RU2713850C1 - Fuel combustion modes monitoring system by means of torch images analysis using classifier based on convolutional neural network - Google Patents

Abstract

FIELD: heat exchange.

SUBSTANCE: invention relates to heat engineering and can be used for prediction and control of fuel flaring, in particular, in furnace devices in coal and gas boilers. System includes a device for recording images of flame chemiluminescence formed during combustion of radicals and images of stimulated luminescence of radicals, a computing device which implements a classifier based on a trained convolution neural network, and a device for displaying results.

EFFECT: technical result is higher efficiency and safety of fuel combustion due to analysis of images of a flame, which enables more accurate, reliable and practically non-inertial diagnostics of fuel combustion mode change.

4 cl, 3 dwg, 1 tbl

Description

The invention relates to the field of power engineering. The invention relates to methods for predicting and controlling fuel flaring, in particular in combustion devices in coal and gas boilers. The invention can find application in any industry related to flare.

To date, in industrial applications, to monitor the combustion process and detect the transition of the process to an unfavorable mode, the flare combustion diagnostics is carried out using cameras, gas analyzers, thermocouples and a number of other sensors.

One approach to monitoring the combustion process is to compare the image of the flame and the combustion mode, i.e. a characteristic range of parameters in which this mode is implemented, for the implementation of which it is necessary to solve the problem of flame image classification based on a multidimensional feature vector.

There are a wide variety of methods for classifying using a multidimensional feature vector as input [Witten I., Frank E., Hall M., Pal C. Data Mining: Practical Machine Learning Tools and Techniques. San Francisco: Morgan Kaufmann, 2017. 654 p.]: Support vector method, neural network (for example, a fully connected network such as a multilayer perceptron), gradient boosting over linear models or decision trees, and others. The effectiveness of such methods for solving computer vision problems largely depends on the expert knowledge of the researcher in creating an indicative description of images.

A known system for monitoring and control of the parameters of the combustion process in the furnace of the boiler [US 6468069, 10/22/2002, F23N 1/02; F23N 5/08; F23N 001/00; F23N 005/08], including a scanning device (a video camera, a CCD camera, a laser beam scanning device and a corresponding photodetector) generating electrical signals that vary depending on the combustion process, a scanning device control system, sensors for monitoring various parameters of the combustion process, a processor, connected to the output of the scanning device that implements a trained recurrent neural network that processes scanning signals and generates combustion classification signals, the controller logic, which receives combustion classification signals, analyzes them, determines and generates combustion control signals. The system uses a neural network classifier that classifies signals from a scanning device. In the description of the system, a classifier based on a fully connected neural network is indicated.

The applied neural network image processing algorithm has low efficiency compared to a convolutional neural network with custom spatial filters in convolutional layers.

Convolutional Neural Networks (CNN - Convolutional Neural Network) [Hagan M.T., Demuth N.V., Beale M.N., de Jesus O. Neural Network Design. USA: Martin Hagan, 2014. 800 p.] Have been successfully applied to a wide class of image classification problems, including handwriting recognition and faces in images.

Basic for the convolutional neural network are two operations performed in separate layers: convolution and subsampling. The convolutional layer significantly reduces the number of parameters compared to a fully connected layer, which can significantly increase the time and quality of image processing,

however, it uses more hyperparameters that must be defined before the training procedure.

A classifier with a convolutional neural network is known from the technical field [US 20150036920, 2015-02-05, G06K 9/6]. The known solution is a general-purpose classifier and does not take into account the features of the training of the neural network in relation to the case of monitoring combustion modes. The known solution does not have sufficient stability of the trained classifier.

The closest in combination of features and the obtained technical result is a device (image sensor / neural network controller) for combustion control systems [Patent US 5249954, 1993-10-05, F23N 5/08], based on the registration of chemiluminescence images generated during the combustion of radicals and processing such images using a neural network. A well-known technical solution for image processing uses a neural network, including an input layer (inputlayer), an output layer (outputlayer) and a hidden layer with customizable link weights (hiddenlayer).

Such a neural network image processing algorithm has low efficiency compared to a convolutional neural network with custom spatial filters in convolutional layers. This makes the device unsuitable for real-time monitoring. Also, the solution described in the patent does not specify the method on the basis of which training and further image analysis are performed.

The objective of the present invention is to develop a system for monitoring fuel combustion modes by analyzing torch images, which allows more accurately, reliably and practically inertialess (at the time of transition) to diagnose a change in fuel combustion mode, that is, a change in combustion parameters, which will improve the efficiency and safety of fuel combustion.

This problem is solved by creating a system for monitoring fuel combustion modes by analyzing torch images using a classifier based on a trained convolutional neural network, including a device for registering chemiluminescence images of flame generated during the combustion of radicals and images of stimulated luminescence of radicals, a computing device that implements a classifier based on trained convolutional neural network, a device for displaying results.

According to the invention, the image recording device is installed so that a portion of the base of the flame is included in the observation area, and the registration of flame images with a given combination of parameters such as the range of the degree of swirling flow, the range of the fuel / air ratio, the range of emission of volatile components, the temperature range of any of the flame sections, or other parameters characterizing the process of burning fuel.

According to the invention, a convolutional neural network consists of functional layers of the following types: convolutional, subsampling, excluding, aligning, fully connected, as well as functional layers of other types.

According to the invention, a convolutional neural network consists of an arbitrary required number of functional layers of various types, combined in a certain sequence so that the subsampling and exclusion layers are located after each convolutional layer, or after a pair of convolutional layers in the following order: convolutional layer (or two convolutional layers in a row ), a subsampling layer excluding the layer.

Each convolutional layer can contain an arbitrary required number and size of spatial filters and their corresponding feature maps, pre-selected by randomly sorting over the parameter grid, while maximizing the classification accuracy for cross-validation. The values of the classifier’s hyperparameters are also selected by random enumeration on the parameter grid.

According to the invention, a convolutional neural network is trained using statistical initialization of weights in the layers of the neural network and augmentation of training data (images) on flame images pre-recorded on the same burner full-time device, which is subsequently used in a system labeled in accordance with the burning mode .

According to the invention, the size of the convolutional core, the number of nuclei in each convolutional layer, as well as the number of neurons in fully connected layers are used as a combination of customizable classifier hyperparameters in a convolutional neural network.

A system for monitoring fuel combustion modes by analyzing a torch image using a classifier based on a trained convolutional neural network is shown schematically in FIG. 1, where: 1 - burner device; 2 - monitoring system; 3 - image registration device; 4 - a computing device that implements a classifier based on a convolutional neural network; 5, 6, 7, 8 9 - layers of the neural network; 5 - input layer of the neural network; 6 - convolutional layer of the neural network; 7 - sub-sampling layer of the neural network; 8 - an exclusive layer of a neural network; 9 - fully connected output layer of the neural network; 10 is a result display device.

The monitoring system of the combustion regimes includes: a device for recording images of 3 chemiluminescence of flame generated during the combustion of radicals and images of stimulated luminescence of radicals, a computing device 4 that implements (software or hardware) a classifier based on a convolutional neural network, a device for displaying the results 10.

The image recording device 3 is a digital camera equipped with an optical filter or a set of optical filters, designed to highlight the range of radiation wavelengths corresponding to the spectral range of flame luminescence.

Computing device 4 — a computer or controller, implements (software or hardware) a classifier based on a convolutional neural network that analyzes flame images received from the device for recording images. The task of the convolutional neural network is to classify the images of swirling flame arriving at its input and assign them to one of the known classes.

The device for displaying results 10 is a monitor, a separate indicator, or a set of indicators that displays the result of the classifier - information about which of the predefined classes the current state of the flame (burning mode) and / or a predetermined range of parameters in which such mode.

The convolutional neural network is designed to solve the problem of classifying flame images obtained from the video stream of a digital camera. The convolutional neural network consists of an arbitrary required number of convolutional functional layers, including:

- convolutional,

- sub-sampling (pooling),

- exclusive (dropout),

- leveling (flattening),

- fully connected (fully-connected),

- functional layers of other types.

The convolutional nero network is arranged in such a way that data is transmitted sequentially through the layers in the direction from the input layer 5 to the output layer 9. The input data for each layer is the output data of the previous layer.

An example configuration of a convolutional neural network is shown in FIG. 2, where (0) is the input layer, the input of which the image is supplied (color or black and white), the output is the same image. In the example, a color three-channel image with a size of 64 × 64 pice.

(1, 2) is a convolutional layer, the input of which is the original image transmitted from the input layer, the output is a set of feature maps (matrices) highlighted by the convolutional layer. In the example, there are 16 feature cards with a size of 64 × 64 pique.

(3) - a sub-sampling layer, at the input of which a set of feature cards transmitted from the convolutional layer is fed, the output is a set of feature cards of a smaller size (matrix with a reduced number of rows and columns). In the example, there are 16 feature cards 32 × 32 pique in size.

(4) - an exclusion layer, at the input of which a set of feature maps is passed, transferred from the exclusion layer, output is a set of feature maps of the same size as at the input, a part of the weights specified by the layer parameters (values in the feature map represented by the matrix) are zero. In the example, there are 16 feature cards 32 × 32 pique in size.

(5, 6) - a convolutional layer, at the input of which a set of feature cards from the exclusion layer is fed, the output is a new set of feature cards (matrices) highlighted with a convolutional layer. In the example, there are 8 feature cards 32 × 32 in size.

(7) - a sub-sampling layer, at the input of which a set of feature cards transferred from the previous convolutional layer is fed, the output is a set of feature cards of a smaller size (matrix with a reduced number of rows and columns). In the example, there are 8 feature cards with a size of 16 × 16 pique.

(8) - the exclusion layer, at the input of which a set of feature cards is passed, transferred from the exclusion layer, the output is a set of feature cards of the same size as at the input, the part of the weights specified by the layer parameters (values in the feature map represented by the matrix) is zero. In the example, there are 8 feature cards with a size of 16 × 16 pique.

(9) - a leveling layer, at the input of which a set of feature cards is supplied from the exclusive layer, the output is a vector constructed based on a set of feature cards. In the example, a vector containing 2048 elements.

(10) - a fully connected layer, the input of which is a vector transferred from the alignment layer, the output is a vector with a smaller number of elements. In the example, a vector containing 512 elements.

(11) - the exclusion layer, the input of which is a vector transmitted from a fully connected layer, the output is a vector of the same dimension as the input. Part of the vector elements is zero. In the example, a vector containing 512 elements.

(12) is a fully-connected output layer, the input of which is a vector transmitted from the exclusive layer, the output is a vector with the number of elements corresponding to the number of defined classes. In the example, there are 13 elements (defined classes).

Convolutional functional layers are combined in a certain sequence so that the subsampling and excluding layers are located after each convolutional layer, or after a pair of convolutional layers in the following order: a convolutional layer (or two convolutional layers in a row), a subsampling layer, a convolutional layer (or two convolutional layers in a row ), a subsampling layer excluding the layer,

Each convolutional layer can contain an arbitrary required number and size of spatial filters, and the corresponding feature map filters, moreover, the selection of the required number of convolutional layers, the number and size of feature cards is carried out by random sorting over the parameter grid, while maximizing the classification accuracy on the part of the set of training images.

As a combination of configurable hyperparameters in a convolutional neural network, the following are used: the size of the convolutional core, the number of nuclei in each convolutional layer, and the number of neurons in fully connected layers. The values of the hyperparameters are selected by random enumeration on the parameter grid,

The convolutional neural network is pre-trained by supplying to its input sets of flame images received from the registration device on the same burner device on which the system will be used for monitoring later on and labeled in accordance with the operating parameters of the combustion process. The convolutional neural network is trained on labeled data using statistical initialization of the initial weights with a given dispersion of values and the use of training data augmentation.

Before starting to use the monitoring system of fuel combustion modes, a neural network architecture is selected, hyper parameters are configured and the neural network is trained in various combustion modes. For this, a set of flame images is previously recorded using the registration device on the same burner device on which the system is subsequently used for monitoring, and marked in accordance with the operating parameters of the combustion process.

After that, on the resulting data set, the architecture is selected, hyper parameters are configured, and the neural network is trained in various combustion modes. In the learning process, the convolutional neural network, which is part of the system, learns to compare the images supplied to the input with the indicated burning mode. After training the neural network, the recording device in streaming mode registers images of the torch and transmits them to the input of the convolutional neural network. The convolutional neural network determines the probability that the received image belongs to one of the “known” combustion modes and, as a result of the operation, provides the name of the combustion mode to which the received image and / or set of parameters that determine the burning mode most likely corresponds.

The formation of a set of source data is as follows. The image registration device (video camera) is placed so that it observes the flame of the burner device in all determined combustion modes and the flame images are recorded from one angle. The area of the base of the flame should be included in the observation area. Sets of flame images are recorded for a given combination of parameters (or a range of parameters) that determine the “mode” of combustion. As such parameters, the following can be used: the range of the degree of swirling the flow, the range of the ratio of the fuel / air, the range of the emission of volatile components, the temperature range at any part of the flame, or other parameters characterizing the process of burning fuel. At least two thousand images are recorded for each mode. To process the obtained images using multiprocessor computing on a graphics processor. For processing, a computer is used on the basis of productive CPUs compatible with the architecture of × 86, with a RAM capacity of at least 4 GB, equipped with a discrete graphics processor that implements the CUDA hardware-software architecture of calculations, with a RAM capacity of at least 6 GB.

Data preprocessing is to reduce the resolution of images to the desired value. The image is reduced to a size that allows processing at the required speed, but not less than to a size of 64 × 64 pique.

The selection of architecture and optimization of hyperparameters of the convolutional neural network is as follows. A convolutional neural network consists of an arbitrary required number of functional layers of various types, including an input and output layer, convolutional layers, subsampling layers, excluding layers, a leveling layer, fully connected layers combined in a certain sequence so that subsampling and excluding layers are located after each convolutional layer , or after a pair of convolutional layers in the following order: a convolutional layer (or two convolutional layers in a row), a subsampling layer excluding the layer. The input is a color RGB image or an image in grayscale. First, the image is convolutional, or sequentially two convolutional operations. After that, the data is transferred to the subsampling layer. The subsampling layers implement the compaction of the feature map by selecting the maximum value with the given values of the transformation core and step (for example, a 3 × 3 transformation core and step 2, which will give an output feature map of size 0.5 n × 0.5 m, where n and m are the size of the input map signs). Exclusion layers nullify randomly selected values in the feature map, excluding them from further processing during the iteration of training, which allows us to regularize the trained model and prevent retraining of the neural network. After the sequence of blocks [convolutional layer - subsampling layer - exclusion layer] in the neural network, there is an alignment layer that converts a set of feature maps into a feature space vector, a fully connected layer, or a sequence of fully connected layers (usually no more than three) highlighting features of a higher level. Between fully connected layers there are exclusion layers, occupying randomly selected values in the feature vector, excluding them from further processing during the iteration of training. The output of the last fully connected layer is fed to the final Softmax layer, which calculates the probability that the image belongs to the class. At the initial stage, the initial network architecture is established, that is, the number of layers of each type and their parameters.

The combination of model hyperparameters is selected by random enumeration over the grid, while the classification accuracy for cross-validation is maximized using the generated data set. It is mandatory to select a combination of hyperparameters that have the greatest influence on the quality of classification: the size of the convolutional nucleus, the number of nuclei in each convolutional layer, the number of neurons in fully connected layers. Other hyperparameters can also be selected for finer tuning: the number of training images processed per iteration of the gradient descent algorithm, the size of the subsample in the subsampling layers, and the probability of eliminating a neuron in the exclusion layer. If the necessary classification accuracy is not achieved, then the number of blocks is changed [convolutional layer - subsampling layer - excluding layer], the optimization procedure is increased and repeated.

Before starting the application, the network is trained on various combustion modes. During the training, the cross-entropy function is used as the loss function, and the proportion of correct answers as the quality metric. To optimize the vector of neural network weights, the adaptive inertia method is used. For training, a complete sample of pre-registered images is divided into three disjoint subsets: training, validation, test selection. When splitting the sample, mixing is used while maintaining the balance of the classes (the relative share of each class) in each subsample. In the learning process, images from the training subsample and a class marker (marker) indicating the burning mode at which the image was recorded are fed to the input of the convolutional neural network. In this case, the neural network learns to compare the flame images supplied to the input with a class mark. Test subsampling is used to check the quality of training (classification accuracy) at each iteration (era). A validation subsample is used to further verify the quality of training after it is completed.

Monitoring of combustion modes is as follows. To monitor the combustion regimes, the image recording device (video camera) is placed in the same position at which the images for the training sample were recorded. A registration device registers torch images in streaming mode and transfers them to a computing device that implements a classifier based on a trained neural network. Images are transmitted to the input of the convolutional neural network. The convolutional neural network determines the probability that the received image belongs to one of the “known” modes (performs classification) and, as a result of the operation, provides the name of the combustion mode to which the received image and / or operating parameters of the combustion process corresponding to this mode most likely correspond. The classification result is displayed on the display device, which is a monitor, a separate indicator, or a set of indicators.

A distinctive feature of the convolutional neural network used in the system is its purpose - to solve the classification problem, and the presence of convolutional functional layers in its composition. A distinctive feature of the system as a whole is monitoring by solving the classification problem and using a classifier based on a convolutional neural network as part of the system.

To determine the combustion regimes, a convolutional neural network trained on marked data was used, and the result of its work was compared with the classification results using several linear models.

Flame images were collected on a laboratory gas burner stand, which is an open circuit connected to cylinders with combustible gas and a line with compressed air. Sets of flame images were recorded for a given combination of two parameters that determined the “mode” of combustion: the degree of swirling flow S and the ratio of fuel / air in the mixture (fuel excess coefficient) F. In total, 13 different combustion modes of the propane-air mixture were recorded. The parameters of the modes are shown in table. 1. The shooting was carried out with a color camera with an image resolution of 2 megapixels. For each mode, about 10,000 images were recorded on a dark background from one angle.

The data preprocessing, classification, and CNN model building algorithms were implemented in the Python programming language using the NumPy, SciPy, Pandas, Scikit-Learn, OpenCV, Theano, Tensorflow, and Keras libraries. For data processing, multiprocessor calculations were used on one computer (as part of a computing cluster based on two IntelXeon E5 v2, 8 x DDR3 16GB processors).

A complete sample of the used flame images consisted of 39,000 marked RGB-images in the size of 1920 × 1080 pixels - 3000 images for each of 13 modes. The sample was divided into three disjoint subsets: training 56% of images, validation 14% and test sample 30%. When splitting the sample, as for linear methods, mixing was used with maintaining the balance of classes in each subsample. Image resolution was reduced to 64 × 64 RGB pixels to reduce network training time. This size of input images was chosen empirically, based on a compromise between classification accuracy and training time and model prediction. The proposed classification method was based on the classical convolutional neural network combining convolutional and subsampling layers.

The combination of model hyperparameters was selected by random enumeration over the grid, while the classification accuracy on cross-validation was maximized, and the final model was tested on a delayed sample. Among the selected hyperparameters were: training images processed in one iteration of the gradient descent algorithm, iterations of the training algorithm throughout the training set, kernel size and number of kernels in convolutional layers, size of the subsample in sub-sampling layers, the probability of eliminating a neuron in the dropout procedure and the number of neurons in a fully connected layer. As a result of the selection of hyperparameters, the classification accuracy was increased by 3.5%.

The convergence of network learning with selected hyperparameters was achieved in 25 epochs. The training time for the final model was 24.2 minutes, the average classification time for one image was 5.1 ms. The procedure for selecting hyperparameters took 40.8 hours.

Classification accuracy averaged across all regimes on a delayed sample was 97.9%.

In fig. The classification error matrix for 13 classes is shown.

The table provides a metric for the quality of image classification using a convolutional neural network for individual modes. The presented error matrix demonstrates the occurrence of an increased error in the classification of modes 0, 7 and modes 3, 4, 8, which were also classified with an error when using linear methods. Nevertheless, even for these modes, the quality of classification using CNN significantly exceeds the average quality of classification using linear methods, and for modes with high variability (1, 11, 12), the proposed model demonstrates confident classification with an accuracy of 99%.

Claims (8)

1. A system for monitoring fuel combustion modes by analyzing torch images using a classifier based on a trained convolutional neural network, including a device for registering chemiluminescence images of flame generated during the combustion of radicals and images of stimulated luminescence of radicals, a computing device that implements a classifier based on a trained convolutional neural network , a device for displaying results, characterized in that the image registration device is installed in such a way that the base of the flame is included in the observation area, and the flame images were recorded from a single angle for a given combination of parameters defining the “mode” of combustion, a convolutional neural network consists of an arbitrary required number of functional layers of various types, combined in a certain sequence so that subsampling and excluding layers are located after each convolutional layer, or after a pair of exact layers in the following order: convolutional layer (or two convolutional layers in a row), subsampling layer, excluding layer, each convolutional layer can contain an arbitrary required number and size of spatial filters and corresponding feature maps pre-selected by randomly sorting over the parameter grid, while maximizing the accuracy of classification on cross-validation and checking the final model on a delayed sample, the values of the classifier’s hyperparameters are selected by random enumeration on the parameter grid, the convolutional neural network is trained on images of the flame previously recorded on the same burner device, which later uses the marked data system in accordance with the burning mode, the convolutional neural network is trained using statistical initialization of weights in the layers of the neural network and augmentation of training data (images ) 2. The monitoring system of fuel combustion regimes according to claim 1, characterized in that the convolutional neural network consists of the functional layers of the following types: convolutional, sub-sampling, excluding, equalizing, fully connected. 3. The system of monitoring fuel combustion regimes according to claim 1, characterized in that the size of the convolutional core, the number of nuclei in each convolutional layer, and the number of neurons in fully connected layers are used as a combination of configurable classifier hyperparameters in the convolutional neural network. 4. The monitoring system of fuel combustion modes according to claim 1, characterized in that as a combination of parameters that determine the "mode" of combustion, use: the range of the degree of swirling flow, the range of the ratio of fuel / air, the range of emission of volatile components, the temperature range at either from flames.

Images