A three-dimensional visualization diagnosis system for furnace combustion state by fusing flame image features and oxygen data

Abstract

The application discloses a kind of fusion flame image features and oxygen quantity data's furnace combustion state three-dimensional visualization diagnosis system, it is related to furnace combustion monitoring and intelligent diagnosis technical field, including feature module, obtain the flame video stream inside furnace, time series registration is carried out to continuous video frame, extract regional luminance variation feature, obtain flame luminance feature matrix, fusion module, obtain the real-time oxygen proportioning data of each area inside furnace and the area coordinate of each area in preset furnace three-dimensional coordinate system, according to real-time oxygen proportioning data and flame luminance feature matrix, obtain regional combustion feature vector;The fusion flame image features and oxygen quantity data's furnace combustion state three-dimensional visualization diagnosis system, it is favorable to improve the stability of furnace combustion process, adjusting pertinence and operating safety.

Description

Technical Field

[0001] This invention relates to the field of furnace combustion monitoring and intelligent diagnosis technology, specifically to a three-dimensional visualization diagnostic system for furnace combustion status that integrates flame image features and oxygen data. Background Technology

[0002] Furnace combustion is widely used in industries such as power, metallurgy, and chemicals, and is a crucial process for converting the chemical energy of fuel into thermal energy. The stability of the combustion state inside the furnace directly affects equipment operating safety, energy utilization efficiency, and pollutant emission levels. As industrial production increasingly demands higher load response speeds, combustion efficiency, and operational reliability, operators need to promptly monitor the combustion state in different areas of the furnace to make appropriate adjustments to fuel supply and air delivery.

[0003] Current furnace combustion monitoring methods typically rely on external parameters such as furnace outlet flue gas temperature, steam pressure, and flue gas oxygen content to determine the combustion state inside the furnace. These parameters are mostly results of the furnace combustion and heat transfer processes, and there is a certain transmission time between changes in the internal combustion state and corresponding fluctuations in these external parameters, making it difficult to reflect dynamic changes in the combustion state in a timely manner. Some furnaces are equipped with thermocouples, infrared temperature measuring devices, or flame image acquisition devices to obtain internal temperature or flame image information. However, thermocouples and infrared temperature measuring devices can usually only obtain measurement data from limited locations, making it difficult to reflect the overall temperature distribution at different heights, cross-sections, and regions within the furnace. While flame images can reflect flame brightness, shape, and fluctuations, continuous images are easily affected by flame swaying, smoke obstruction, and airflow disturbances. Image features at different times may shift positionally, making it difficult to maintain a stable correspondence between flame changes and the actual area of ​​the furnace. Furthermore, the relationship between flame brightness and combustion temperature is not a fixed one-to-one correspondence. When the oxygen content in a localized area of ​​the furnace is insufficient, incomplete combustion of fuel may occur. The flame may still have a certain brightness, but the effective heat release will decrease, and unburned components may accumulate. When the oxygen content in a localized area is too high, excess cold air will absorb and carry away heat, causing a deviation between the flame brightness and the actual temperature. Therefore, judging the combustion state in the furnace solely based on flame images or oxygen content data is easily affected by changes in fuel concentration, air supply conditions, and the degree of combustion, making it difficult to accurately distinguish between insufficient oxygen, excess air, and normal combustion.

[0004] Furthermore, fuel concentration, oxygen levels, and airflow direction typically vary across different areas within the furnace. However, existing monitoring systems often rely on independent flame image data, oxygen data, and regional location information, making it difficult to establish stable temporal and spatial correlations. Consequently, monitoring results usually only reflect the overall furnace condition or a few localized conditions, making it difficult to accurately estimate the actual temperature of each area, let alone generate a three-dimensional temperature distribution that characterizes temperature changes within the furnace. Existing furnace monitoring interfaces typically display combustion status using parameter lists, trend curves, or two-dimensional images, making it difficult to visually display the spatial location of abnormal areas within the furnace and the distribution range of abnormal conditions. When localized overheating, uneven burning, wall-attached combustion, or incomplete combustion occur within the furnace, operators struggle to promptly pinpoint the specific location of the risk area and to comprehensively assess the correlation between flame condition, oxygen level, and temperature deviations. Regarding combustion regulation, existing systems typically adjust fuel and air volume uniformly based on the overall furnace temperature, average oxygen level, or outlet parameters, failing to consider the actual combustion needs of different areas. Unified regulation may cause the temperature in high-temperature areas to continue to rise, while the temperature in low-temperature or oxygen-deficient areas is not effectively improved, thereby increasing the risk of local coking, damage to heated surfaces, accumulation of unburned components, and unstable combustion. Summary of the Invention

[0005] The purpose of this invention is to provide a three-dimensional visualization diagnostic system for furnace combustion status that integrates flame image features and oxygen data, thereby solving the problems existing in the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a three-dimensional visualization diagnostic system for furnace combustion status that integrates flame image features and oxygen data, comprising: a feature module, which acquires a flame video stream inside the furnace, performs temporal registration on continuous video frames, extracts regional brightness change features, and obtains a flame brightness feature matrix; a fusion module, which acquires real-time oxygen ratio data for each region inside the furnace and the regional coordinates of each region in a preset furnace three-dimensional coordinate system, and obtains a regional combustion feature vector based on the real-time oxygen ratio data and the flame brightness feature matrix; and a determination module, which determines that if the oxygen ratio value in the regional combustion feature vector is lower than a preset oxygen threshold, the corresponding region is in a state of unburned component accumulation, and a first state label is obtained; if the oxygen ratio value in the regional combustion feature vector is higher than the preset oxygen threshold, the corresponding region is determined to have... In the case of excess cold air, a second state marker is obtained. The temperature module uses a multilayer sensor to calculate the actual temperature of each region inside the furnace based on the first and second state markers and the regional combustion feature vectors. It then constructs a three-dimensional true temperature distribution matrix inside the furnace based on the region coordinates and the actual temperature values. The deviation module obtains a pre-set target three-dimensional temperature distribution matrix and determines the temperature calculation deviation matrix by comparing the three-dimensional true temperature distribution matrix with the target three-dimensional temperature distribution matrix. The risk module determines that if the deviation value in the temperature calculation deviation matrix is ​​greater than a preset deviation threshold, the corresponding region is considered to have a local overheating risk, and the coordinates of the risk region are obtained. The adjustment module determines the fuel concentration adjustment amount and airflow direction adjustment amount for each region inside the furnace based on the risk region coordinates and the temperature calculation deviation matrix.

[0007] Preferably, the feature module acquires a flame video stream inside the furnace, performs temporal registration on consecutive video frames based on the furnace region mapping relationship, and extracts regional brightness change features through a convolutional neural network to obtain a flame brightness feature matrix. This includes acquiring the flame video stream, performing coordinate transformation on the flame video stream according to the furnace region mapping relationship to obtain consecutive video frames, and performing temporal registration on the consecutive video frames to obtain temporally registered video frames; extracting regional brightness change features of the temporally registered video frames through a convolutional neural network; if the regional brightness change features are greater than a preset threshold, constructing an initial matrix based on the regional brightness change features, and performing dimensionality reduction processing on the initial matrix to obtain a flame brightness feature matrix.

[0008] Preferably, the fusion module acquires real-time oxygen ratio data and the coordinates of each region in a preset three-dimensional coordinate system of the furnace. Based on the real-time oxygen ratio data and the flame brightness feature matrix, it obtains regional combustion feature vectors. This includes acquiring the real-time oxygen ratio data and coordinates of each region inside the furnace; constructing a spatial mapping matrix based on the real-time oxygen ratio data and the coordinates; extracting the oxygen distribution gradient; concatenating the oxygen distribution gradient with the flame brightness feature matrix to obtain a brightness-oxygen fusion tensor; performing principal component analysis on the brightness-oxygen fusion tensor to obtain a dimensionality-reduced feature sequence; and if the dimension of the dimensionality-reduced feature sequence is greater than a preset dimension threshold, inputting the dimensionality-reduced feature sequence into a support vector machine for classification processing to obtain regional combustion feature vectors.

[0009] Preferably, the determination module determines that if the oxygen ratio value in the regional combustion feature vector is lower than a preset oxygen threshold, the corresponding region is in a state of unburned component accumulation, and a first state label is obtained; if the oxygen ratio value in the regional combustion feature vector is higher than the preset oxygen threshold, the corresponding region is in a state of excess cold air, and a second state label is obtained. This includes acquiring the regional combustion feature vector and extracting the oxygen ratio value from the regional combustion feature vector; performing noise reduction processing on the oxygen ratio value to obtain a smoothed oxygen value; performing a difference calculation between the smoothed oxygen value and the preset oxygen threshold to obtain an oxygen deviation sequence; and determining the numerical state of the oxygen deviation sequence, including determining that if the value in the oxygen deviation sequence is less than zero, the corresponding region is in a state of unburned component accumulation, and a first state label is obtained; and determining that if the value in the oxygen deviation sequence is greater than zero, the corresponding region is in a state of excess cold air, and a second state label is obtained.

[0010] Preferably, the temperature module uses a multilayer perceptron to calculate the temperature of each region inside the furnace based on the first state marker, the second state marker, and the regional combustion feature vector. It then constructs a three-dimensional true temperature distribution matrix inside the furnace based on the region coordinates and the true temperature value of each region. This includes acquiring the first state marker, the second state marker, and the regional combustion feature vector; fusing the first state marker, the second state marker, and the regional combustion feature vector to obtain a feature fusion vector; calculating the initial temperature value of each region inside the furnace using a multilayer perceptron based on the feature fusion vector; determining whether the initial temperature value is greater than a preset temperature threshold; if the initial temperature value is greater than the preset temperature threshold, performing nonlinear smoothing on the initial temperature value to obtain the true temperature value of each region; acquiring the region coordinates of each region; performing spatial location matching based on the region coordinates and the true temperature value to obtain a spatial mapping matrix; and performing three-dimensional reconstruction based on the spatial mapping matrix to construct the three-dimensional true temperature distribution matrix inside the furnace.

[0011] Preferably, the deviation module acquires a pre-set target three-dimensional temperature distribution matrix and determines the temperature estimation deviation matrix by comparing the three-dimensional real temperature distribution matrix with the target three-dimensional temperature distribution matrix. This includes mapping the target three-dimensional temperature distribution matrix with the three-dimensional real temperature distribution matrix to obtain a target mesh temperature matrix and a real mesh temperature matrix; comparing the target mesh temperature matrix with the real mesh temperature matrix to calculate a temperature difference matrix; if the absolute value of the deviation of the temperature difference matrix is ​​greater than a preset threshold, then generating a deviation correction weight based on the temperature difference matrix; using the deviation correction weight to weight the temperature difference matrix to obtain a weighted temperature difference matrix; and determining the temperature estimation deviation matrix based on the weighted temperature difference matrix.

[0012] Preferably, in the risk module, if the deviation value in the temperature estimation deviation matrix is ​​greater than a preset deviation threshold, it determines that there is a local overheating risk in the corresponding area. Obtaining the risk area coordinates includes obtaining the deviation value of the temperature estimation deviation matrix. If the deviation value is greater than the preset deviation threshold, it determines that there is a local overheating risk and generates an initial overheating node set. Based on the initial overheating node set, a local overheating connected component is constructed, and boundary contour nodes are selected from the local overheating connected component. The boundary contour nodes are enveloped to generate a risk area polygon, and the geometric centroid of the risk area polygon is extracted as the risk area coordinates of the local overheating risk.

[0013] Preferably, the adjustment module determines the fuel concentration adjustment amount and airflow direction adjustment amount for each region inside the furnace based on the risk area coordinates and the temperature deviation matrix. This includes performing spatial grid division processing on the furnace interior region according to the risk area coordinates to obtain a three-dimensional grid topology; performing heat flux mapping processing on the three-dimensional grid topology and the temperature deviation matrix to obtain a combustion thermal distribution matrix; extracting combustion state features of independent spatial nodes using the combustion thermal distribution matrix; determining whether the combustion state features are greater than a preset feature threshold; if the combustion state features are greater than the preset feature threshold, generating a deviation compensation matrix; calculating the concentration gradient vector and airflow deflection angle based on the deviation compensation matrix; and determining the fuel concentration adjustment amount and airflow direction adjustment amount for each region inside the furnace based on the concentration gradient vector and airflow deflection angle.

[0014] Preferably, the system also includes an instruction module, which determines independent adjustment instructions for each region inside the furnace based on the fuel concentration adjustment and airflow direction adjustment. Specifically, this includes extracting the feeder speed increment and secondary damper opening based on the fuel concentration adjustment and airflow direction adjustment, and obtaining the actuator response delay. If the actuator response delay is greater than a preset delay threshold, timing alignment processing is performed on the feeder speed increment and secondary damper opening to obtain a dynamic compensation sequence. Spatial interference weights are calculated based on the dynamic compensation sequence to construct a decoupling control matrix. The decoupling control matrix is ​​used to perform instruction conversion processing on the dynamic compensation sequence to determine independent adjustment instructions for each region inside the furnace.

[0015] Preferably, the system also includes a display module, which maps the three-dimensional real temperature distribution matrix, the first state marker, the second state marker, the risk area coordinates, and the independent adjustment commands to a preset three-dimensional furnace model to generate a three-dimensional visualization diagnostic result of the furnace combustion state. Specifically, this includes acquiring the three-dimensional real temperature distribution matrix and the preset three-dimensional furnace model; performing spatial interpolation and color assignment on the three-dimensional real temperature distribution matrix based on the grid node coordinates of the furnace three-dimensional model to obtain a basic temperature rendering layer; performing local highlighting processing on the basic temperature rendering layer based on the first state marker and the second state marker to obtain a state enhancement rendering layer; generating a multi-dimensional feature fusion layer in the state enhancement rendering layer based on the risk area coordinates and the independent adjustment commands; and performing layer overlay sequence processing on the multi-dimensional feature fusion layer and the furnace three-dimensional model through a rendering viewpoint matrix to generate a three-dimensional visualization diagnostic result of the furnace combustion state.

[0016] As can be seen from the above technical solution, the present invention has the following beneficial effects: This 3D visualization diagnostic system for furnace combustion status, which integrates flame image features and oxygen data, performs temporal registration of continuous flame video frames inside the furnace and extracts regional brightness variation features using a convolutional neural network. This allows the flame image features to form a relatively stable correspondence with the actual furnace area. By integrating real-time oxygen ratio data and flame brightness features from each region, it reduces the bias caused by judging the combustion status solely based on flame brightness or single oxygen data, and regionalizes the accumulation of unburned components and excess cold air. Furthermore, it uses a multilayer perceptron to calculate the temperature of each region and constructs a 3D temperature distribution matrix inside the furnace using regional coordinates, which helps to more comprehensively reflect the temperature changes at different locations inside the furnace. By comparing the temperature distribution with the target temperature distribution, local overheating areas can be identified in a timely manner and the coordinates of the corresponding risk areas can be obtained, thereby improving the timeliness and accuracy of furnace combustion anomaly diagnosis. Furthermore, based on the risk areas and temperature deviations, the fuel concentration adjustment amount and airflow direction adjustment amount for each area can be determined, and corresponding independent adjustment commands can be generated. This helps to avoid local over-adjustment or under-adjustment problems caused by uniform adjustment using overall average parameters. At the same time, mapping the temperature distribution, combustion status markings, risk areas, and adjustment commands to the three-dimensional furnace model can intuitively display the spatial distribution of combustion status and abnormal areas inside the furnace. This provides operators with a more accurate and intuitive basis for combustion analysis and regional adjustment, thereby improving the stability, targeted adjustment, and operational safety of the furnace combustion process. Attached Figure Description

[0017] Figure 1 This is a flowchart of the present invention; Figure 2 This is a timing diagram of the present invention; Figure 3 This is a curve comparing the temperature estimation deviation of the present invention; Figure 4 This is a box plot of the absolute value of the temperature deviation of the present invention; Figure 5 This is a bar chart comparing the operating indicators before and after the regional adjustment of this invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1: like Figure 1 and Figure 5As shown, the present invention provides a technical solution: a three-dimensional visualization diagnostic system for furnace combustion status that integrates flame image features and oxygen content data, comprising: The feature module acquires the flame video stream inside the furnace, performs temporal registration of continuous video frames based on the furnace area mapping relationship, and extracts the regional brightness change features through a convolutional neural network to obtain the flame brightness feature matrix. The fusion module acquires real-time oxygen ratio data for each region inside the furnace and the region coordinates of each region in the preset three-dimensional coordinate system of the furnace. Based on the real-time oxygen ratio data and the flame brightness feature matrix, it obtains the region combustion feature vector. The determination module determines that if the oxygen ratio in the regional combustion feature vector is lower than the preset oxygen threshold, the corresponding region is determined to have an accumulation of unburned components, and a first state label is obtained; if the oxygen ratio in the regional combustion feature vector is higher than the preset oxygen threshold, the corresponding region is determined to have an excess of cold air, and a second state label is obtained. The temperature module uses a multilayer perceptron to calculate the temperature based on the first state marker, the second state marker, and the regional combustion feature vector, to obtain the true temperature values ​​of each region inside the furnace. Based on the regional coordinates and true temperature values ​​of each region, a three-dimensional true temperature distribution matrix inside the furnace is constructed. The deviation module obtains the pre-set target three-dimensional temperature distribution matrix and determines the temperature estimation deviation matrix by comparing the actual three-dimensional temperature distribution matrix with the target three-dimensional temperature distribution matrix. The risk module determines that there is a local overheating risk in the corresponding area if the deviation value in the temperature estimation deviation matrix is ​​greater than the preset deviation threshold, and obtains the coordinates of the risk area. The adjustment module calculates the deviation matrix based on the risk area coordinates and temperature, and determines the fuel concentration adjustment amount and airflow direction adjustment amount for each area inside the furnace. The instruction module determines independent adjustment instructions for each area inside the furnace based on the fuel concentration adjustment amount and the airflow direction adjustment amount. The display module maps the three-dimensional real temperature distribution matrix, the first state marker, the second state marker, the risk area coordinates, and the independent adjustment commands to the preset three-dimensional furnace model, generating a three-dimensional visualization diagnostic result of the furnace combustion state.

[0020] In this embodiment, the system first establishes a three-dimensional coordinate system for the furnace based on the actual geometry of the furnace. The three-dimensional coordinate system uses the width, depth, and height of the furnace as the three coordinate directions, and the fixed structural position inside the furnace as the origin.

[0021] Specifically, according to the pre-set area division rules, the internal space of the furnace is divided into multiple corresponding diagnostic areas. Each diagnostic area is equipped with a unique area number and area coordinates. The area coordinates are used to indicate the spatial location of the diagnostic area inside the furnace.

[0022] Furthermore, the feature module continuously acquires video streams of the flames inside the furnace through a camera device installed at the furnace observation position.

[0023] Specifically, each frame of flame image captured by the camera device is accompanied by a corresponding capture time. The feature module reads consecutive video frames in chronological order of capture time and selects one frame of flame image as the reference video frame.

[0024] In this process, due to equipment vibration, camera displacement, flue gas refraction and flame disturbance during furnace operation, the position of the same furnace structure in the image will change in consecutive video frames. The feature module extracts fixed structural features such as the edge of the furnace inner wall, the edge of the burner and the edge of the observation hole from the reference video frame and the current video frame, respectively, and then compares the positional differences of the same fixed structural features in the two frames.

[0025] In addition, the feature module determines the changes in horizontal displacement, vertical displacement, rotation angle, and image size of the current video frame relative to the reference video frame based on the positional differences of the fixed structural features, and performs translation correction, rotation correction, and scaling correction on the current video frame based on the above changes.

[0026] Furthermore, after the correction is completed, the fixed structural features in the current video frame are located in the same image position as the corresponding fixed structural features in the reference video frame, thereby completing the temporal registration between consecutive video frames.

[0027] Furthermore, the system pre-establishes a mapping relationship between the furnace areas.

[0028] Specifically, the furnace region mapping relationship records the correspondence between each furnace diagnostic region and the corresponding image range in the video frame. After completing the temporal registration, the feature module segments each frame of flame image according to the furnace region mapping relationship to obtain the region flame image that corresponds one-to-one with each furnace diagnostic region.

[0029] Furthermore, the feature module combines regional flame images belonging to the same furnace diagnostic area from multiple consecutive video frames into a regional flame image sequence according to a fixed time window.

[0030] Each region's flame image sequence is arranged according to the acquisition time of the video frames, thus preserving the order in which the flame brightness changes over time.

[0031] Furthermore, the feature module inputs the flame image sequences of each region into the convolutional neural network.

[0032] Specifically, the convolutional neural network first performs pixel brightness analysis on the regional flame image to identify bright areas, dark areas, flame boundaries, and brightness transition areas in the flame image; then, the convolutional neural network compares the brightness changes between consecutive video frames and extracts the brightness change features of each diagnostic region within the time window.

[0033] Specifically, the brightness variation characteristics include the average brightness of all pixels in the region, the brightness value of the pixel with the highest brightness in the region, the brightness value of the pixel with the lowest brightness in the region, the average brightness difference between adjacent video frames, the amount of brightness change per unit time, the degree of brightness dispersion between different positions in the region, and the amount of positional movement of the bright area in consecutive video frames.

[0034] Furthermore, the convolutional neural network arranges the brightness change features corresponding to each diagnostic region in a fixed order to form the flame brightness feature data of that diagnostic region.

[0035] Then, the feature module arranges the flame brightness feature data of all diagnostic areas in the order of area number to obtain the flame brightness feature matrix. Each row in the flame brightness feature matrix corresponds to a furnace diagnostic area, and each column corresponds to a brightness change feature.

[0036] As a preferred embodiment, the fusion module acquires real-time oxygen ratio data for each diagnostic zone inside the furnace.

[0037] Specifically, real-time oxygen ratio data is collected by oxygen detection devices located inside the furnace or connected to various areas inside the furnace. Each oxygen detection location is pre-associated with a corresponding area number and area coordinates, so that each set of oxygen ratio data can be mapped to a specific furnace diagnostic area.

[0038] In addition, the fusion module aligns the real-time oxygen ratio data with the flame brightness feature matrix according to the data acquisition time.

[0039] Specifically, the fusion module reads the acquisition time corresponding to the flame brightness feature matrix, selects the set of data with the shortest interval to the acquisition time from the oxygen ratio data, and uses the selected oxygen ratio data as the oxygen data used in the current diagnostic cycle.

[0040] Furthermore, the fusion module performs numerical standardization processing on the flame brightness characteristic data and real-time oxygen ratio data.

[0041] Standardization processes convert different types of data into a unified numerical range, eliminating numerical scale differences between flame brightness data and oxygen ratio data caused by different units of measurement.

[0042] Furthermore, after standardization, the fusion module connects the flame brightness feature data and real-time oxygen ratio data of the same diagnostic area in a fixed order to form a regional combustion feature vector for that diagnostic area. Each regional combustion feature vector is associated with a unique area number and area coordinates to represent the flame brightness change status and oxygen supply status of the corresponding area at the current moment.

[0043] The determination module stores oxygen thresholds in advance.

[0044] Specifically, the oxygen threshold is determined based on the fuel type, combustion mode, operating load, and oxygen control requirements under normal combustion conditions in the furnace; the determination module reads the oxygen ratio value in the combustion feature vector of each region one by one, and compares the oxygen ratio value with the oxygen threshold.

[0045] Specifically, when the oxygen content ratio of a diagnostic area is lower than the oxygen content threshold, the determination module determines that the diagnostic area is in a state of unburned component accumulation and generates a first state mark for the area; the first state mark is associated with the area number, area coordinates and current diagnostic time of the area.

[0046] When the oxygen content ratio of a diagnostic area is higher than the oxygen content threshold, the determination module determines that the diagnostic area has an excess of cold air and generates a second status mark for the area; the second status mark is also associated with the area number, area coordinates and current diagnosis time of the area.

[0047] When the oxygen content ratio of a diagnostic area is equal to the oxygen content threshold, the determination module does not generate a first state flag or a second state flag, and records the diagnostic area as having an oxygen content ratio in a preset baseline state.

[0048] In this embodiment, the temperature module combines the regional combustion feature vector, the first state label, and the second state label of each diagnostic region into temperature estimation input data, and inputs the temperature estimation input data into a pre-trained multilayer perceptron.

[0049] Specifically, the multilayer perceptron includes an input layer, multiple hidden layers, and an output layer. The input layer receives flame brightness variation features, oxygen supply ratio values, a first state label, and a second state label. Each hidden layer performs layer-by-layer weighting, superposition, and nonlinear transformation on the input data according to pre-trained weights and biases, thereby extracting the correlation between flame brightness variation, oxygen supply status, and furnace temperature. The output layer outputs the temperature estimation result for the corresponding diagnostic region based on the processing result of the last hidden layer.

[0050] Furthermore, the multilayer sensor is trained using historical furnace operation data before being put into use.

[0051] Specifically, the training data includes historical flame brightness characteristics, historical oxygen ratio data, historical first state markers, historical second state markers, and measured regional temperatures at the same time and in the same furnace area.

[0052] In addition, the measured temperature of the area was collected by a calibrated furnace temperature measuring device.

[0053] During training, the multilayer perceptron performs temperature estimation on historical data and compares the estimated temperature with the corresponding measured temperature in the region. The system adjusts the weights and biases of each layer within the multilayer perceptron based on the difference between the two. This process is repeated until the error between the temperature estimation result and the measured temperature in the region meets the pre-set training termination condition.

[0054] Furthermore, after training is completed, the system fixes the weights and biases of each layer within the multilayer perceptron.

[0055] Specifically, during real-time diagnosis, the temperature module inputs the current temperature estimation data of each diagnostic area into a multilayer sensor with fixed parameters, and determines the temperature estimation result output by the multilayer sensor as the true temperature value of the corresponding area.

[0056] In one possible embodiment, the temperature module writes the actual temperature value of each diagnostic area into the corresponding position in the furnace three-dimensional coordinate system according to the area number and area coordinates.

[0057] The furnace three-dimensional coordinate system forms multiple spatial positions according to preset width, depth and height intervals, and each spatial position corresponds to a matrix unit in the three-dimensional real temperature distribution matrix.

[0058] Furthermore, the temperature module writes the actual temperature value of each diagnostic area into the matrix cell corresponding to the coordinates of that area. After all the actual temperature values ​​of the diagnostic areas are written, a three-dimensional actual temperature distribution matrix inside the furnace is formed. Each matrix cell in the three-dimensional actual temperature distribution matrix corresponds to a spatial location inside the furnace and the actual temperature value of that spatial location.

[0059] In addition, the deviation module pre-stores the target three-dimensional temperature distribution matrix.

[0060] Specifically, the target three-dimensional temperature distribution matrix is ​​determined based on the temperature that each region should reach when the furnace is in the target combustion state; the target three-dimensional temperature distribution matrix and the three-dimensional real temperature distribution matrix adopt the same furnace three-dimensional coordinate system, region division rules and matrix arrangement.

[0061] Furthermore, the deviation module reads the actual temperature value in the three-dimensional true temperature distribution matrix and the target temperature in the target three-dimensional temperature distribution matrix one by one according to the position of the matrix unit; the deviation module subtracts the target temperature value at the same spatial position from the actual temperature value, and determines the difference as the temperature estimation deviation at that spatial position.

[0062] Specifically, when the temperature calculation deviation is positive, it means that the actual temperature value of the corresponding area exceeds the target temperature value; when the temperature calculation deviation is zero, it means that the actual temperature value of the corresponding area is equal to the target temperature value; and when the temperature calculation deviation is negative, it means that the actual temperature value of the corresponding area has not reached the target temperature value.

[0063] In this embodiment, the deviation module arranges the temperature estimation deviations of all regions according to the arrangement of the three-dimensional real temperature distribution matrix, thereby forming a temperature estimation deviation matrix.

[0064] In this matrix, each element of the temperature estimation deviation matrix is ​​associated with the regional coordinates and temperature estimation deviation of the corresponding region.

[0065] In addition, the risk module pre-stores deviation thresholds. The risk module reads the deviation values ​​in the temperature estimation deviation matrix one by one and compares the deviation values ​​with the deviation thresholds.

[0066] Specifically, when the deviation value of a diagnostic area is greater than the deviation threshold, the risk module determines that the diagnostic area has a local overheating risk and reads the area coordinates of the diagnostic area; the risk module determines the read area coordinates as the risk area coordinates, and records the corresponding area number, actual temperature value, target temperature value and temperature calculation deviation.

[0067] When multiple adjacent diagnostic areas are all determined to have a local overheating risk, the risk module merges the multiple diagnostic areas into a continuous risk area based on the adjacency relationship between the area coordinates, and records the coordinates of all areas contained in the continuous risk area.

[0068] Specifically, the adjustment module reads the coordinates of the risk area and the temperature calculation deviation matrix, and determines the fuel concentration adjustment amount for each furnace area according to the pre-set fuel adjustment rules; the fuel adjustment rules specify the correspondence between the temperature calculation deviation and the fuel concentration adjustment amount.

[0069] Specifically, for areas identified as having a risk of local overheating, the adjustment module determines the amount of fuel concentration reduction based on the temperature calculation deviation of that area; when the temperature calculation deviation increases by a preset unit, the amount of fuel concentration reduction increases by a corresponding preset unit; once the amount of fuel concentration reduction reaches the maximum allowable adjustment value of the fuel adjustment device, the adjustment module will no longer increase the amount of fuel concentration reduction.

[0070] In addition, for areas where there is no risk of local overheating, the adjustment module determines whether the fuel concentration should remain at the current value based on the temperature deviation of the area. When the actual temperature value of the area is equal to the target temperature value, the fuel concentration adjustment amount is set to zero. When the actual temperature value of the area is lower than the target temperature value, the adjustment module determines the amount of fuel concentration increase according to the preset fuel adjustment rules.

[0071] Furthermore, the adjustment module also determines the airflow direction adjustment amount for each region based on the temperature-calculated deviation matrix.

[0072] Specifically, the adjustment module compares the temperature calculation deviations of a diagnostic area with its adjacent areas in the width direction, depth direction, and height direction, and determines the temperature deviation changes in each direction.

[0073] In this embodiment, the adjustment module determines the direction in which the temperature calculation deviation value decreases as the target airflow adjustment direction, so that the airflow is adjusted from the area with a large temperature calculation deviation value to the area with a small temperature calculation deviation value; the adjustment module determines the airflow adjustment amplitude based on the temperature calculation deviation difference between adjacent areas; when the temperature calculation deviation difference increases by a preset unit, the airflow adjustment amplitude increases by a corresponding preset unit.

[0074] Furthermore, once the airflow adjustment range reaches the maximum allowable adjustment value of the airflow regulating device, the regulating module will no longer increase the airflow adjustment range.

[0075] Specifically, the adjustment module determines the airflow adjustment components in the width, depth and height directions respectively, and combines the airflow adjustment components in the three directions into the airflow direction adjustment amount for the corresponding area.

[0076] In addition, the adjustment module performs amplitude limiting on the fuel concentration adjustment and airflow direction adjustment.

[0077] Specifically, the fuel concentration adjustment amount shall not exceed the allowable adjustment range of the fuel regulating device, and the airflow direction adjustment amount shall not exceed the allowable operating range of the damper, guide vane or air distribution device; when the calculated adjustment amount exceeds the allowable adjustment range, the regulating module will set the adjustment amount to the maximum allowable adjustment value of the corresponding actuator.

[0078] In one aspect, the instruction module reads the area number, fuel concentration adjustment amount, and airflow direction adjustment amount of each area, and combines the above data into an independent adjustment instruction for the corresponding area; each independent adjustment instruction includes the target area number, fuel adjustment direction, fuel adjustment value, airflow adjustment direction, and airflow adjustment value.

[0079] Furthermore, the instruction module sends independent adjustment instructions to the fuel regulating device and airflow regulating device in the corresponding area based on the target area number.

[0080] Specifically, the fuel regulating device changes the fuel supply parameters of the corresponding area according to the fuel adjustment direction and fuel adjustment value; the airflow regulating device changes the airflow direction and airflow distribution parameters of the corresponding area according to the airflow adjustment direction and airflow adjustment value.

[0081] In this embodiment, the display module reads the three-dimensional real temperature distribution matrix, the first state marker, the second state marker, the risk area coordinates, and the independent adjustment command, and maps the above data to the preset three-dimensional furnace model according to the area coordinates.

[0082] The display module configures the display colors for each area in the three-dimensional model of the furnace according to the pre-set correspondence between temperature ranges and colors; each temperature range corresponds to a fixed color, enabling operators to identify the actual temperature range of each area based on the color.

[0083] The display module displays a first status indicator in areas where unburned components accumulate, a second status indicator in areas where there is an excess of cold air, and a risk boundary indicator in areas where there is a risk of local overheating.

[0084] In addition, the display module displays the fuel adjustment value in each area that needs to be adjusted, and uses directional arrows to indicate the direction of airflow adjustment and numerical markers to indicate the magnitude of airflow adjustment.

[0085] To verify this application, a two-stage experimental verification scheme of "refined verification on a scaled-down hot test bench + on-site industrial verification of power plant boiler" was adopted.

[0086] I. Experimental Platform and Overall Scheme 1. Scaled-down hot test bench (Phase 1) The test bench is a 1:8 scale furnace hot combustion test device with furnace inner cavity dimensions of 2.4m (height) × 0.9m (width) × 0.9m (depth). It adopts a natural gas / pulverized coal dual-purpose burner, which can stably reproduce the zonal air distribution and stratified coal feeding conditions.

[0087] K-type armored thermocouples (measuring range 0~1300℃, accuracy ±0.75%) are arranged in a grid of 32 nodes (4×4×2) on the four walls and cross-section of the furnace to serve as the true temperature reference value. Three industrial endoscopic high-temperature cameras (25fps, 1920×1080 resolution) were placed at different heights in the furnace to capture flame video streams. A flue gas extraction zirconia oxygen analysis probe (response time < 3s, accuracy ±0.1% O2) was inserted near 32 grid nodes corresponding to the thermocouples to obtain real-time oxygen ratio data for each region. By using the air distribution regulating valve and the pulverizer frequency converter, six typical operating conditions (normal air distribution, local oxygen deficiency, local oxygen enrichment, unilateral combustion, overall overheating, and dynamic disturbance) were artificially set to verify the system's ability to identify and regulate under various combustion anomalies.

[0088] 2. On-site industrial validation (Phase Two) A 30-day field comparison test was conducted using a 220t / h circulating fluidized bed / pulverized coal co-firing boiler. The furnace was arranged in 4 layers along the height direction, with 6 measuring points in each layer, for a total of 24 grid nodes. The test was jointly verified using the existing DCS system's flame television and oxygen analyzer (2 points per layer, 8 points in total) and a newly added portable infrared temperature measuring array (for phased verification, with a measurement error of ±1.5%).

[0089] The on-site data collection frequency was 1 time / min, and the data was collected continuously for 30 days (a total of 43,200 regional combustion feature vector samples) for statistical significance testing.

[0090] 3. Comparison Methods To demonstrate the improvements brought about by system "integration," the experiment set up three diagnostic methods for horizontal comparison under the same operating conditions: Method A (traditional oxygen method): The temperature and combustion state are estimated based solely on the oxygen ratio data of each region, combined with empirical formulas. Method B (traditional image method): Estimates temperature solely based on the flame brightness feature matrix through gray-temperature empirical mapping; Method C (the fusion diagnostic method of this patent): According to the process of this application, the flame brightness feature matrix and oxygen ratio data are fused, and the temperature is calculated by a multilayer sensor, the deviation is compared, the risk is located and the three-dimensional visualization is output.

[0091] All three methods use the thermocouple measured temperature (stage one) and the infrared temperature measurement array verification data (stage two) as the true value benchmark.

[0092] II. Experiment 1: Verification of the Accuracy of Regional Temperature Estimation 1. Experimental Methods Under six operating conditions on a scaled-down hot test bench, temperature calculations were performed on 32 grid nodes using methods A, B, and C, respectively. These calculations were compared with the actual thermocouple measurements to calculate the mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²). 2 After each operating condition has been running stably for 20 minutes, 10 sets of samples are continuously taken and the average value is calculated to reduce the impact of transient fluctuations.

[0093] 2. Experimental Data Table 1. Comparison of temperature estimation accuracy of three methods under six operating conditions (32 measuring points, unit: °C) Data shows that the MAE (Magnitude Emission) of methods A and B, using a single information source, was in the range of 38–64℃ under six operating conditions. Furthermore, the error increased significantly with increasing operating condition disturbances (unilateral burn-off, dynamic disturbances), indicating that relying solely on oxygen content or image grayscale for temperature estimation has limited resistance to disturbances. Method C of this patent, through tensor splicing of flame brightness feature matrix and oxygen ratio data, principal component dimensionality reduction, and joint calculation using a multilayer perceptron, stabilized the MAE in all operating conditions between 14.8–22.4℃, with the mean value decreasing by approximately 63.9% and 63.6% compared to methods A and B, respectively. Moreover, R… 2 All values ​​are above 0.94, indicating that the temperature calculation results are highly consistent with the actual temperature field.

[0094] III. Experiment 2: Verification of Combustion State Regionalization 1. Experimental Methods Using the actual combustion states (accumulation of unburned components / normal / excess cold air) calibrated by thermocouples and oxygen probes as labels, the accuracy, precision, recall, and F1 score of method C (the judgment module of this patent) were statistically analyzed on 192 samples (32 nodes × 6 conditions) under 6 working conditions on a scaled platform. The results were compared with those obtained by using only a single oxygen threshold (method A simplified criterion).

[0095] 3. Experimental Data Table 2 Comparison of combustion state recognition performance (192 regional samples) The single oxygen threshold criterion is prone to misjudgment (misjudgment rate of 21.4%) under conditions such as partial flame obstruction and probe response lag. However, the method of this patented patent improves the overall recognition accuracy to 93.8% by fusing the regional combustion feature vector composed of flame brightness features and oxygen distribution gradient, and combining denoising and threshold difference calculation to determine the first / second state marker. The number of misjudged areas is reduced from 41 to 12, reducing the deviation caused by judging the combustion state based solely on flame brightness or single oxygen data.

[0096] IV. Experiment 3: Verification of Local Overheat Risk Location 1. Experimental Methods Three known overheated areas were artificially created under two working conditions: overall overheating and unilateral overheating on the scaled platform (the center coordinates were confirmed by actual measurement using thermocouples). The risk module of this patent (connected domain construction, boundary contour extraction, envelope processing, and centroid extraction) was used to calculate the coordinates of the risk areas and compare them with the actual overheated center coordinates to calculate the three-dimensional positioning deviation (Euclidean distance).

[0097] 2. Experimental Data Table 3. Location accuracy of local overheating risk areas (unit: mm, furnace scale coordinate system) The three-dimensional positioning deviation of the three overheated areas is 22.7–25.9 mm, which translates to approximately 180–210 mm on the actual furnace scale at a 1:8 scale. This is within an acceptable range compared to the meter-level grid division accuracy of typical power plant boiler furnaces. Moreover, the average positioning time is only 1.8 s, meeting the real-time requirements of online diagnosis, promptly identifying local overheated areas and obtaining the coordinates of corresponding risk areas, thereby improving the timeliness and positioning accuracy of furnace combustion anomaly diagnosis.

[0098] V. Experiment 4: Verification of Regional Adjustment Effect (30-day Industrial Trial) 1. Experimental Methods At the 220t / h boiler site, 30 days of operating data were divided into two phases: the first 15 days maintained the original overall average parameter uniform adjustment method (i.e., traditional adjustment, referred to as before adjustment); the latter 15 days used the independent adjustment commands output by this patented system to perform regional closed-loop adjustment of the pulverizer speed and secondary air damper opening (referred to as after adjustment). Daily operating indicators such as the standard deviation of the furnace cross-section temperature field (measuring temperature uniformity), fly ash carbon content, flue gas CO concentration, and boiler thermal efficiency were extracted and compared using the 15-day mean ± standard deviation.

[0099] 2. Experimental Data Table 4 Comparison of operational indicators before and after regional adjustments (15-day mean ± standard deviation, n=15) Table 5. Results of independent samples t-tests for key indicators before and after adjustment (n1=n2=15, two-tailed test) After adjustment, the standard deviation of the furnace cross-sectional temperature field decreased from 68.4 ℃ to 31.6 ℃, a reduction of 53.8%. This indicates that after the system generates regional fuel concentration adjustment and airflow direction adjustment based on the risk area coordinates and temperature deviation, it effectively avoids local over-adjustment or under-adjustment caused by overall average parameter adjustment, resulting in a more uniform temperature distribution within the furnace. The carbon content of fly ash and the CO concentration in flue gas decreased simultaneously, indicating that regional fine-tuning improved fuel burnout and reduced local oxygen-rich / oxygen-deficient areas. The boiler thermal efficiency increased by 1.4 percentage points, and the number of over-temperature alarms decreased from 9 to 2. The t-test results showed that the differences in all indicators were significant at the p<0.001 level, ruling out the possibility of random fluctuations leading to improvement, avoiding local over-adjustment or under-adjustment, improving the stability of the furnace combustion process, the targeting of adjustments, and operational safety.

[0100] Example 2: The feature module acquires the flame video stream inside the furnace, performs temporal registration on continuous video frames based on the furnace region mapping relationship, and extracts regional brightness change features through a convolutional neural network to obtain a flame brightness feature matrix. This includes acquiring the flame video stream, performing coordinate transformation on the flame video stream according to the furnace region mapping relationship to obtain continuous video frames, and performing temporal registration on the continuous video frames to obtain temporally registered video frames; extracting regional brightness change features from the temporally registered video frames through a convolutional neural network; if the regional brightness change features are greater than a preset threshold, constructing an initial matrix based on the regional brightness change features, and performing dimensionality reduction on the initial matrix to obtain the flame brightness feature matrix.

[0101] In this embodiment, the video stream of the flame inside the furnace is collected by a camera device fixedly installed at the furnace observation port.

[0102] The camera device is set to a frame rate of 25 frames per second and an output image resolution of 1920 pixels by 1080 pixels. Each frame is accompanied by a timestamp generated by the same clock source. The feature module decodes the flame video stream according to the order of the timestamps to obtain the original flame image sequence.

[0103] Furthermore, after the camera device is installed, the correspondence between the camera's shooting range and the furnace area is calibrated.

[0104] During the specific calibration, fixed calibration points are selected on the inner wall of the furnace, the edge of the burner outlet, and the edge of the observation port. The regional coordinates of each calibration point in the three-dimensional model of the furnace are recorded, and the pixel coordinates of the same calibration point are determined in the calibration image acquired by the camera device. Four boundary calibration points are selected for each furnace diagnostic area, and the four boundary calibration points correspond to the four corner points of the diagnostic area in the video image.

[0105] In this embodiment, the four sets of furnace area coordinates and image pixel coordinates are substituted into the perspective projection relationship to solve for eight coordinate transformation coefficients used to describe lateral displacement, longitudinal displacement, image rotation, image scaling and perspective distortion.

[0106] The solution process is as follows: Substitute the four sets of corresponding coordinates into the coordinate transformation relationship to form eight linear equations containing eight unknown coefficients. Use Gaussian elimination to eliminate the unknown terms in each equation in turn to obtain eight coordinate transformation coefficients. Store the coordinate transformation coefficients, diagnostic region numbers and diagnostic region coordinates together to form the furnace region mapping relationship.

[0107] Furthermore, after reading the original flame image, the feature module performs coordinate transformation on each frame of the original flame image according to the furnace area mapping relationship.

[0108] The coordinate transformation uses the furnace diagnostic area as the processing unit, converting the quadrilateral image range corresponding to each diagnostic area into a square area image of 224 pixels by 224 pixels.

[0109] During coordinate transformation, the feature module calculates the corresponding position of each pixel in the original flame image in reverse, based on the position of each pixel in the transformed region image.

[0110] Specifically, when the corresponding position falls between four original pixels, the horizontal and vertical distances between the position and the four original pixels are calculated respectively, and the interpolation weights of the four original pixels are determined based on the distances. The smaller the distance of the original pixel, the larger the interpolation weight. The sum of the four interpolation weights is set to 1. The brightness values ​​of the four original pixels are multiplied by the corresponding interpolation weights respectively, and the results are summed to obtain the brightness value of the converted pixel. After the calculation of all pixels is completed, a standard area image corresponding one-to-one with the furnace diagnosis area is obtained. The feature module arranges the standard area images according to the timestamp to obtain continuous video frames.

[0111] Furthermore, after the coordinate transformation of consecutive video frames is completed, the feature module performs temporal registration on them; each time 50 consecutive video images are processed, and the first frame of the 50 frames is determined as the reference frame, and the remaining 49 frames are used as the frames to be registered in sequence.

[0112] For the reference frame and the frame to be registered, the brightness values ​​of the red channel, green channel and blue channel are multiplied by a preset conversion coefficient, and then the products of the three terms are added together to obtain the grayscale image. The conversion coefficient for the red channel is set to 0.299, the conversion coefficient for the green channel is set to 0.587, and the conversion coefficient for the blue channel is set to 0.114.

[0113] In this embodiment, the feature module extracts 300 corner points from the grayscale image.

[0114] The process for determining corner points is as follows: Select a circular neighborhood with a radius of 3 pixels centered on each pixel, and compare the brightness difference between each pixel in the circular neighborhood and the center pixel; when the absolute value of the brightness difference between 12 consecutive neighboring pixels and the center pixel is greater than 20 gray units, the center pixel is recorded as a candidate corner point; arrange the candidate corner points in descending order of the sum of the absolute values ​​of the brightness differences, and retain the top 300 candidate corner points.

[0115] Furthermore, for each corner point, the feature module selects 256 sets of pixel pairs within a 31-pixel by 31-pixel range centered on that corner point, and compares the brightness values ​​of the two pixels in each set of pixel pairs.

[0116] Specifically, when the brightness value of the first pixel is greater than the brightness value of the second pixel, the corresponding position is recorded as 1; when the brightness value of the first pixel is less than or equal to the brightness value of the second pixel, the corresponding position is recorded as 0; after comparing 256 sets of pixel pairs in sequence, a binary feature description code with a length of 256 bits is formed.

[0117] Furthermore, each corner point in the frame to be registered is compared one by one with the corner points in the reference frame.

[0118] During comparison, a bitwise XOR operation is performed on the binary feature description codes of the two corner points, and the number of different values ​​is counted; the corner point pair with the fewest different values ​​is determined as a candidate matching point pair; when the number of different values ​​is greater than 40, the candidate matching point pair is deleted.

[0119] In this embodiment, the feature module uses a random sampling consistency method to screen candidate matching point pairs.

[0120] Specifically, each time, four sets of matching points are selected from the candidate matching point pairs. A set of perspective transformation parameters is calculated based on the four sets of matching points. Then, the remaining candidate matching points in the frame to be registered are mapped to the reference frame using the set of perspective transformation parameters. When the pixel distance between the mapped position and the corresponding corner point in the reference frame is no more than 2 pixels, the corresponding candidate matching point pair is recorded as a valid matching point pair.

[0121] The above sampling calculation is performed 200 times consecutively, and the perspective transformation parameters with the largest number of valid matching point pairs are selected. The feature module then uses this set of perspective transformation parameters and all valid matching point pairs to recalculate the final perspective transformation parameters, and performs translation, rotation, scaling and perspective correction on the frame to be registered according to the final perspective transformation parameters. The corrected image still determines the brightness value of each pixel according to the bilinear interpolation method.

[0122] Furthermore, after processing 49 frames to be registered sequentially, the edges of the furnace inner wall, the burner outlet, and the observation port occupy the same pixel position in 50 frames of images, thus obtaining time-series registered video frames.

[0123] Specifically, the feature module extracts the corresponding regional images of each diagnostic region from 50 time-registered video frames according to the furnace region mapping relationship; for the same diagnostic region, the corresponding pixels in two adjacent regional images are subtracted in time order, and the absolute value of the subtraction result is taken to obtain 49 brightness difference images; each pixel value in the brightness difference image is used to characterize the amount of brightness change at the same image position at adjacent acquisition times.

[0124] As a preferred embodiment, 49 frames of brightness difference images are sequentially input into a convolutional neural network.

[0125] Specifically, the convolutional neural network has 5 convolutional layers, and the kernel size of each convolutional layer is 3 pixels by 3 pixels; the first convolutional layer has 16 convolutional kernels, the second convolutional layer has 32 convolutional kernels, the third convolutional layer has 64 convolutional kernels, the fourth convolutional layer has 64 convolutional kernels, and the fifth convolutional layer has 64 convolutional kernels; after each convolutional layer completes the convolution operation, the output values ​​less than 0 are set to 0, and the output values ​​greater than or equal to 0 are retained.

[0126] Among them, after the first to the fourth convolutional layers, maximum pooling with a size of 2 pixels by 2 pixels is set respectively.

[0127] Specifically, max pooling divides the feature map into non-overlapping 2-pixel regions and retains the feature value with the largest value in each region; the fifth convolutional layer outputs 64 feature maps, and the arithmetic mean of all feature values ​​in each feature map is calculated to obtain 64 brightness change feature values.

[0128] Furthermore, the convolutional neural network is trained before the equipment is put into operation.

[0129] Specifically, the training samples consist of 5,000 sets of temporally registered region images, with each set including 50 consecutive frames of region images. For each set of training samples, the average pixel brightness, bright area change, and brightness centroid displacement of 49 frames of brightness difference images are calculated, and the three calculation results are normalized to a value range of 0 to 1 as training supervision values.

[0130] In this embodiment, the change in bright area is determined as follows: pixels with a value greater than 30 gray units in the brightness difference image are recorded as changed pixels. The number of changed pixels in each frame of the brightness difference image is counted. The number of changed pixels in the previous frame is subtracted from the number of changed pixels in the next frame, and the absolute value of the difference is taken.

[0131] The process for determining the displacement of the brightness centroid is as follows: multiply the horizontal and vertical coordinates of each changed pixel by the brightness difference value of the corresponding pixel, sum all the products, and then divide them by the sum of the brightness difference values ​​of all changed pixels to obtain the horizontal and vertical coordinates of the brightness centroid; calculate the straight-line distance between the brightness centroids in two adjacent brightness difference images, and calculate the arithmetic mean of the distances corresponding to the 49 brightness difference images.

[0132] In addition, during training, the output of the convolutional neural network is subtracted from the three training supervision values ​​one by one, the differences are squared, and the arithmetic mean of the three squared results is calculated to obtain the training error of a single training sample. The parameters of each convolutional kernel are adjusted in reverse according to the training error. When the verification error of 10 consecutive training rounds is no greater than 0.01, training is stopped and all parameters of the convolutional neural network are fixed. The 64 feature values ​​obtained after averaging the fifth convolutional layer of the convolutional neural network are used as the regional brightness change features.

[0133] Specifically, for 49 frames of brightness difference images of the same diagnostic region, the convolutional neural network outputs 49 sets of regional brightness change features respectively; the feature module summarizes the 49 sets of regional brightness change features according to the feature location.

[0134] Specifically, the values ​​at the first feature position in the 49 sets of regional brightness change features are added together and then divided by 49 to obtain the first summary feature value; the remaining 63 summary feature values ​​are calculated in the same way; the resulting 64 summary feature values ​​constitute the regional brightness change features of the diagnostic region in the current processing cycle.

[0135] In addition, in order to determine whether the regional brightness change features meet the matrix construction conditions, the feature module squares each of the 64 aggregated feature values, sums the 64 squared results, divides the sum by 64, and then takes the square root of the result to obtain the brightness change intensity value corresponding to the diagnostic region; the brightness change intensity value is used as a value to be directly compared with the preset threshold.

[0136] As a preferred embodiment, the preset threshold is determined separately for each diagnostic region.

[0137] Specifically, after the furnace is installed and commissioned, flame video is continuously collected for 30 minutes under the target combustion conditions. The target combustion conditions are defined as the furnace load maintaining the set value, the temperature change per minute in each diagnostic area not exceeding 2 degrees Celsius within 30 minutes, the oxygen ratio change per minute in each diagnostic area not exceeding 0.2 numerical units, and the furnace control device not generating a combustion abnormality alarm.

[0138] In this embodiment, a 30-minute flame video is processed by forming a sample group every 50 frames, and 900 sets of brightness change intensity values ​​are obtained for each diagnostic area.

[0139] The process involves summing 900 sets of brightness change intensity values ​​and dividing by 900 to obtain the baseline average value for the diagnostic region. Then, each set of brightness change intensity values ​​is subtracted from the baseline average value, and the resulting 900 differences are squared. The sum of these squared values ​​is then divided by 900, and the square root of the result is taken to obtain the baseline standard deviation for the diagnostic region. The baseline average value is then added to three times the baseline standard deviation, and the resulting value is the preset threshold for the diagnostic region. The preset threshold is stored corresponding to the region number and remains unchanged during operation due to the current video data.

[0140] In addition, the feature module compares the intensity value of the brightness change in each diagnostic area with the preset threshold corresponding to that area.

[0141] Specifically, when the intensity of the brightness change is greater than a preset threshold, the brightness change features of the corresponding 64 regions are retained; when the intensity of the brightness change is equal to or less than the preset threshold, the brightness change features of the region are not written into the initial matrix of the current processing cycle.

[0142] Furthermore, after comparing all diagnostic regions, the feature module arranges the diagnostic regions that meet the matrix construction conditions in ascending order of region number.

[0143] Specifically, each diagnostic region corresponds to one row in the initial matrix, and the brightness change features of the 64 regions are written into 64 columns according to the output order of the convolutional neural network, thus forming the initial matrix; the number of rows in the initial matrix is ​​equal to the number of diagnostic regions whose brightness change intensity value is greater than a preset threshold in the current processing cycle, and the number of columns is fixed at 64; the correspondence between the row number of the initial matrix and the diagnostic region number is stored separately for subsequent fusion module to call.

[0144] As a preferred implementation, the initial matrix is ​​reduced in dimensionality using principal component analysis.

[0145] Specifically, the dimensionality reduction parameters are determined based on historical flame video data before the equipment is put into operation. When the parameters are determined, 10,000 sets of regional brightness change features are extracted from the historical data and arranged into a training matrix in 64 columns. The arithmetic mean of each column in the training matrix is ​​calculated, and then the corresponding arithmetic mean is subtracted from each feature value in the column to obtain the centered training matrix.

[0146] Specifically, for any two columns in the centered training matrix, the feature values ​​in the same row of the two columns are multiplied one by one, and the sum of 10,000 products is divided by 9,999 to obtain the covariance between the two columns. The pairwise calculation between the 64 features is completed in the same way to form a 64-row, 64-column covariance matrix.

[0147] Furthermore, the covariance matrix is ​​decomposed into eigenvalues ​​to obtain 64 eigenvalues ​​and 64 sets of characteristic direction coefficients corresponding to each eigenvalue.

[0148] Specifically, 64 sets of feature direction coefficients are arranged in descending order of feature values, and the first 12 sets of feature direction coefficients are retained; the 12 sets of feature direction coefficients constitute fixed dimension reduction projection parameters; the arithmetic mean of each column of the training matrix and the dimension reduction projection parameters are written into the feature module.

[0149] In addition, when processing the initial matrix in real time, the feature module first subtracts the training arithmetic mean of the corresponding column from the eigenvalue of each column in the initial matrix to obtain the centered initial matrix.

[0150] Specifically, for each row in the centered initial matrix, the 64 eigenvalues ​​of that row are multiplied by the corresponding 64 coefficients in the first set of eigendirection coefficients, and then the 64 products are added together to obtain the first dimensionality-reduced eigenvalue. The above process is repeated for the remaining 11 sets of eigendirection coefficients to obtain a total of 12 dimensionality-reduced eigenvalues.

[0151] Furthermore, after processing all rows in the centralized initial matrix in the same way, a flame brightness feature matrix is ​​formed.

[0152] The flame brightness feature matrix has the same number of rows as the initial matrix, and the number of columns is fixed at 12. Each row still corresponds to the same diagnostic region, and the 12 columns record the brightness change features of that region after dimensionality reduction.

[0153] Example 3: The fusion module acquires real-time oxygen ratio data and the coordinates of each region in a preset three-dimensional coordinate system within the furnace. Based on the real-time oxygen ratio data and the flame brightness feature matrix, it obtains regional combustion feature vectors. This includes acquiring real-time oxygen ratio data and regional coordinates for each region within the furnace, constructing a spatial mapping matrix based on the real-time oxygen ratio data and regional coordinates, and extracting the oxygen distribution gradient. The oxygen distribution gradient and the flame brightness feature matrix are then tensor-concatenated to obtain a brightness-oxygen fusion tensor. Principal component analysis is performed on the brightness-oxygen fusion tensor to obtain a dimensionality-reduced feature sequence. If the dimension of the dimensionality-reduced feature sequence is greater than a preset dimension threshold, the dimensionality-reduced feature sequence is input into a support vector machine for classification processing to obtain regional combustion feature vectors.

[0154] In this embodiment, the internal space of the furnace is pre-divided into regions according to the width, depth and height directions of the furnace.

[0155] In this embodiment, the furnace is divided into 8 regions along the width direction, 4 regions along the depth direction, and 2 regions along the height direction, forming a total of 64 furnace regions. Each furnace region is assigned a unique region number, and the region coordinates are the coordinates of the geometric center of the region in the three-dimensional coordinate system of the furnace.

[0156] Furthermore, the oxygen detection device collects oxygen ratio data for each furnace area at a sampling frequency of 10 times per second.

[0157] The fusion module uses 2 seconds as one data processing cycle, reads 20 oxygen ratio values ​​obtained from the same area within the processing cycle, adds the 20 oxygen ratio values ​​together, and divides the sum by 20. The resulting value is used as the real-time oxygen ratio data of the area within the current processing cycle.

[0158] Specifically, the fusion module arranges the data of each region in the order of region number from 1 to 64, and each region corresponds to 1 row in the spatial mapping matrix; each row is written with the region number, width direction coordinate, depth direction coordinate, height direction coordinate and real-time oxygen ratio data in sequence, so that each real-time oxygen ratio data corresponds to a certain position in the three-dimensional coordinate system of the furnace.

[0159] Furthermore, after the spatial mapping matrix is ​​formed, the fusion module calculates the changes in real-time oxygen ratio data along the width, depth, and height directions, respectively.

[0160] Specifically, for a region located within the width direction, the real-time oxygen ratio data of the adjacent regions in the forward and reverse directions of the region are read. The real-time oxygen ratio data of the adjacent regions in the forward direction is subtracted from the real-time oxygen ratio data of the adjacent regions in the reverse direction. The difference is then divided by the coordinate distance between the geometric centers of the two adjacent regions to obtain the change in oxygen content along the width direction of the region.

[0161] Specifically, for a region located at the width boundary, one adjacent region is read, and the oxygen ratio data of the region arranged in the reverse direction of the width coordinate is subtracted from the oxygen ratio data of the region arranged in the positive direction of the width coordinate. The difference is then divided by the coordinate distance between the geometric centers of the two regions to obtain the change in oxygen content of the boundary region along the width direction.

[0162] In this embodiment, the changes in oxygen content in the depth and height directions are calculated in the same way.

[0163] Specifically, the fusion module arranges the three oxygen quantity changes in the order of width, depth and height to form the oxygen quantity distribution gradient of the corresponding region. Positive numbers in the oxygen quantity distribution gradient indicate that the oxygen quantity ratio data increases along the corresponding coordinate, negative numbers indicate that the oxygen quantity ratio data decreases along the corresponding coordinate, and a value of 0 indicates that the oxygen quantity ratio data of adjacent regions in the corresponding direction is the same.

[0164] In the formed flame brightness feature matrix, each row is associated with a region number, and each row contains 12 flame brightness features. The fusion module matches each row in the flame brightness feature matrix with the oxygen distribution gradient of the same region based on the region number, and does not directly perform data connection according to the matrix row order.

[0165] Furthermore, after completing the regional correspondence, the fusion module arranges the 12 flame brightness features of the same region first, and arranges the 3 oxygen quantity changes in the width, depth and height directions in sequence, forming a single-cycle fusion feature containing 15 data items.

[0166] In addition, the fusion module continuously saves data for 10 processing cycles. Each processing cycle lasts for 2 seconds, therefore, 10 processing cycles correspond to 20 consecutive seconds of furnace operation data.

[0167] Specifically, the fusion module arranges 10 sets of single-cycle fusion features in the order of acquisition time from front to back to form a brightness-oxygen fusion tensor. The first data direction of this tensor represents the furnace area, the second data direction represents 15 fusion features, and the third data direction represents 10 processing cycles.

[0168] Before principal component analysis in this embodiment, the fusion module expands the brightness-oxygen fusion tensor corresponding to each region into a feature sequence.

[0169] During the expansion, the 15 fusion features of the first processing cycle are written first, then the 15 fusion features of the second processing cycle are written, and so on until the 15 fusion features of the tenth processing cycle are written. Each region thus forms an original feature sequence containing 150 data items.

[0170] Furthermore, the fusion module calls the feature mean and feature standard deviation saved during the training phase to standardize the original feature sequence.

[0171] Specifically, for each feature location, the current feature value at that location is subtracted from the arithmetic mean of the training samples at the same location, and the difference is divided by the standard deviation of that location to obtain the standardized feature value. When the standard deviation of the training samples at a certain feature location is 0, it indicates that the feature location has not changed in value among all training samples, and the feature location is not included in subsequent calculations.

[0172] In this embodiment, the principal component projection parameters are determined before the equipment is put into operation.

[0173] Specifically, 3,000 sets of brightness and oxygen content fusion tensors were selected as training data. Each set of training data was expanded into 150 features and standardized in the manner described above.

[0174] Specifically, for any two feature locations in the standardized training data, the values ​​of the same training sample at the two feature locations are multiplied together, the sum of 3000 products is calculated, and the sum is divided by 2999 to obtain the covariance between the two feature locations. The same processing method is used to complete the calculation between all feature locations to form the covariance matrix.

[0175] In this process, the covariance matrix is ​​decomposed into eigenvalues ​​to obtain the directions of each principal component and their corresponding eigenvalues.

[0176] Specifically, the principal components are arranged in order of eigenvalue from largest to smallest, and the eigenvalue is accumulated one by one starting from the first eigenvalue. The accumulation stops when the ratio of the accumulated value to the sum of all eigenvalues ​​reaches 0.95 for the first time.

[0177] In this embodiment, the cumulative ratio of the first 12 principal component directions reaches 0.953, so 12 principal component directions are retained.

[0178] Furthermore, the fusion module projects 150 standardized features of each region onto the aforementioned 12 principal component directions.

[0179] In the specific projection, the 150 standardized features are multiplied one by one with the 150 directional coefficients at the corresponding positions in the same principal component direction, and then the 150 products are added together to obtain one dimensionality-reduced feature corresponding to the principal component direction. After projecting the 12 principal component directions in sequence, a dimensionality-reduced feature sequence containing 12 dimensionality-reduced features is obtained.

[0180] In this embodiment, the preset dimension threshold is determined during the support vector machine training phase.

[0181] Specifically, the 3000 sets of training data are divided into 5 equal parts, each containing 600 sets of data. Each time, 4 parts are selected for model training, and the remaining part is used for classification verification. This process is repeated 5 times to ensure that each set of data is verified once.

[0182] Specifically, the input dimensions of the dimensionality-reduced feature sequences were set to 4 to 12, support vector machines were trained one by one, and the classification accuracy corresponding to each input dimension was calculated.

[0183] Specifically, the classification accuracy is determined by dividing the number of correctly classified samples by 3000; when the input dimension is 10, the classification accuracy is 0.981; when the input dimension is 11, the classification accuracy is 0.983; when the input dimension is 12, the classification accuracy is 0.984; the input dimension corresponding to the first time the classification accuracy reaches 0.98 is determined as the preset dimension threshold. Therefore, the preset dimension threshold in this embodiment is set to 10.

[0184] When the real-time obtained dimensionality reduction feature sequence contains 12 dimensionality reduction features, its dimension is greater than the preset dimension threshold. The fusion module then inputs the dimensionality reduction feature sequence into the support vector machine for classification processing.

[0185] As a preferred embodiment, the support vector machine is configured with three binary classifiers to distinguish between the baseline combustion state and the unburned component accumulation state, the baseline combustion state and the cold air excess state, and the unburned component accumulation state and the cold air excess state, respectively. Each binary classifier uses a radial basis function kernel.

[0186] During model training, the penalty parameters were set to 0.25, 0.5, 1, 2, 4, 8, and 16, respectively, and the kernel function parameters were set to 0.001, 0.005, 0.01, 0.05, 0.1, 0.5, and 1, respectively. Cross-validation was performed 5 times for each of the 49 parameter combinations. The parameter combination with the highest classification accuracy was used as the running parameter for the corresponding binary classifier. When the classification accuracy of two parameter combinations was the same, the parameter combination with the smaller penalty parameter value was selected.

[0187] During the classification process, the support vector machine calculates the squared distance between the current dimensionality-reduced feature sequence and each support vector, multiplies the squared distance by the negative of the kernel function parameter, and then performs an exponential operation on the resulting value to obtain the kernel function value between the current dimensionality-reduced feature sequence and the corresponding support vector.

[0188] In this process, each kernel function value is multiplied by its corresponding support vector coefficient, and the sum of all products is then added to the bias value to obtain the classification decision value of the binary classifier.

[0189] In this embodiment, each binary classifier determines a state category based on the sign of the classification decision value and casts one vote for that state category.

[0190] After the three binary classifiers complete their processing, the state category with the largest number of votes is determined as the classification result for the current region. When two state categories have the same number of votes, the output of the binary classifier used to distinguish the two state categories is read, and the classification result is determined according to the positive or negative value of the classification decision.

[0191] Furthermore, the fusion module arranges the real-time oxygen ratio data of the current region, 3 oxygen distribution gradients, 12 dimensionality reduction features, classification decision values ​​of 3 binary classifiers and classification result numbers in a fixed order to form a regional combustion feature vector; the classification result number of the baseline combustion state is set to 0, the classification result number of the unburned component accumulation state is set to 1, and the classification result number of the cold air excess state is set to 2.

[0192] Example 4: The determination module determines the presence of unburned components in the corresponding region if the oxygen ratio value in the regional combustion feature vector is lower than a preset oxygen threshold, thus obtaining a first state label. If the oxygen ratio value in the regional combustion feature vector is higher than the preset oxygen threshold, it determines the presence of excess cold air in the corresponding region, thus obtaining a second state label. This process includes acquiring the regional combustion feature vector and extracting the oxygen ratio value from it; denoising the oxygen ratio value to obtain a smoothed oxygen value; performing a difference calculation between the smoothed oxygen value and the preset oxygen threshold to obtain an oxygen deviation sequence; and determining the state of the oxygen deviation sequence, including determining the presence of unburned components in the corresponding region if the value in the oxygen deviation sequence is less than zero, thus obtaining a first state label; and determining the presence of excess cold air in the corresponding region if the value in the oxygen deviation sequence is greater than zero, thus obtaining a second state label.

[0193] In this embodiment, the determination module receives the regional combustion feature vectors output by the fusion module. Each regional combustion feature vector carries a region number, region coordinates, data generation time, and real-time oxygen ratio value.

[0194] The determination module reads the real-time oxygen ratio value from the regional combustion feature vector based on the pre-stored feature arrangement position, and writes the value into the oxygen data cache of the corresponding region according to the region number.

[0195] In addition, each oxygen data cache stores the oxygen ratio values ​​for five consecutive processing cycles in the same area according to the data generation time; after a new oxygen ratio value is written, the value with the earliest generation time is deleted, so that the cache always retains the data of the current processing cycle and the previous four processing cycles.

[0196] In this embodiment, the noise reduction process is divided into median filtering and smoothing.

[0197] When performing median filtering, the determination module arranges the five oxygen ratio values ​​in the buffer in ascending order and takes the third oxygen ratio value as the median filtering result for the current processing cycle. This process can eliminate the influence of a single upward or downward jump on the oxygen determination result.

[0198] Furthermore, the determination module saves the median filtering results of three consecutive processing cycles in the same region and performs a weighted calculation on the three median filtering results.

[0199] The median filtering result of the current processing cycle is multiplied by 3, the median filtering result of the previous processing cycle is multiplied by 2, and the median filtering result of the previous two processing cycles is multiplied by 1. The products of the three terms are added together, and the sum is divided by 6. The resulting value is used as the smoothed oxygen value for the current processing cycle. By increasing the calculation weight of the data in the current processing cycle, the smoothed oxygen value can reflect the current oxygen status, while limiting the numerical fluctuation within a single processing cycle.

[0200] In addition, during the first four processing cycles after the judgment module starts, if the buffer has not yet formed five oxygen ratio values, the earliest obtained effective oxygen ratio value is used to fill the missing data positions.

[0201] In the case where three median filtering results have not yet been formed, the earliest formed median filtering result is used to fill the missing data positions, so that the judgment module can perform noise reduction processing from the time the first effective oxygen ratio value is obtained.

[0202] Furthermore, the preset oxygen threshold is determined separately for each furnace region.

[0203] Specifically, after the furnace completes combustion debugging, it runs continuously for 30 minutes under the target combustion conditions. The conditions for confirming the target combustion conditions are: the absolute value of the difference between the furnace load and the set load does not exceed 1% of the set load; the absolute value of the difference between the total fuel flow rate and the set total fuel flow rate does not exceed 1% of the set total fuel flow rate; the absolute value of the difference between the total air flow rate and the set total air flow rate does not exceed 1% of the set total air flow rate; the temperature change in each zone within 1 minute does not exceed 2 degrees Celsius; and the furnace control device does not output any abnormal combustion signals.

[0204] Under the above operating conditions, the judgment module obtains the smoothed oxygen value of each region at a frequency of once every 2 seconds, and obtains 900 smoothed oxygen values ​​for each region within 30 minutes.

[0205] Specifically, for any region, the 900 smoothed oxygen values ​​of that region are added together sequentially, and the sum is divided by 900. The resulting arithmetic mean is determined as the preset oxygen threshold for that region. The preset oxygen threshold is stored in correspondence with the region number and is called according to the region number during the determination.

[0206] Furthermore, when the type of fuel used in the furnace changes, the burner arrangement changes, or the set load changes, the original preset oxygen threshold is discontinued, and a 30-minute calibration process is re-executed; the newly obtained arithmetic mean replaces the original preset oxygen threshold to ensure that the oxygen threshold is consistent with the current furnace operating conditions.

[0207] Furthermore, after oxygen denoising processing, the determination module reads the smoothed oxygen value and the corresponding preset oxygen threshold for the current region. The difference between the smoothed oxygen value and the preset oxygen threshold is used as the oxygen deviation value for the current region. The determination module sequentially saves the oxygen deviation values ​​for each processing cycle according to the data generation time, thus forming an oxygen deviation sequence for the corresponding region. Each value in the oxygen deviation sequence is associated with a region number, region coordinates, and a smoothed oxygen value.

[0208] Specifically, when the current oxygen deviation value in the oxygen deviation sequence is less than 0, it indicates that the smoothed oxygen value of the current region is lower than the preset oxygen threshold of the region; the determination module determines the region as an unburned component accumulation state and generates a first state marker; the state value of the first state marker is set to 1, and the state value of the second state marker is set to 0.

[0209] When the current oxygen deviation value in the oxygen deviation sequence is greater than 0, it means that the smoothed oxygen value of the current region is higher than the preset oxygen threshold of the region.

[0210] The determination module identifies the area as having an excess of cold air and generates a second state marker. The state value of the second state marker is set to 1, while the state value of the first state marker is set to 0.

[0211] When the oxygen deviation value is equal to 0, it means that the smoothed oxygen value is consistent with the preset oxygen threshold.

[0212] The determination module sets the state values ​​of both the first state marker and the second state marker to 0, and does not determine the area as a state of unburned component accumulation or a state of excess cold air.

[0213] Example 5: The temperature module, based on the first state marker, the second state marker, and the regional combustion feature vector, uses a multilayer perceptron to calculate the temperature, obtaining the true temperature values ​​of each region inside the furnace. Based on the region coordinates and true temperature values ​​of each region, it constructs a three-dimensional true temperature distribution matrix inside the furnace. This includes acquiring the first state marker, the second state marker, and the regional combustion feature vector; fusing the first state marker, the second state marker, and the regional combustion feature vector to obtain a feature fusion vector; using the feature fusion vector to calculate the initial temperature values ​​of each region inside the furnace using a multilayer perceptron; determining whether the initial temperature values ​​are greater than a preset temperature threshold; if the initial temperature values ​​are greater than the preset temperature threshold, performing nonlinear smoothing on the initial temperature values ​​to obtain the true temperature values ​​of each region; acquiring the region coordinates of each region; performing spatial location matching based on the region coordinates and true temperature values ​​to obtain a spatial mapping matrix; and performing three-dimensional reconstruction based on the spatial mapping matrix to construct the three-dimensional true temperature distribution matrix inside the furnace.

[0214] In this embodiment, the temperature module receives the first state flag and the second state flag output by the determination module, as well as the regional combustion feature vector output by the fusion module.

[0215] Specifically, the first state marker, the second state marker, and the regional combustion feature vector are all accompanied by a region number and a data generation time. The temperature module first compares the region number and then compares the data generation time, and only uses data with the same region number and belonging to the same processing cycle as the temperature estimation input for the same region.

[0216] Furthermore, both the first state marker and the second state marker are represented in numerical form.

[0217] When unburned components accumulate in the corresponding region, the first state marker is set to 1; otherwise, it is set to 0. When there is an excess of cold air in the corresponding region, the second state marker is set to 1; otherwise, it is set to 0. The temperature module arranges the data in the order of first state marker, second state marker, oxygen ratio, oxygen distribution gradient, dimensionality reduction features, and classification decision results to form a feature fusion vector corresponding to the region.

[0218] To eliminate the influence of different feature dimensions on the temperature estimation results, the temperature module calculates the arithmetic mean and standard deviation of each continuous numerical feature in the training samples during the multilayer perceptron training phase.

[0219] Specifically, when calculating the arithmetic mean, all training values ​​at the same feature location are summed and then divided by the number of training samples; when calculating the standard deviation, the corresponding arithmetic mean is subtracted from each training value, the difference is squared, all squared results are summed and then divided by the number of training samples, and the square root of the result is taken.

[0220] In addition, during real-time calculation, the temperature module subtracts the training arithmetic mean of the feature from the current feature value in the feature fusion vector, and then divides the difference by the corresponding standard deviation to obtain the standardized feature value; the first state label, the second state label, and the classification result number retain their original values; when the training standard deviation of a certain continuous numerical feature is equal to 0, it indicates that the feature has not changed in the training samples, and the feature is not input into the multilayer perceptron.

[0221] Specifically, the multilayer perceptron includes an input layer, three hidden layers, and an output layer; the first hidden layer has 32 nodes, the second hidden layer has 16 nodes, the third hidden layer has 8 nodes, and the output layer has 1 node; the number of nodes in the input layer is consistent with the dimension of the feature fusion vector after standardization.

[0222] Specifically, for any node in the first hidden layer, the values ​​of each node in the input layer are multiplied by the connection weights corresponding to that hidden layer node, and then all the products are added together, along with the bias value of that node.

[0223] Additionally, if the result is less than 0, the node outputs 0; if the result is greater than or equal to 0, the node outputs the result. The second and third hidden layers process the data output by the previous hidden layer in the same way.

[0224] Furthermore, the output layer multiplies the eight output values ​​of the third hidden layer by their corresponding connection weights, adds the eight products together, and adds them to the output layer bias value. The result is used as the initial temperature value of the corresponding furnace region. The combustion feature vector of each region is processed separately to obtain the initial temperature value of each region inside the furnace.

[0225] Furthermore, the multilayer sensor was trained using historical furnace operation data before being put into operation.

[0226] Specifically, each training sample includes a first state marker, a second state marker, a regional combustion feature vector, and a measured temperature value for the same region and the same processing cycle. The measured temperature value is collected by a temperature detection device that has completed metrological calibration, and the detection position of the temperature detection device is consistent with the coordinates of the corresponding region.

[0227] This embodiment uses 10,000 sets of samples, of which 7,000 sets are used to adjust model parameters, 1,500 sets are used to determine training stopping conditions, and 1,500 sets are used to verify temperature estimation results.

[0228] Specifically, during the training process, the multilayer perceptron first generates a calculated temperature value based on each group of training samples, then subtracts the corresponding measured temperature value from the calculated temperature value, and squares the difference. The squared results corresponding to 7000 groups of samples are added together and then divided by 7000. The result is used as the error value of the current training round.

[0229] Furthermore, the connection weights and bias values ​​are adjusted in reverse based on the error values, with the parameter adjustment step size set to 0.001.

[0230] Specifically, after each training round, the validation error is calculated using 1500 validation samples. Training is terminated when the validation error for 10 consecutive training rounds is not less than the previously obtained minimum validation error, and the connection weights and bias values ​​corresponding to the minimum validation error are saved.

[0231] In addition, during model validation, the absolute value of the difference between the estimated and measured temperature values ​​for 1500 test samples was calculated. All absolute values ​​of the difference were summed and divided by 1500. If the average absolute value of the difference did not exceed 5 degrees Celsius, the corresponding model parameters were used as the operating parameters for the temperature module.

[0232] Furthermore, after obtaining the initial temperature value, the temperature module retrieves the corresponding preset temperature threshold based on the area number.

[0233] Specifically, the preset temperature thresholds are determined separately for each area of ​​the furnace. When the furnace has completed combustion adjustment and is in the target operating condition, the measured temperature values ​​of each area are collected continuously for 30 minutes at a sampling interval of 2 seconds, and 900 measured temperature values ​​are obtained for each area.

[0234] Specifically, for any given region, 900 measured temperature values ​​are summed, and the sum is divided by 900 to obtain the average measured temperature for that region. The average measured temperature is then subtracted from each of the 900 measured temperature values, and the resulting difference is squared. The sum of the 900 squared values ​​is then divided by 900, and the square root of the result is taken to obtain the standard deviation of the measured temperature for that region. The average measured temperature is then added to three times the standard deviation of the measured temperature, and the resulting value is determined as the preset temperature threshold for that region.

[0235] As a preferred embodiment, during the threshold calibration period, the absolute value of the difference between the furnace load and the set load shall not exceed 1% of the set load, the absolute value of the difference between the total fuel flow and the set total fuel flow shall not exceed 1% of the set total fuel flow, the absolute value of the difference between the total air flow and the set total air flow shall not exceed 1% of the set total air flow, and the measured temperature change in each zone within 1 minute shall not exceed 2 degrees Celsius.

[0236] If the fuel type, burner arrangement, or set load changes, the preset temperature threshold calibration will be re-executed.

[0237] Specifically, when the initial temperature value is less than or equal to the preset temperature threshold of the corresponding area, the temperature module directly uses the initial temperature value as the actual temperature value.

[0238] When the initial temperature value is greater than the preset temperature threshold, the temperature module performs non-linear smoothing on the portion exceeding the threshold.

[0239] In addition, during nonlinear smoothing, the initial temperature value is first subtracted from the preset temperature threshold to obtain the over-threshold temperature difference; then the over-threshold temperature difference is divided by the smoothing scale, 1 is added to the resulting quotient, and the natural logarithm is performed on the value after adding 1; the result of the natural logarithm is multiplied by the smoothing scale, and then added to the preset temperature threshold, and the resulting value is used as the true temperature value.

[0240] This process ensures that the initial temperature value remains continuous at a preset temperature threshold and compresses the temperature increment corresponding to the temperature difference exceeding the threshold.

[0241] Furthermore, the smoothing scale was determined using 1500 sets of validation samples after the multilayer perceptron was trained.

[0242] Specifically, 20 degrees Celsius, 40 degrees Celsius, 60 degrees Celsius, 80 degrees Celsius and 100 degrees Celsius were used as candidate smoothing scales in turn to perform nonlinear smoothing on data in the verification sample whose initial temperature value was greater than the preset temperature threshold.

[0243] As a preferred embodiment, for each candidate smoothing scale, the smoothed temperature value is subtracted from the corresponding measured temperature value, and the resulting difference is squared. All squared results are summed, divided by the number of samples involved in the calculation, and then the square root of the result is taken. The candidate smoothing scale with the smallest calculated result is used as the smoothing scale during runtime. In this embodiment, 60 degrees Celsius is used.

[0244] Furthermore, after the actual temperature values ​​of each region are generated, the temperature module reads the corresponding region coordinates.

[0245] Specifically, each set of region coordinates includes the coordinates in the width direction, depth direction, and height direction of the furnace. The temperature module writes the region number, the region coordinates in the three directions, and the actual temperature value into the same data record, and arranges all data records according to the region number to obtain a spatial mapping matrix. Each row in the spatial mapping matrix corresponds to one furnace region, which is used to represent the correspondence between the actual temperature value and the spatial position of the furnace.

[0246] Before 3D reconstruction, a regular voxel mesh is established according to the internal space dimensions of the furnace.

[0247] In this embodiment, the furnace is divided into 40 positions along its width, 20 positions along its depth, and 10 positions along its height, forming a 40x20x10 matrix unit. Each matrix unit corresponds to a fixed spatial position inside the furnace.

[0248] When the spatial position of a matrix cell coincides with the coordinates of a region, the temperature module directly writes the actual temperature value of that region into the corresponding matrix cell.

[0249] When a matrix element is located between the coordinates of adjacent regions, trilinear interpolation is used to determine the temperature value of that matrix element.

[0250] In addition, when performing trilinear interpolation, eight center points of regions surrounding the matrix unit to be processed are selected.

[0251] Specifically, the distances between the matrix cell to be processed and the center planes of the regions on both sides of the width direction are calculated separately, and the ratio of the distance on the opposite side to the sum of the distances on both sides is used as the weight in the width direction, so that the sum of the two weights in the width direction is equal to 1; the weights in the depth direction and the height direction are calculated in the same way.

[0252] Furthermore, the interpolation weights of the center point of any region are obtained by multiplying the three weights corresponding to the width, depth and height directions.

[0253] Specifically, after obtaining the interpolation weights of the eight regional center points in the same way, the actual temperature value of each regional center point is multiplied by the corresponding interpolation weight, and the eight products are added together. The result is then written into the matrix cell to be processed.

[0254] For a matrix unit located at the furnace boundary where it is impossible to obtain the center points of the eight regions, the spatial straight-line distance between the matrix unit and the center points of each adjacent region is calculated, and the center point of the region with the smallest spatial straight-line distance is selected. The actual temperature value corresponding to the center point of the region is then written into the matrix unit.

[0255] Furthermore, the temperature module sequentially assigns temperature values ​​to all matrix units to obtain a three-dimensional real temperature distribution matrix inside the furnace; each matrix unit in the three-dimensional real temperature distribution matrix corresponds to a determined spatial location and real temperature value within the furnace.

[0256] Example 6: The deviation module acquires a pre-defined target 3D temperature distribution matrix. By comparing the actual 3D temperature distribution matrix with the target 3D temperature distribution matrix, it determines the temperature estimation deviation matrix. This includes mapping the target 3D temperature distribution matrix to the actual 3D temperature distribution matrix to obtain the target mesh temperature matrix and the actual mesh temperature matrix; comparing the target mesh temperature matrix with the actual mesh temperature matrix to calculate the temperature difference matrix; if the absolute value of the deviation in the temperature difference matrix is ​​greater than a preset threshold, a deviation correction weight is generated based on the temperature difference matrix; the deviation correction weight is used to weight the temperature difference matrix to obtain a weighted temperature difference matrix, and the temperature estimation deviation matrix is ​​determined based on the weighted temperature difference matrix.

[0257] In this embodiment, the deviation module pre-stores the target three-dimensional temperature distribution matrix and receives the three-dimensional real temperature distribution matrix output by the temperature module.

[0258] Specifically, the target three-dimensional temperature distribution matrix records the target temperature value corresponding to each spatial position when the furnace is under the target combustion condition; the three-dimensional real temperature distribution matrix records the real temperature value corresponding to each spatial position within the current processing cycle; both matrices include the furnace coordinate origin, coordinate axis direction, spatial coordinates, and data generation time.

[0259] Before matrix comparison, the deviation module performs a unified mesh mapping on the target three-dimensional temperature distribution matrix and the three-dimensional real temperature distribution matrix.

[0260] In this embodiment, the furnace is divided into 40 grid positions along the width direction, 20 grid positions along the depth direction, and 10 grid positions along the height direction, forming a total of 8,000 grid units. The spatial position of each grid unit is determined by the width coordinates, depth coordinates, and height coordinates of its geometric center.

[0261] In addition, for any temperature data in the target three-dimensional temperature distribution matrix, the deviation module reads the spatial coordinates corresponding to that temperature data.

[0262] Specifically, when the spatial coordinates coincide with the center coordinates of a certain grid cell, the temperature data is directly written into the corresponding grid cell; when the spatial coordinates are located between the centers of adjacent grid cells, trilinear interpolation is used to determine the target temperature value of the corresponding grid cell.

[0263] In a preferred embodiment, when performing trilinear interpolation, eight temperature data points surrounding the grid cell to be processed are first determined.

[0264] Specifically, along the width direction, the distance between the center of the grid to be processed and the plane containing the temperature data points on both sides is calculated, and the ratio of the distance on the opposite side to the sum of the distances on both sides is used as the corresponding directional weight, so that the sum of the two weights in the width direction is equal to 1; the weights in the depth and height directions are determined in the same way.

[0265] Furthermore, the interpolation weight for any temperature data point is obtained by multiplying the three weights corresponding to that temperature data point in the width, depth, and height directions.

[0266] The process involves multiplying the temperature values ​​of eight temperature data points by their corresponding interpolation weights, summing the eight products, and then writing the result into the grid cell to be processed. After assigning values ​​to all grid cells, the target grid temperature matrix is ​​formed.

[0267] In this embodiment, the three-dimensional real temperature distribution matrix is ​​mapped according to the same grid division and interpolation method to form a real grid temperature matrix.

[0268] Specifically, the target grid temperature matrix and the actual grid temperature matrix are both 40 x 20 x 10, and grid cells located at the same matrix position correspond to the same spatial position inside the furnace.

[0269] In a preferred embodiment, the deviation module compares the target grid temperature matrix with the actual grid temperature matrix item by item according to the grid cell position.

[0270] For any given grid cell, the temperature difference is calculated by subtracting the target grid temperature at the same location from the actual grid temperature. A temperature difference greater than 0 indicates that the actual temperature exceeds the target temperature; a temperature difference less than 0 indicates that the actual temperature is lower than the target temperature; and a temperature difference equal to 0 indicates that they are the same. All temperature differences are arranged according to their positions within the same grid cell, forming a temperature difference matrix.

[0271] Furthermore, the preset threshold is determined separately for each grid cell.

[0272] Specifically, the threshold calibration is performed when the furnace is under the target combustion condition. The calibration time is 30 minutes, the sampling interval is 2 seconds, and 900 real temperature values ​​are obtained for each grid cell. The target combustion condition meets the following conditions: the absolute value of the difference between the furnace load and the set load does not exceed 1% of the set load, the absolute value of the difference between the total fuel flow rate and the set total fuel flow rate does not exceed 1% of the set total fuel flow rate, the absolute value of the difference between the total air flow rate and the set total air flow rate does not exceed 1% of the set total air flow rate, and the furnace control device does not output a combustion abnormality signal.

[0273] For any given grid cell, the target temperature value of that grid cell is subtracted from 900 actual temperature values, and the absolute value of the difference is taken to obtain 900 historical deviation absolute values. The 900 historical deviation absolute values ​​are then added together and divided by 900 to obtain the arithmetic mean of the historical deviation absolute values.

[0274] Subsequently, the arithmetic mean is subtracted from the absolute value of each historical deviation, and the resulting difference is squared. The 900 squared results are summed, divided by 900, and then the square root is taken to obtain the standard deviation of the absolute value of the historical deviation for that grid cell. The arithmetic mean is added to three times the standard deviation, and the resulting value is determined as the preset threshold for that grid cell. If the fuel type, burner arrangement, set load, or target three-dimensional temperature distribution matrix changes, the threshold calibration is re-performed.

[0275] Furthermore, the deviation module takes the absolute value of each temperature difference in the temperature difference matrix and compares the obtained absolute value of deviation with the preset threshold corresponding to the same grid cell.

[0276] Specifically, when the absolute value of the deviation is less than or equal to a preset threshold, the deviation correction weight of the grid cell is set to 1.

[0277] In addition, when the absolute value of the deviation is greater than the preset threshold, the absolute value of the deviation is divided by the preset threshold, and the resulting quotient is used as the initial deviation correction weight for that grid cell.

[0278] When the initial deviation correction weight is greater than 3, the deviation correction weight is set to 3; when the initial deviation correction weight is less than or equal to 3, the initial deviation correction weight is used directly.

[0279] In addition, when the preset threshold is equal to 0, the deviation module calls the allowable measurement error determined by the temperature detection device after metrological calibration, and uses the allowable measurement error as the calculation threshold of the grid cell; thus avoiding division by 0 during the calculation process.

[0280] Furthermore, after the deviation correction weight is determined, the deviation module multiplies the temperature difference of each grid cell by the corresponding deviation correction weight, and the result is used as the weighted temperature difference of that grid cell.

[0281] When the temperature difference is positive, the weighted temperature difference remains positive; when the temperature difference is negative, the weighted temperature difference remains negative; when the temperature difference is 0, the weighted temperature difference is 0.

[0282] In this embodiment, all weighted temperature difference values ​​are arranged according to a unified grid position to form a weighted temperature difference matrix.

[0283] Specifically, for grid cells whose absolute deviation value does not exceed the preset threshold, the deviation correction weight is 1, and the weighted temperature difference is consistent with the original temperature difference; for grid cells whose absolute deviation value exceeds the preset threshold, the deviation correction weight is greater than 1, which is used to improve the deviation characterization value of the grid cell in subsequent risk assessment.

[0284] Furthermore, the deviation module determines the weighted temperature difference matrix as the temperature estimation deviation matrix for the current processing cycle, and associates each matrix cell with width coordinates, depth coordinates, height coordinates, target temperature value, actual temperature value, temperature difference, deviation correction weight, and weighted temperature difference.

[0285] Example 7: The risk module determines that a local overheating risk exists in the corresponding area if the deviation value in the temperature estimation deviation matrix is ​​greater than a preset deviation threshold. Obtaining the risk area coordinates involves acquiring the deviation value of the temperature estimation deviation matrix. If the deviation value is greater than the preset deviation threshold, a local overheating risk is determined, and an initial set of overheating nodes is generated. A local overheating connected component is constructed based on the initial set of overheating nodes, and boundary contour nodes are selected from this component. Envelope processing is performed on the boundary contour nodes to generate a risk area polygon, and the geometric centroid of the risk area polygon is extracted as the risk area coordinates for the local overheating risk.

[0286] In this embodiment, the risk module receives the temperature estimation deviation matrix output by the deviation module.

[0287] Specifically, each matrix unit in the temperature estimation deviation matrix corresponds to a node in the three-dimensional mesh of the furnace, and records the width, depth, and height coordinates of the node, as well as the weighted temperature deviation value. The risk module reads the data of each node according to the arrangement order of the three-dimensional mesh to ensure that the deviation value corresponds to the spatial position of the furnace.

[0288] Furthermore, the preset deviation threshold is determined separately for each furnace grid node.

[0289] Specifically, after the furnace combustion adjustment is completed, it runs continuously for 30 minutes under the target combustion conditions. Temperature deviation matrices are acquired at 2-second intervals, with 900 historical deviation values ​​obtained for each grid node. During the calibration period, the absolute value of the difference between the furnace load and the set load does not exceed 1% of the set load, the absolute value of the difference between the total fuel flow rate and the set total fuel flow rate does not exceed 1% of the set total fuel flow rate, the absolute value of the difference between the total air flow rate and the set total air flow rate does not exceed 1% of the set total air flow rate, and the furnace control device does not generate an over-temperature alarm.

[0290] For any given grid node, the 900 historical deviation values ​​are summed, and the sum is divided by 900 to obtain the average historical deviation of that node.

[0291] Next, subtract the average historical deviation from each historical deviation value, square the resulting difference, add the 900 squared results together, divide by 900, and then take the square root of the result to obtain the standard deviation of the historical deviation. Add the average historical deviation to three times the standard deviation of the historical deviation, and the resulting value is determined as the preset deviation threshold for that grid node. If the fuel type, burner arrangement, set load, or target three-dimensional temperature distribution matrix changes, the above calibration process is repeated.

[0292] In this embodiment, during real-time diagnosis, the risk module compares the deviation value of each grid node with the preset deviation threshold corresponding to that node.

[0293] When the deviation value is greater than the preset deviation threshold, the node is marked as an over-temperature node, and its node number, three-dimensional coordinates, deviation value and data generation time are recorded; when the deviation value is less than or equal to the preset deviation threshold, no over-temperature node mark is generated; after the comparison of all grid nodes is completed, the initial over-temperature node set is formed by all over-temperature nodes in the current processing cycle.

[0294] Furthermore, after the initial set of overheated nodes is formed, the risk module constructs a local overheated connected domain based on the face adjacency relationship in the three-dimensional mesh.

[0295] For any overheated node, check its six adjacent nodes in the positive width direction, negative width direction, positive depth direction, negative depth direction, positive height direction, and negative height direction respectively; if the adjacent nodes also belong to the initial overheated node set, the two nodes are assigned to the same local overheated connected domain.

[0296] In the specific processing, the node with the smallest node number is selected from the unclassified overheated nodes as the starting node, and it is written into the current connected component and the queue to be checked. Each time, one node is taken out from the queue to be checked, and its six neighboring nodes are checked. If the neighboring node belongs to the initial overheated node set and has not yet been written into the current connected component, the neighboring node is written into both the current connected component and the queue to be checked. When the queue to be checked is empty, the construction of a local overheated connected component is completed. Continue to select a new starting node from the unclassified overheated nodes until all nodes in the initial overheated node set have been classified.

[0297] In addition, to eliminate sampling jumps in individual grid nodes, the risk module verifies the number of nodes contained in each local overheated connected domain; the node number threshold is determined using historical diagnostic data.

[0298] Specifically, 1000 sets of data confirmed by furnace maintenance records and on-site temperature measurement results were selected, of which 500 sets were actual local over-temperature data and 500 sets were temperature fluctuation data. The number of continuous over-temperature nodes formed by each set of data was counted. The connected domains formed by temperature fluctuation data contained 1 to 3 nodes, and the connected domains formed by actual local over-temperature data contained more than 4 nodes. Therefore, the node number threshold was set to 4. Connected domains with fewer than 4 nodes were deleted, and connected domains with 4 or more nodes entered the boundary contour extraction process.

[0299] For the retained local overheated connected domains, the risk module determines whether each node belongs to the boundary contour node.

[0300] Specifically, if a certain overheating node does not have a valid grid node in any of the six adjacent directions, or if the adjacent node in that direction does not belong to the current local overheating connected domain, then the overheating node is determined as a boundary profile node; only when a node's adjacent nodes in all six directions belong to the current local overheating connected domain is it determined as an internal node; after judging all nodes according to the above conditions, a set of boundary profile nodes is formed.

[0301] Furthermore, the risk region polygon is generated according to the spatial extension direction of the local overheated connected domain.

[0302] Specifically, the risk module reads the maximum and minimum coordinate values ​​of the boundary contour nodes in the width, depth, and height directions, and subtracts the minimum coordinate value from the maximum coordinate value to obtain the coordinate span in the three directions. The direction with the smallest coordinate span value is determined as the projection direction, and the boundary contour nodes are projected onto the plane formed by the other two coordinate directions. By using the direction with the smallest coordinate span value for projection, the contour range of the local overheated connected domain on the extended plane can be preserved.

[0303] Furthermore, after projection is completed, the risk module uses convex hull processing to form a polygon of the risk area.

[0304] Specifically, the node with the smallest horizontal coordinate among the projected nodes is selected as the starting node; when there are more than two nodes with the same horizontal coordinate, the node with the smallest vertical coordinate is selected; based on the starting node, the remaining nodes are arranged in ascending order of their rotation angle relative to the positive horizontal direction.

[0305] In this embodiment, the projection nodes are written into the contour sequence in the order of arrangement.

[0306] Specifically, for each new node written, the orientation of the last three consecutive nodes in the contour sequence is checked; the difference in lateral coordinates between the first and second nodes is multiplied by the difference in lateral coordinates between the second and third nodes, and then the product of the difference in lateral coordinates between the first and second nodes is subtracted; if the calculation result is less than or equal to 0, the middle node of the three nodes is deleted, and the last three nodes after deletion are recalculated; if the calculation result is greater than 0, the middle node is retained. After all projected nodes have been processed, the last nodes of the contour sequence are connected to the starting node to form a closed risk region polygon.

[0307] Furthermore, the risk module calculates the geometric centroid based on the vertex coordinates of the risk region polygon.

[0308] Specifically, according to the arrangement order of the polygon vertices, each vertex and the next vertex form an edge, and the last vertex and the first vertex form a closed edge. For each edge, the horizontal coordinate of the previous vertex is multiplied by the vertical coordinate of the next vertex, and then the product of the horizontal coordinate of the next vertex and the vertical coordinate of the previous vertex is subtracted to obtain the area contribution value of the edge. All area contribution values ​​are added together and divided by 2 to obtain the directed area of ​​the risk region polygon.

[0309] When calculating the centroid's horizontal coordinates, the horizontal coordinates of the two vertices of each edge are added together, and then multiplied by the area contribution value of that edge. All products are then summed and divided by six times the directed area. The centroid's vertical coordinates are processed in the same way: the vertical coordinates of the two vertices of each edge are added together, multiplied by the area contribution value of that edge, and all products are then summed and divided by six times the directed area.

[0310] As a preferred embodiment, the coordinates of the geometric centroid in the projection direction are determined by the average coordinates of all nodes in the current local hyperthermic connected domain in that direction.

[0311] Specifically, the coordinates of all nodes in the projection direction are added together and then divided by the number of nodes contained in the connected domain to obtain the projection direction coordinates; the centroid horizontal coordinates, centroid vertical coordinates and projection direction coordinates are restored according to the directional relationship of the furnace three-dimensional coordinate system, and the resulting three coordinate values ​​form the risk area coordinates of the local overheat risk.

[0312] In addition, when the directed area of ​​the risk region polygon is equal to 0, it indicates that the boundary contour nodes after projection are located on the same straight line.

[0313] Specifically, the risk module adds up the width, depth, and height coordinates of all nodes within the local overheated connected domain, divides each coordinate by the number of nodes, and uses the average of the three coordinates as the risk area coordinates.

[0314] Example 8: The adjustment module determines the fuel concentration adjustment and airflow direction adjustment for each region inside the furnace based on the risk area coordinates and temperature deviation matrix. This includes performing spatial grid partitioning on the furnace interior regions based on the risk area coordinates to obtain a three-dimensional grid topology. The three-dimensional grid topology is then mapped to the temperature deviation matrix using heat flux to obtain a combustion thermal distribution matrix. The combustion thermal distribution matrix is ​​used to extract combustion state features of independent spatial nodes. It then determines whether the combustion state features are greater than a preset feature threshold. If the combustion state features are greater than the preset feature threshold, a deviation compensation matrix is ​​generated. Based on the deviation compensation matrix, the concentration gradient vector and airflow deflection angle are calculated, and the fuel concentration adjustment and airflow direction adjustment for each region inside the furnace are determined based on these parameters.

[0315] In this embodiment, the adjustment module receives the risk area coordinates output by the risk module and the temperature calculation deviation matrix output by the deviation module.

[0316] Specifically, the risk area coordinates include the furnace width, depth, and height coordinates; each matrix cell in the temperature estimation deviation matrix includes the grid position, temperature deviation value, and data generation time; the adjustment module only processes data whose data generation time belongs to the same diagnostic cycle.

[0317] Furthermore, the adjustment module divides the internal space of the furnace according to the grid size used in the temperature calculation deviation matrix.

[0318] In this embodiment, 40 grid positions are set along the width direction of the furnace, 20 grid positions are set along the depth direction, and 10 grid positions are set along the height direction, forming a total of 8000 spatial nodes. Each spatial node records the node number, three-dimensional coordinates, and the adjacent node numbers in the forward width direction, reverse width direction, forward depth direction, reverse depth direction, forward height direction, and reverse height direction, thereby forming a three-dimensional grid topology.

[0319] Specifically, when determining risk nodes based on the coordinates of the risk area, the spatial distance between the risk area coordinates and each spatial node is calculated.

[0320] Specifically, the coordinate differences in the width direction, depth direction, and height direction are calculated separately. The three coordinate differences are squared, and the square root of the sum of the three squared results is taken to obtain the spatial distance between the risk area coordinates and the corresponding spatial node. The spatial node with the smallest spatial distance value is determined as the risk center node.

[0321] In this embodiment, when the temperature calculation deviation matrix is ​​mapped to the three-dimensional mesh topology, the adjustment module searches for the corresponding node according to the spatial coordinates.

[0322] When the coordinates of the matrix unit are consistent with the coordinates of the spatial node, the temperature deviation value of the matrix unit is directly written into the corresponding spatial node; when the coordinates are inconsistent, the spatial node with the smallest spatial distance is selected and written; the position occupied by the furnace solid structure is marked as an invalid node, and invalid nodes do not participate in the heat flux calculation.

[0323] As a preferred embodiment, the heat flux mapping is determined according to the temperature deviation changes between adjacent nodes.

[0324] For any valid node, the temperature deviation between that node and its adjacent node in the positive width direction is read. The temperature deviation of the current node is subtracted from the temperature deviation of the adjacent node in the positive width direction. The difference is divided by the distance between the centers of the two nodes and then multiplied by the equivalent heat transfer coefficient to obtain the positive width heat flux component. The heat flux components in the other five adjacent directions are calculated in the same way. When a node is located at the furnace boundary and there is no adjacent node in the corresponding direction, the heat flux component in that direction is set to 0.

[0325] Furthermore, the equivalent heat transfer coefficient is calibrated during the furnace commissioning stage.

[0326] Specifically, 100 sets of stable furnace operation data were selected. Each set of data included the measured temperature difference between adjacent nodes, the distance between node centers, and the heat flux detection value. The heat flux detection value was multiplied by the distance between node centers and then divided by the absolute value of the measured temperature difference to obtain a calibration coefficient. Data with a measured temperature difference of 0 were not included in the calibration. All valid calibration coefficients were added together and then divided by the number of valid data sets. The result was used as the equivalent heat transfer coefficient.

[0327] In addition, for each spatial node, the heat flux components in the positive and negative width directions are added together to obtain the net heat flux in the width direction; the heat flux components in the positive and negative depth directions are added together to obtain the net heat flux in the depth direction; the heat flux components in the positive and negative height directions are added together to obtain the net heat flux in the height direction; the three net heat fluxes are squared respectively, and the square root of the sum of the three squared results is taken to obtain the heat flux intensity of the node.

[0328] Furthermore, the adjustment module arranges the temperature deviation values ​​of each node, the net heat flux in three directions, and the heat flux intensity according to the three-dimensional grid position to obtain the combustion thermal distribution matrix; each set of data in the combustion thermal distribution matrix corresponds to a unique spatial node.

[0329] Specifically, the combustion state characteristics of independent spatial nodes are determined by node temperature deviation, heat flux intensity, and local temperature deviation dispersion. The calculation range of local temperature deviation dispersion includes the current node and its effective adjacent nodes. First, the temperature deviation values ​​within the calculation range are added together and then divided by the number of effective nodes to obtain the average local temperature deviation. The average value is then subtracted from the temperature deviation value of each node, the difference is squared, all squared results are added together and divided by the number of effective nodes. The square root of the result is then taken to obtain the local temperature deviation dispersion.

[0330] In addition, to unify the numerical scale of the three data points, the adjustment module calls the temperature deviation benchmark value, heat flux benchmark value, and discrete quantity benchmark value respectively.

[0331] Specifically, the temperature deviation characteristic is obtained by dividing the absolute value of the current node temperature deviation by the temperature deviation reference value; the heat flux characteristic is obtained by dividing the heat flux intensity by the heat flux reference value; the discrete characteristic is obtained by dividing the local temperature deviation discrete value by the discrete value reference value; the temperature deviation characteristic is multiplied by 0.5, the heat flux characteristic is multiplied by 0.3, the discrete characteristic is multiplied by 0.2, and the products of the three terms are added together. The resulting value is used as the combustion state characteristic of the node.

[0332] The above three benchmark values ​​were determined under the target combustion conditions.

[0333] Specifically, the furnace runs continuously and stably for 30 minutes, and the adjustment module acquires 900 sets of node data at a 2-second acquisition interval. For each node, the 900 absolute values ​​of temperature deviation, 900 heat flux intensities, and 900 discrete values ​​of local temperature deviation are added together and then divided by 900. The three arithmetic mean values ​​are used as the temperature deviation reference value, heat flux reference value, and discrete value reference value, respectively. When any reference value is equal to 0, it is replaced by the minimum resolution value of the corresponding detection device.

[0334] As a preferred embodiment, the preset feature thresholds are calibrated separately for each spatial node.

[0335] Specifically, using 900 combustion state characteristics obtained in the same calibration period, the 900 combustion state characteristics are first added together and then divided by 900 to obtain the average value of the combustion state characteristics; then, each combustion state characteristic is subtracted from the average value, the difference is squared, the 900 squared results are added together and divided by 900, and the square root of the result is taken to obtain the standard deviation of the combustion state characteristics; the average value of the combustion state characteristics is added to three times the standard deviation of the combustion state characteristics, and the resulting value is determined as the preset characteristic threshold of the corresponding node; after changes in fuel type, set load, burner arrangement or target temperature distribution, the benchmark value and the preset characteristic threshold are recalibrated.

[0336] In addition, during real-time processing, when the combustion state characteristic of a node is less than or equal to the corresponding preset characteristic threshold, the deviation compensation value of that node is set to 0.

[0337] Specifically, when the node combustion state feature is greater than the preset feature threshold, the combustion state feature is subtracted from the preset feature threshold to obtain the feature excess amount; then the feature excess amount is divided by the preset feature threshold to obtain the compensation ratio.

[0338] Specifically, when the compensation ratio is greater than 2, the value is taken as 2; when the compensation ratio is less than or equal to 2, the calculation result is retained. The temperature deviation value of the node is multiplied by the sum of the compensation ratio and 1, and the result is determined as the node deviation compensation value. The deviation compensation values ​​of all nodes are arranged according to the position of the three-dimensional mesh to form a deviation compensation matrix. The positive and negative directions of the deviation compensation value are consistent with the temperature deviation value.

[0339] Furthermore, the concentration gradient vector is calculated based on the deviation compensation values ​​of adjacent nodes in the deviation compensation matrix.

[0340] Specifically, for nodes located inside the width direction, the width direction concentration gradient component is obtained by subtracting the deviation compensation value of the adjacent node in the opposite direction from the deviation compensation value of the adjacent node in the positive direction of the width, and then dividing by the distance between the centers of the two adjacent nodes; the depth direction concentration gradient component and the height direction concentration gradient component are determined by the same process.

[0341] In this embodiment, for a node located at the grid boundary, the deviation compensation value of the current node is subtracted from the deviation compensation value of the only adjacent node, and then divided by the distance between the centers of the two nodes to obtain the concentration gradient component in the corresponding direction; the concentration gradient components in the three directions together constitute the concentration gradient vector.

[0342] Furthermore, the adjustment module squares each of the three concentration gradient components, adds the three squared results together, and takes the square root to obtain the concentration gradient intensity.

[0343] When the concentration gradient intensity is equal to 0, no airflow direction adjustment is generated; when the concentration gradient intensity is greater than 0, the opposite direction of the concentration gradient vector is determined as the airflow adjustment direction.

[0344] Furthermore, the airflow deflection angle includes both horizontal and vertical deflection angles.

[0345] Specifically, when calculating the horizontal deflection angle, the negative of the concentration gradient component in the depth direction is used as the vertical input, and the negative of the concentration gradient component in the width direction is used as the horizontal input, and a four-quadrant arctangent operation is performed; when calculating the vertical deflection angle, the concentration gradient components in the width and depth directions are squared first, the two squared results are added together and the square root is taken; then the negative of the concentration gradient component in the height direction is used as the vertical input, and the square root result is used as the horizontal input, and an arctangent operation is performed.

[0346] When the horizontal deflection angle exceeds the horizontal rotation range of the airflow regulating device, the horizontal rotation limit value is taken; when the vertical deflection angle exceeds the vertical rotation range, the vertical rotation limit value is taken.

[0347] As a preferred embodiment, the fuel concentration adjustment amount is determined based on the node deviation compensation value.

[0348] Specifically, for the same local overheated connected domain, the largest positive deviation compensation value is read; the positive deviation compensation value of the current node is divided by this maximum value, and the resulting ratio is multiplied by the maximum concentration reduction allowed by the fuel regulator. The result is used as the fuel concentration reduction of the current node; when the deviation compensation value of the current node is less than or equal to 0, the fuel concentration adjustment is set to 0; when there is no positive deviation compensation value in the connected domain, no fuel concentration adjustment is performed.

[0349] Furthermore, the airflow direction adjustment amount consists of the horizontal deflection angle, the vertical deflection angle, and the airflow adjustment amplitude.

[0350] Specifically, the airflow adjustment range is determined by dividing the concentration gradient intensity of the current node by the maximum concentration gradient intensity within the same local overheated connected domain, and then multiplying the resulting ratio by the maximum allowable opening adjustment of the airflow regulating device; when the maximum concentration gradient intensity within the connected domain is equal to 0, the airflow adjustment range is set to 0.

[0351] When the same furnace area contains multiple spatial nodes, the fuel concentration adjustment of all nodes in the area is added together and then divided by the number of nodes to obtain the fuel concentration adjustment of the area. The horizontal deflection angle, vertical deflection angle and airflow adjustment amplitude are calculated as arithmetic averages in the same way, and the results constitute the airflow direction adjustment of the area.

[0352] Example 9: The instruction module determines independent adjustment instructions for each region inside the furnace based on the fuel concentration adjustment and airflow direction adjustment. This includes extracting the feeder speed increment and secondary damper opening based on the fuel concentration adjustment and airflow direction adjustment, and obtaining the actuator response delay. If the actuator response delay is greater than a preset delay threshold, timing alignment processing is performed on the feeder speed increment and secondary damper opening to obtain a dynamic compensation sequence. Spatial interference weights are calculated based on the dynamic compensation sequence to construct a decoupling control matrix. The decoupling control matrix is ​​used to perform instruction conversion processing on the dynamic compensation sequence to determine independent adjustment instructions for each region inside the furnace.

[0353] In this embodiment, the instruction module receives the fuel concentration adjustment amount and airflow direction adjustment amount output by the adjustment module, and retrieves the feedback data of the pulverizer, secondary damper and actuator of the corresponding area according to the area number.

[0354] Specifically, the airflow direction adjustment includes the airflow adjustment amplitude, horizontal deflection angle, and vertical deflection angle; when the area number, adjustment amount, and diagnostic cycle number of the actuator feedback data are consistent, the instruction generation process begins.

[0355] Among them, the fuel concentration adjustment is converted into the feeder speed increment.

[0356] Specifically, before the pulverizer is put into operation, the pulverizer speed is set sequentially to 10%, 20%, 30% up to 100% of the rated speed, and each speed is maintained for 60 seconds. The mass of pulverized coal output during that period is recorded, and the mass of pulverized coal is divided by 60 to obtain the mass of pulverized coal output per second at the corresponding speed. Ten sets of speed and pulverized coal output are saved in ascending order of speed to form the pulverizer calibration data.

[0357] Furthermore, the instruction module reads the current rotation speed of the pulverizer and obtains the current coal powder output based on the pulverizer calibration data.

[0358] The coal powder output adjustment is obtained by multiplying the current coal powder output by the fuel concentration adjustment; the coal powder output adjustment is then added to the current coal powder output to obtain the target coal powder output.

[0359] In addition, when the target pulverized coal output is between two adjacent calibrated outputs, the smaller calibrated output is subtracted from the target pulverized coal output and then divided by the difference between the two adjacent calibrated outputs to obtain the interpolation ratio; the difference between the two adjacent calibrated speeds is multiplied by the interpolation ratio and added to the smaller calibrated speed to obtain the target feeder speed; the target feeder speed is subtracted from the current speed, and the difference is used as the feeder speed increment.

[0360] When the target pulverized coal output is less than the minimum output in the calibration data, the target feeder speed is the minimum allowable speed; when the target pulverized coal output is greater than the maximum output in the calibration data, the target feeder speed is the maximum allowable speed.

[0361] When the absolute value of the feeder speed increment exceeds the allowable change in a single control cycle, the speed increment for this cycle is determined according to the allowable change, and the remaining increment is written into the next control cycle.

[0362] Furthermore, the airflow direction adjustment is converted into the opening adjustment value of each secondary damper.

[0363] Specifically, the instruction module performs a cosine calculation on the horizontal deflection angle, multiplies the result by the airflow adjustment amplitude, and obtains the width direction airflow adjustment component; performs a sine calculation on the horizontal deflection angle, multiplies the result by the airflow adjustment amplitude, and obtains the depth direction airflow adjustment component; performs a sine calculation on the vertical deflection angle, multiplies the result by the airflow adjustment amplitude, and obtains the height direction airflow adjustment component.

[0364] When the width direction airflow adjustment component is positive, it is allocated to the width positive direction secondary air damper; when the width direction airflow adjustment component is negative, its absolute value is taken and allocated to the width negative direction secondary air damper; the depth direction and height direction airflow adjustment components are allocated to the corresponding direction secondary air dampers according to the same rule.

[0365] In this embodiment, the correspondence between the opening degree and the air supply volume of each secondary damper is established in advance.

[0366] During the specific calibration, the dampers are set in the order of 0%, 10%, 20% up to 100%, and each opening is maintained for 30 seconds. The air volume collected within 30 seconds is added together and divided by the number of samplings to obtain the average air volume corresponding to that opening.

[0367] Furthermore, the target air volume is determined based on the air volume adjustment component and the current air volume.

[0368] Specifically, when the target air volume is between two adjacent rated air volumes, the smaller rated air volume is subtracted from the target air volume, and then divided by the difference between the two adjacent rated air volumes to obtain the opening interpolation ratio; the difference between the two adjacent rated openings is multiplied by the opening interpolation ratio, and then added to the smaller rated opening to obtain the target secondary damper opening; the target secondary damper opening is subtracted from the current feedback opening, and the difference is used as the secondary damper opening adjustment value.

[0369] Specifically, when the target secondary damper opening is less than 0%, it is set to 0%; when it is greater than 100%, it is set to 100%. When the opening adjustment value exceeds the allowable action amount for a single control cycle of the damper, the adjustment for the current cycle is performed according to the allowable action amount, and the unexecuted part is carried over to the next control cycle.

[0370] Furthermore, the actuator response delay is determined based on the instruction sending time and the feedback arrival time.

[0371] Specifically, when the instruction module sends a powder mill speed instruction or a secondary damper opening instruction, it records the instruction sending time and continuously reads the actuator feedback value; when the feedback change reaches 90% of the target change, it records the feedback arrival time; the time obtained by subtracting the instruction sending time from the feedback arrival time is the actuator response delay for this time.

[0372] In addition, the instruction module saves the response delay of the last 20 valid actions of each actuator. The current response delay of the actuator is obtained by adding the 20 response delays together and dividing by 20. If the actuator has not completed 90% of the target change or a fault signal occurs during the action, the action will not be written into the response delay record.

[0373] In this embodiment, the preset delay threshold is determined during the actuator debugging phase.

[0374] Specifically, 50 step actions are performed for each powder feeder and each secondary damper, with the amplitude of each action set to 5% of the actuator's full range; the response delays of the 50 actions are recorded, summed, and divided by 50 to obtain the average delay; the average delay is subtracted from each response delay, the difference is squared, the sum of the 50 squared results is divided by 50, and the square root of the result is taken to obtain the standard deviation of the delay; the average delay is added to three times the standard deviation of the delay, and the resulting value is determined as the preset delay threshold for the actuator; the preset delay threshold is re-determined after the actuator is repaired, the drive mechanism is replaced, or the control parameters are changed.

[0375] When the current response delay of any powder feeder or secondary damper in the same area is greater than the corresponding preset delay threshold, the instruction module performs timing alignment processing on the powder feeder speed increment and the secondary damper opening adjustment value.

[0376] As a preferred embodiment, during timing alignment, the response delays of all actuators participating in this adjustment within the same region are read, and the response delay with the largest value is determined as the reference delay; the current control time plus the reference delay is used as the unified effective time; for any actuator, the current response delay of the actuator is subtracted from the unified effective time, and the resulting time is used as the instruction sending time of the actuator, so that the adjustment action of the pulverizer and the secondary air damper reaches the furnace region at the unified effective time.

[0377] Furthermore, the adjustment values ​​for the feeder speed increment and the secondary air damper opening are divided into four levels, with the adjustment amounts for each level being 25%, 50%, 75%, and 100% of the total adjustment amount, respectively. The effective time interval between two adjacent levels is set to 0.5 seconds.

[0378] For each adjustment level, the current response delay of the actuator is subtracted from the effective time of the target level to obtain the command sending time of the corresponding level; the speed commands of the powder feeder and the opening commands of the secondary damper at each level are arranged according to the command sending time, thus forming a dynamic compensation sequence.

[0379] In addition, when the current response delay of all actuators is less than or equal to the corresponding preset delay threshold, a dynamic compensation sequence is still formed according to the four adjustment levels of 25%, 50%, 75% and 100%, and each level of instruction is sent sequentially at 0.5-second intervals.

[0380] Furthermore, spatial interference weights are used to characterize the impact of the action of a powder feeder or secondary damper in one region on the temperature deviation in other regions.

[0381] Specifically, the spatial interference weight was obtained through single actuator tests during the furnace commissioning phase; only one actuator was adjusted at a time, while the remaining actuators remained in their original state; the feeder speed adjustment was set to 5% of the rated speed, and the secondary damper opening adjustment was set to 5% of the full opening.

[0382] In this embodiment, before the actuator operates, the temperature prediction deviation of each area is collected for 20 consecutive seconds. The deviation values ​​of the same area are added together and divided by the number of samplings to obtain the average deviation before the operation. After the actuator operates, the temperature prediction deviation of each area is collected for another 20 consecutive seconds, and the average deviation after the operation is calculated in the same way.

[0383] Furthermore, by subtracting the average deviation before the action from the average deviation after the action, we obtain the change in regional deviation caused by the actuator's action.

[0384] The temperature influence coefficient of the actuator on the corresponding area is obtained by dividing the change in regional deviation by the actual increase in the speed of the powder feeder or the increase in the opening of the secondary damper.

[0385] Specifically, for the same actuator, the absolute values ​​of its temperature influence coefficients on all regions are taken and summed to obtain the total influence coefficient. The temperature influence coefficient of the actuator on any region is divided by the total influence coefficient, and the resulting value is used as the spatial interference weight of the actuator on that region. The spatial interference weight retains the positive and negative signs. A positive number indicates that the actuator action increases the temperature deviation of the corresponding region, and a negative number indicates that the actuator action decreases the temperature deviation of the corresponding region.

[0386] As a preferred embodiment, the spatial interference weights are arranged according to the affected area numbers, and the actuator data is arranged according to the powder feeder and secondary damper numbers to form a spatial interference matrix. Each row in the spatial interference matrix corresponds to one affected area, and each column corresponds to one actuator adjustment input.

[0387] Furthermore, the instruction module constructs a decoupling control matrix based on the spatial interference matrix.

[0388] In the specific construction, an identity matrix of the same order is concatenated to the right of the spatial interference matrix, and column-by-column elimination is performed starting from the first column. For the current column, the non-zero element with the largest absolute value in the current row and all rows below it is selected as the pivot. The row containing the pivot is swapped with the current row, and then all elements in the row containing the pivot are divided by the pivot to convert the pivot to 1. Subsequently, the row containing the pivot is multiplied by the elements in the current column of the other rows, and then subtracted from the corresponding rows to convert all elements in the current column except the pivot to 0.

[0389] Furthermore, after processing all columns, the left-side spatial interference matrix is ​​converted into an identity matrix, and the right-side matrix is ​​the decoupling control matrix.

[0390] If a column does not contain any non-zero principal elements, 0.01 is added to each diagonal position of the spatial interference matrix, and column-by-column elimination is performed again to obtain a definite decoupling control matrix.

[0391] Furthermore, when executing command conversion through the decoupled control matrix, the feeder speed increment and secondary damper opening adjustment value in the dynamic compensation sequence are first arranged according to the actuator number.

[0392] Specifically, for any row in the decoupling control matrix, each element of that row is multiplied by the adjustment value at the same position in the dynamic compensation sequence, and then all the products are added together. The result is used as the decoupling adjustment amount for the corresponding actuator.

[0393] Specifically, for the powder feeder, the decoupled speed increment is added to the current powder feeder speed to obtain the independent target speed; for the secondary damper, the decoupled opening adjustment value is added to the current damper opening to obtain the independent target opening; when the independent target speed exceeds the allowable speed range of the powder feeder, the corresponding speed boundary value is taken; when the independent target opening is less than 0%, 0% is taken, and when it is greater than 100%, 100% is taken.

[0394] Furthermore, the instruction module writes the area number, actuator number, instruction sending time, target effective time, independent target speed or independent target opening degree, single-cycle change limit and diagnostic cycle number into the control data frame to form independent adjustment instructions for each area inside the furnace.

[0395] The independent adjustment command is sent to the corresponding powder feeder controller and secondary damper actuator according to the sending time determined by the dynamic compensation sequence.

[0396] Example 10: The display module maps the 3D real temperature distribution matrix, first state markers, second state markers, risk area coordinates, and independent adjustment commands to a preset 3D furnace model, generating a 3D visualization diagnostic result of the furnace combustion state. This includes acquiring the 3D real temperature distribution matrix and the preset 3D furnace model; performing spatial interpolation and color assignment on the 3D real temperature distribution matrix based on the grid node coordinates of the 3D furnace model to obtain a basic temperature rendering layer; performing local highlighting on the basic temperature rendering layer based on the first and second state markers to obtain a state enhancement rendering layer; generating a multi-dimensional feature fusion layer in the state enhancement rendering layer based on the risk area coordinates and independent adjustment commands; and performing layer overlay sequence processing on the multi-dimensional feature fusion layer and the 3D furnace model through a rendering viewpoint matrix to generate a 3D visualization diagnostic result of the furnace combustion state.

[0397] In this embodiment, the display module receives a three-dimensional real temperature distribution matrix, a first state marker, a second state marker, risk area coordinates, and independent adjustment instructions; each data item is accompanied by a diagnostic cycle number, a region number, and a data generation time; the display module first checks the diagnostic cycle number and only maps data belonging to the same diagnostic cycle to the same furnace three-dimensional model; data whose data generation time exceeds the end time of the current diagnostic cycle is transferred to the next diagnostic cycle for processing.

[0398] Furthermore, the pre-set three-dimensional model of the furnace is established according to the actual dimensions of the furnace inner wall, burner outlet, observation port, and effective combustion space.

[0399] Specifically, the model's coordinate origin and three coordinate directions are consistent with the furnace's three-dimensional coordinate system used in the three-dimensional real temperature distribution matrix; the furnace's three-dimensional model is composed of triangular patches, and each model mesh node stores the width direction coordinates, depth direction coordinates, height direction coordinates, and the furnace region number to which it belongs.

[0400] Furthermore, the display module reads the coordinates of the grid nodes of the furnace 3D model one by one, and determines the 8 matrix nodes surrounding the grid node in the 3D real temperature distribution matrix.

[0401] Specifically, along the width direction, the distances from the model mesh nodes to the planes containing the matrix nodes on both sides are calculated. The distance from the model mesh nodes to the opposite side plane is then divided by the total distance between the two side planes to obtain the width direction weights corresponding to the matrix nodes on both sides. The sum of the two width direction weights is 1. The depth direction weights and height direction weights are determined using the same method.

[0402] As a preferred embodiment, the spatial interpolation weight of each matrix node is obtained by multiplying the width direction weight, depth direction weight, and height direction weight corresponding to that matrix node.

[0403] The display module multiplies the actual temperature values ​​of the eight matrix nodes by their respective spatial interpolation weights, and then adds the eight products together to obtain the interpolated temperature value of the current model grid node.

[0404] Specifically, when the coordinates of the model mesh nodes are consistent with the coordinates of the matrix nodes in the three-dimensional real temperature distribution matrix, the real temperature value of the corresponding matrix node is directly read.

[0405] For model mesh nodes located at the furnace boundary where it is impossible to obtain 8 matrix nodes, calculate the spatial distance between the model mesh node and the surrounding effective matrix nodes, and use the temperature value of the matrix node with the smallest spatial distance to complete the assignment.

[0406] Furthermore, the temperature and color ranges are determined during the display module debugging phase.

[0407] Specifically, the minimum allowable temperature specified in the furnace operation procedure is set as the lower limit of color mapping, and the maximum allowable temperature is set as the upper limit of color mapping; temperatures below the lower limit of color mapping are treated as the lower limit, and temperatures above the upper limit of color mapping are treated as the upper limit.

[0408] Furthermore, when assigning color values, the interpolated temperature value of the model mesh node is first subtracted from the lower limit of the color mapping, and then the difference is divided by the difference between the upper and lower limits of the color mapping to obtain the temperature position ratio.

[0409] The display module is set with five consecutive color intervals in the order of dark blue, cyan, green, yellow, orange and red, and the color interval of the current temperature is determined according to the temperature position ratio.

[0410] Specifically, within a defined color range, the interpolation ratio is obtained by subtracting the initial ratio of the color range from the current temperature position ratio, and then dividing by the ratio width of the color range.

[0411] Specifically, the red, green, and blue channel values ​​of the next reference color are subtracted from the corresponding channel values ​​of the previous reference color, respectively. The three differences are then multiplied by the interval interpolation ratio and added to the corresponding channel values ​​of the previous reference color to obtain the three color channel values ​​of the current model mesh node. After all model mesh nodes have completed spatial interpolation and color assignment, a basic temperature rendering layer is formed.

[0412] Furthermore, the display module searches for all grid nodes belonging to the corresponding region in the three-dimensional model of the furnace based on the region number associated with the first and second state markers.

[0413] When the state value of the first state marker is 1, the area is designated as a display area for the accumulation of unburned components; when the state value of the second state marker is 1, the area is designated as a display area for excess cold air.

[0414] For the unburned component accumulation display area, multiply each channel value of the base temperature color by 0.6, multiply the corresponding channel value of the purple indicator color by 0.4, and then add the two products of the same channel to obtain the local highlight color.

[0415] For areas with excess cold air, the base temperature color is mixed with the cyan-white indicator color using the same calculation method; when the status value of the status marker is 0, the base temperature color of the corresponding area is not changed.

[0416] Furthermore, after the local highlighting is completed, the display module checks the adjacent nodes of each grid node within the status area.

[0417] Specifically, when at least one adjacent node of the current grid node does not belong to the same state region, the current grid node is determined as the state boundary node; the display module generates a closed outline according to the spatial arrangement order of the state boundary nodes, wherein the unburned component accumulation area adopts a purple outline and the cold air excess area adopts a bluish-white outline; the local highlight color and the state boundary outline together form a state enhancement rendering layer.

[0418] In addition, during the risk area coordinate mapping, the display module calculates the width, depth, and height coordinate differences between the risk area coordinates and each model grid node, respectively. The three coordinate differences are squared, and the sum of the three squared results is taken as the square root to obtain the spatial distance between the risk area coordinates and each model grid node. The grid node with the smallest spatial distance value is identified as the risk identification node.

[0419] Furthermore, the display module generates risk identifiers centered around risk identifier nodes.

[0420] Specifically, the risk marker is a red semi-transparent sphere with a radius equal to twice the center distance between adjacent mesh nodes in the furnace 3D model and a transparency of 0.35. When the risk module outputs the risk area boundary at the same time, it connects the boundary nodes according to the arrangement order to form a closed risk profile and sets the risk area coordinates inside the profile.

[0421] In this embodiment, the independent adjustment command is mapped to the corresponding furnace area according to the area number.

[0422] Specifically, when the feeder speed increment is greater than 0, an upward fuel adjustment indicator is displayed in the center of the area; when the feeder speed increment is less than 0, a downward fuel adjustment indicator is displayed. The target speed and speed adjustment amount are simultaneously displayed next to the indicator.

[0423] Furthermore, the secondary damper adjustment command is represented by a three-dimensional arrow.

[0424] Specifically, the starting point of the arrow is set at the geometric center of the corresponding furnace area; based on the horizontal and vertical deflection angles in the independent adjustment command, the components of the arrow in the width, depth and height directions are calculated respectively, and the three directional components are combined to form the direction of the arrow.

[0425] Furthermore, the arrow length is determined based on the adjustment amount of the secondary air damper opening.

[0426] Specifically, the length ratio is obtained by dividing the absolute value of the opening adjustment by the maximum allowable opening adjustment of the secondary damper; then the length ratio is multiplied by the preset maximum arrow length to obtain the current arrow length; when the opening adjustment is 0, the airflow adjustment arrow is not displayed; the risk indicator, fuel adjustment indicator, airflow adjustment arrow and corresponding numerical information are superimposed on the status enhancement rendering layer to form a multi-dimensional feature fusion layer.

[0427] As a preferred embodiment, the rendering viewpoint matrix is ​​established based on the coordinates of the viewpoint, the coordinates of the target being viewed, and the vertical direction of the screen.

[0428] Specifically, the display module subtracts the corresponding coordinates of the observation point from the three coordinates of the observed target to obtain the observation direction components; it squares each of the three observation direction components, sums the squares, and takes the square root to obtain the observation direction length; then it divides each observation direction component by the observation direction length to obtain the unit observation direction.

[0429] Specifically, the horizontal direction of the image is obtained by cross-product of the unit observation direction and the preset vertical direction of the image; the horizontal direction of the image is then cross-producted with the unit observation direction to obtain the corrected vertical direction of the image.

[0430] Next, the display module writes the horizontal direction of the screen, the vertical direction of the corrected screen, the unit viewing direction, and the coordinates of the viewing point into a 4x4 matrix to form a rendering viewpoint matrix; after being transformed by this matrix, each mesh node in the furnace 3D model obtains its display position and display depth relative to the current viewing point.

[0431] Furthermore, the layer overlay is executed in the following order: furnace 3D model, basic temperature rendering layer, status enhancement content, risk label content, and independent adjustment instruction content.

[0432] Specifically, the 3D model of the furnace is used to display the outline of the furnace structure; the base temperature rendering layer covers the effective combustion space; the status highlight color and status boundary outline cover the base temperature color; the risk indicator covers the status enhancement content; and the fuel adjustment indicator, airflow adjustment arrow and adjustment value are located at the top of the display order.

[0433] Layers with transparency are processed using channel blending.

[0434] For any display pixel, the channel values ​​of the upper color are multiplied by the upper transparency, the corresponding channel value of the lower color is multiplied by 1 and the upper transparency is subtracted, and the two results of the same channel are added together to obtain the color value after layer overlay.

[0435] Furthermore, when there are multiple display objects in the same viewing direction, the display module compares the display depth of each display object.

[0436] Specifically, non-transparent objects retain the object closest to the observation point; transparent objects are color-blended in order of distance from the observation point to maintain the spatial occlusion relationship between temperature colors, status markers, risk areas, and adjustment instructions.

[0437] Furthermore, after the layers are overlaid, the display module outputs a 3D image that includes the three-dimensional structure of the furnace, the actual temperature of the area, the first state marker, the second state marker, the location of the risk area, and independent adjustment instructions.

[0438] When any furnace region is selected, the display module finds the corresponding grid node based on the selected location and displays the actual temperature value, status marker, risk area coordinates, target speed of the pulverizer, and target opening of the secondary air damper for that region, thereby forming a three-dimensional visualization diagnostic result of the furnace combustion status.

[0439] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A three-dimensional visualization diagnostic system for furnace combustion status that integrates flame image features and oxygen content data, characterized in that, include: The feature module acquires the flame video stream inside the furnace, performs temporal registration on continuous video frames, extracts regional brightness change features, and obtains the flame brightness feature matrix. The fusion module acquires real-time oxygen ratio data for each region inside the furnace and the region coordinates of each region in the preset three-dimensional coordinate system of the furnace. Based on the real-time oxygen ratio data and the flame brightness feature matrix, it obtains the region combustion feature vector. The determination module determines that if the oxygen ratio in the regional combustion feature vector is lower than the preset oxygen threshold, the corresponding region is determined to have an accumulation of unburned components, and a first state label is obtained; if the oxygen ratio in the regional combustion feature vector is higher than the preset oxygen threshold, the corresponding region is determined to have an excess of cold air, and a second state label is obtained. The temperature module uses a multilayer perceptron to calculate the temperature based on the first state marker, the second state marker, and the regional combustion feature vector, to obtain the true temperature values ​​of each region inside the furnace. Based on the regional coordinates and true temperature values ​​of each region, a three-dimensional true temperature distribution matrix inside the furnace is constructed. The deviation module obtains the pre-set target three-dimensional temperature distribution matrix and determines the temperature estimation deviation matrix by comparing the actual three-dimensional temperature distribution matrix with the target three-dimensional temperature distribution matrix. The risk module determines that there is a local overheating risk in the corresponding area if the deviation value in the temperature estimation deviation matrix is ​​greater than the preset deviation threshold, and obtains the coordinates of the risk area. The adjustment module calculates the deviation matrix based on the risk area coordinates and temperature, and determines the adjustment amount of fuel concentration and airflow direction in each area inside the furnace.

2. The three-dimensional visualization diagnostic system for furnace combustion status that integrates flame image features and oxygen data according to claim 1, characterized in that: The feature module acquires the flame video stream inside the furnace, performs temporal registration of consecutive video frames based on the furnace region mapping relationship, and extracts regional brightness change features through a convolutional neural network to obtain a flame brightness feature matrix, including: Acquire the flame video stream, perform coordinate transformation on the flame video stream according to the furnace area mapping relationship to obtain continuous video frames, and perform temporal registration on the continuous video frames to obtain temporally registered video frames. Regional brightness variation features of time-registered video frames are extracted using a convolutional neural network. If the regional brightness change characteristics are greater than a preset threshold, an initial matrix is ​​constructed based on the regional brightness change characteristics, and the initial matrix is ​​then dimensionality-reduced to obtain the flame brightness feature matrix.

3. The three-dimensional visualization diagnostic system for furnace combustion status that integrates flame image features and oxygen data according to claim 1, characterized in that: The fusion module acquires real-time oxygen ratio data for each region inside the furnace and the region coordinates of each region in a preset three-dimensional furnace coordinate system. Based on the real-time oxygen ratio data and the flame brightness feature matrix, it obtains the region combustion feature vector, including: Obtain real-time oxygen ratio data and region coordinates for each area inside the furnace. Construct a spatial mapping matrix based on the real-time oxygen ratio data and region coordinates, and extract the oxygen distribution gradient. The oxygen distribution gradient and the flame brightness feature matrix are tensor concatenated to obtain the brightness-oxygen fusion tensor. Principal component analysis was performed on the brightness-oxygen fusion tensor to obtain the dimension-reduced feature sequence; If the dimension of the reduced feature sequence is greater than the preset dimension threshold, the reduced feature sequence is input into the support vector machine for classification processing to obtain the region combustion feature vector.

4. The three-dimensional visualization diagnostic system for furnace combustion status that integrates flame image features and oxygen data according to claim 1, characterized in that: The determination module determines that if the oxygen ratio in the regional combustion feature vector is lower than a preset oxygen threshold, the corresponding region is in a state of unburned component accumulation, and a first state label is obtained; if the oxygen ratio in the regional combustion feature vector is higher than the preset oxygen threshold, the corresponding region is in a state of excess cold air, and a second state label is obtained. Obtain the regional combustion feature vector and extract the oxygen ratio value from the regional combustion feature vector; Denoising the oxygen ratio values ​​yields smoothed oxygen values. The oxygen deviation sequence is obtained by performing a difference calculation between the smoothed oxygen value and the preset oxygen threshold. Determine the numerical state of the oxygen deviation sequence, including if the value in the oxygen deviation sequence is less than zero, then determine that there is an accumulation of unburned components in the corresponding area and obtain the first state label; If the value in the oxygen deviation sequence is greater than zero, it is determined that there is an excess of cold air in the corresponding area, and a second state label is obtained.

5. The three-dimensional visualization diagnostic system for furnace combustion status that integrates flame image features and oxygen data according to claim 1, characterized in that: The temperature module, based on the first state marker, the second state marker, and the regional combustion feature vector, uses a multilayer perceptron to calculate the temperature, obtaining the true temperature values ​​of each region inside the furnace. Based on the region coordinates and true temperature values ​​of each region, it constructs a three-dimensional true temperature distribution matrix inside the furnace, including: First state marker, second state marker and regional combustion feature vector are obtained, and the first state marker, second state marker and regional combustion feature vector are fused to obtain feature fusion vector. Based on the feature fusion vector, multilayer perceptron is used to calculate the initial temperature value of each region inside the furnace. Determine whether the initial temperature value is greater than the preset temperature threshold. If the initial temperature value is greater than the preset temperature threshold, perform non-linear smoothing on the initial temperature value to obtain the true temperature value of each region. The coordinates of each region are obtained, and spatial location matching is performed based on the region coordinates and the actual temperature value to obtain a spatial mapping matrix. Based on the spatial mapping matrix, a three-dimensional reconstruction is performed to construct a three-dimensional real temperature distribution matrix inside the furnace.

6. The three-dimensional visualization diagnostic system for furnace combustion status that integrates flame image features and oxygen data according to claim 1, characterized in that: The deviation module acquires a pre-set target three-dimensional temperature distribution matrix and determines the temperature estimation deviation matrix by comparing the actual three-dimensional temperature distribution matrix with the target three-dimensional temperature distribution matrix. The target 3D temperature distribution matrix and the real 3D temperature distribution matrix are mapped to obtain the target mesh temperature matrix and the real mesh temperature matrix; Compare the temperature matrix of the target mesh with the temperature matrix of the real mesh, and calculate the temperature difference matrix. If the absolute value of the deviation of the temperature difference matrix is ​​greater than the preset threshold, then a deviation correction weight is generated based on the temperature difference matrix. The temperature difference matrix is ​​weighted using deviation correction weights to obtain a weighted temperature difference matrix, and the temperature estimation deviation matrix is ​​determined based on the weighted temperature difference matrix.

7. The three-dimensional visualization diagnostic system for furnace combustion status that integrates flame image features and oxygen data according to claim 1, characterized in that: The risk module determines that there is a local overheating risk in the corresponding area if the deviation value in the temperature calculation deviation matrix is ​​greater than a preset deviation threshold, and obtains the coordinates of the risk area, including: Obtain the deviation value of the temperature estimation deviation matrix. If the deviation value is greater than the preset deviation threshold, it is determined that there is a risk of local over-temperature and an initial set of over-temperature nodes is generated. Construct a local overheated connected domain based on the initial set of overheated nodes, and select boundary contour nodes from the local overheated connected domain; Envelope processing is performed on the boundary contour nodes to generate risk region polygons, and the geometric centroid of the risk region polygons is extracted as the risk region coordinates for local overtemperature risk.

8. The three-dimensional visualization diagnostic system for furnace combustion status that integrates flame image features and oxygen data according to claim 1, characterized in that: The adjustment module, based on the risk area coordinates and temperature deviation matrix, determines the fuel concentration adjustment and airflow direction adjustment for each area inside the furnace, including: Based on the coordinates of the risk area, the internal area of ​​the furnace is divided into spatial grids to obtain a three-dimensional grid topology. The three-dimensional grid topology is then mapped with the temperature calculation deviation matrix to obtain the combustion thermal distribution matrix. Combustion state characteristics of independent spatial nodes are extracted using a combustion thermodynamic distribution matrix; Determine whether the combustion state characteristics are greater than a preset characteristic threshold; If the combustion state characteristics are greater than the preset characteristic threshold, a deviation compensation matrix is ​​generated; The concentration gradient vector and airflow deflection angle are calculated based on the deviation compensation matrix, and the fuel concentration adjustment and airflow direction adjustment of each area inside the furnace are determined based on the concentration gradient vector and airflow deflection angle.

9. The three-dimensional visualization diagnostic system for furnace combustion status that integrates flame image features and oxygen data according to claim 1, characterized in that, It also includes an instruction module, which determines independent adjustment instructions for different areas inside the furnace based on the fuel concentration adjustment and airflow direction adjustment, specifically including: The feeder speed increment and secondary damper opening are extracted based on the fuel concentration adjustment and airflow direction adjustment, and the actuator response delay is obtained. If the actuator response delay is greater than the preset delay threshold, the timing alignment processing of the feeder speed increment and the secondary damper opening is performed to obtain a dynamic compensation sequence. The spatial interference weights are calculated based on the dynamic compensation sequence, and a decoupling control matrix is ​​constructed. By decoupling the control matrix, the dynamic compensation sequence is processed to convert the execution instructions and determine the independent adjustment instructions for each region inside the furnace.

10. A three-dimensional visualization diagnostic system for furnace combustion status that integrates flame image features and oxygen content data according to claim 1, characterized in that, It also includes a display module that maps the three-dimensional real temperature distribution matrix, first state marker, second state marker, risk area coordinates, and independent adjustment commands to a preset three-dimensional furnace model, generating a three-dimensional visualization diagnostic result of the furnace combustion state, specifically including: Obtain the three-dimensional real temperature distribution matrix and the preset three-dimensional furnace model. Based on the grid node coordinates of the three-dimensional furnace model, perform spatial interpolation and color assignment on the three-dimensional real temperature distribution matrix to obtain the basic temperature rendering layer. The base temperature rendering layer is locally highlighted based on the first and second state markers to obtain the state enhancement rendering layer. A multi-dimensional feature fusion layer is generated in the state enhancement rendering layer based on the risk area coordinates and independent adjustment instructions. By rendering the viewpoint matrix, a layer overlay sequence is performed on the multi-dimensional feature fusion layer and the three-dimensional furnace model to generate a three-dimensional visualization diagnostic result of the furnace combustion status.

Images