoder reconstructs image using transposed convolution layers (Conv2DTranspose).Using only convolutional layers may seem unreasonable, but in this case the aim of the study is to compare different architectures on model performance.The last architecture analysed is a combined autoencoder with convolution layers, merge layers and upsampling layers (see Figure 4.).The encoder is the following sequence of layers: Conv2D -Con2D -MaxPooling2D -Conv2D -Conv2D -MaxPooling2D -Flatten -Dense.In this way, the convolution (Conv2D) and flattening (MaxPolling2D) layers alternate, ending with a fullcoherent layer (Dense).The output from the last convolutional layer of the decoder is a reconstructed image with an initial resolution of 256x256.
Most of the original image, containing the shaded area, was cropped.The resolution of the original data was reduced to 256x256 pixels.The aim of the experiments was to train a convolutional neural network with robustness to image distortion and data noise.An effective solution, augmentation, was applied where the training data was artificially extended with distorted versions of the images during training.To implement this procedure, we used the Keras framework, which provides an interface to extend the training set -the ImageDataGenerator class.We initialize the data generator and choose what types of transformations we want to apply to the images, then we run the training data through the generator by calling the fit method and then the flow method, obtaining a constantly expanding iterator for the image sets we replenish.There is even a special method model.fit_generatorthat will train our model using this iterator, which greatly simplifies the code.ImageDataGenerator also gives us the ability to do horizontal vertical shifts, random rotations, scaling, warping and mirroring, adding Gaussian noise, etc.

Results
The developed deep neural network autoencoder is a combination of convolution layers, full-coupled layers and upsampling (upsampling) layers.An autoencoder trained to reconstruct combustion modes with high excess air ratio values has been used to identify abnormal combustion modes with lower excess air ratio by the reconstruction error value (see Fig. 7).
It was shown that the developed convolution autoencoder model (teacherless model), captures many abnormal situations of coal dust combustion.However, the number of common situations classified as anomalies (false positives) is indeed large, but the sensitivity of the detector can be adjusted by changing the probability threshold.The average classification accuracy was 77% and the average completeness was 66%.The area under the ROC curve was AUC ROC = 0.62, under the accuracy and completeness curve AUC PRC = 0.54.
Compared to the autoencoder (the model without a teacher), the teaching model with a teacher performs better (see Figure 8).From the error matrix it can be seen that the worst accuracy for class α=0.21 is 80% and the best for class α = 0.35 is 92%, the average accuracy for all classes is 87%.

Conclusion
Thus, the possibility of using computer vision approaches, in particular deep neural networks for the detection of abnormal coal dust combustion modes in the combustion chamber corresponding to low values of the excess air ratio is shown.Despite the high number of false positives for the convolution autoencoder, the model can be useful for anomaly detection.The accuracy and completeness of the model can be greatly improved by collecting more detailed data with combustion parameters monitored, making image labelling more accurate.Further work will be carried out in this direction.
In contrast to the convolution autoencoder, the classifier allows combustion mode and air excess ratio to be determined from flame images with an average accuracy of 87%, which is satisfactory for industrial use.It is worth noting that regularisation, initialisation of weights, selection of batch size and cyclic learning rate had the biggest impact on model quality.

Fig. 2 .
Fig. 2. Examples of flame combustion patterns with different values of excess air ratio α.
The data was divided into a training sample, a test sample and a validation sample.The training sample included 10000 images, the test sample included 100 images and the validation sample included 500 images.Only combustion modes with high fuel surplus factor, which are characterised by high combustion intensity of the coal flare, have been included in the training sample.The architecture of the combined neural network model (similar to AlexNet) for the task of classification of combustion modes by excess air ratio is presented in fig. 3.

Fig. 3 .
Fig. 3. Combined model architecture for classification of combustion modes according to excess air ratio α.