Full-automatic labeling combustion condition recognition method and system based on flame intelligent analysis

Abstract

The present application relates to a full-automatic labeling combustion condition recognition method and system based on flame intelligent analysis. The condition-specific characteristic parameters of the concerned condition are obtained based on flame image operation, and the condition-specific characteristic parameters of the concerned condition are matched with the condition characteristic parameter threshold library to obtain the corresponding condition state. The condition-specific characteristic parameter threshold of the corresponding condition is recorded in the condition characteristic parameter threshold library. New sample data are obtained based on the operation of newly collected flame images, and the model is incrementally injected regularly. Through iterative training, online fine-tuning of the judgment model is performed. When the condition recognition accuracy is lower than the preset threshold, full-volume retraining is triggered. The system used includes an image acquisition device, an image preprocessing module, an automatic labeling module, a machine learning module, a prediction module and a display module. The present application can be mainly used for condition recognition of waste incinerators, and can also be used in other suitable occasions.

Description

Technical Field

[0001] This invention relates to a fully automated method for identifying combustion conditions based on intelligent flame analysis, and also to a fully automated system for identifying combustion conditions based on intelligent flame analysis for implementing this method. Background Technology

[0002] When constructing recognition or prediction models using supervised learning, a large number of samples often need to be labeled. Although automatic labeling can be achieved using rules in some relatively simple cases, manual labeling remains a common method, especially in complex industrial monitoring scenarios. For example, in the operation of waste-to-energy plants, the combustion process in the incinerator is the core link in realizing the "resource recovery" and "volume reduction" of waste treatment, and its stability and efficiency directly determine the economic benefits of the project. Stable and complete combustion can directly increase the power generation per ton of waste, increasing grid-connected electricity revenue, which is the main source of income. However, the composition of municipal solid waste fed into the incinerator in the waste-to-energy industry is extremely complex. It includes high-calorific-value plastics and paper, as well as high-moisture, low-calorific-value kitchen waste, and is also mixed with chlorinated substances and trace amounts of heavy metals. This high heterogeneity leads to vastly different combustion characteristics, posing a fundamental challenge to establishing stable and uniform combustion conditions. It is necessary to adjust operating parameters such as feeding speed, grate movement frequency, and air volume ratio of each section according to actual conditions to maintain a good and stable combustion state in the furnace. However, due to the lack of effective means to automatically identify or judge the working conditions inside the furnace, the adjustment of operating parameters still relies on the operator's manual judgment of the flame image inside the furnace. The control accuracy is relatively low, the response time is relatively long, the stability is relatively poor, and the labor intensity of the operator is relatively high.

[0003] Under the current technological background, a feasible solution is to use existing furnace flame cameras (high-temperature digital cameras) to obtain furnace flame images, and to build an identification (judgment) model that can automatically identify furnace operating conditions through machine learning. Based on the automatic identification results, the boiler operating parameters can be manually adjusted or intelligently controlled. However, this requires labeling a large number of historical images for operating conditions. However, the existing manual labeling method is not only inefficient, making it difficult to achieve such a large number of labels in practice, but also has low accuracy, affecting the quality of the identification model. The main reasons for this situation include: (1) The data volume is extremely large, and the cost of manual labeling is unbearable: the incinerator flame video is a continuous real-time stream, generating dozens of images per second, and the daily data volume can reach millions. Manual labeling of each image requires a lot of manpower and time, which is extremely costly and cannot be scaled up; (2) The operating conditions change rapidly, and manual labeling is delayed and misses judgments: the manual judgment of flame operating conditions has a reaction delay and fatigue miss judgments, and cannot accurately capture transient operating conditions, resulting in inaccurate labeling results; (3) The manual standards are not uniform, which leads to model training Deviation: Different operators and engineers have different judgment criteria for "dust" and "material deviation". Manual annotation is highly subjective and the rules learned by the model are chaotic, which directly reduces the recognition accuracy. (4) The characteristics of garbage are constantly changing and require continuous sample iteration: The composition, calorific value and humidity of garbage change every day. The model must continuously iterate with new data. Manual annotation cannot keep up with the data update speed and the model will quickly fail. (5) It is difficult to meet the hard requirements of high precision and real-time performance of the model: This model requires an accuracy of ≥95% for working condition recognition and a response of ≤100ms. Only by training based on massive, accurate and standardized labeled samples can the indicators be achieved. Manual annotation cannot provide a sufficient number of high-quality samples. Summary of the Invention

[0004] The purpose of this invention is to provide a fully automatic labeled combustion condition identification method based on intelligent flame analysis and a fully automatic labeled combustion condition identification system based on intelligent flame analysis for implementing this method, so as to build a furnace condition judgment model based on automatic labeling of furnace conditions in flame images and realize automatic identification of the conditions of interest.

[0005] The technical solution of this invention is: a fully automatic labeled combustion condition identification method based on intelligent flame analysis. This method obtains condition-specific feature parameters (usually multiple parameters, or parameter groups, or, when appropriate, a single parameter) of the desired operating condition based on flame image calculations. It then matches these condition-specific feature parameters with a condition feature parameter threshold library (e.g., intelligent matching) to obtain the corresponding operating condition state (normal or abnormal). The condition feature parameter threshold library records the threshold values ​​(normal operating condition thresholds) of the corresponding condition-specific feature parameters (parameter groups). If the condition-specific feature parameters are within the corresponding threshold range, it is a normal operating condition; if the condition-specific feature parameters exceed the corresponding threshold range, it is an abnormal operating condition.

[0006] Preferably, the operating conditions of interest include long fire line, short fire line, and off-center fire conditions (hereinafter referred to as off-center fire conditions). The specific characteristic parameters of the long fire line, short fire line, and off-center fire conditions include the center of mass of the combustion zone (hereinafter referred to as the center of mass), the position of the fire line, and the flame intensity.

[0007] Preferably, the operating conditions of interest include thin material layer and thick material layer operating conditions (hereinafter referred to as material layer thickness operating conditions). The operating condition-specific characteristic parameters for thin material layer and thick material layer operating conditions include average flame brightness, flame intensity, and fire line area.

[0008] Preferably, the working conditions of interest include clumping conditions, and the working condition-specific feature parameters of clumping conditions include flame texture complexity and flame variance (pixel grayscale value variance).

[0009] Preferably, the operating conditions of interest include dust-generating conditions, and the specific characteristic parameters of dust-generating conditions include the number of corner points and the dust obstruction rate.

[0010] Preferably, when judging the dust-raising condition (abnormal condition), the valid data from the previous moment is automatically locked.

[0011] Preferably, sample data adapted for model training (construction) is constructed based on historical data. The model input data in the sample data is flame images (digital flame images), basic feature parameters (parameter sets) of flame images, or working condition-specific feature parameters (parameter sets) of flame images. The model output data is the working condition status of interest (usually the working condition status of multiple working conditions, such as the working condition status of four working conditions: dust, material deviation, material thickness, and clumping). The working condition judgment model is constructed based on the sample data.

[0012] Preferably, the working condition judgment model is a lightweight CNN+LSTM hybrid model trained using a CNN-LSTM hybrid neural network structure, or an XGBoost judgment model trained using an XGBoost model structure.

[0013] Preferably, the judgment model is automatically iterated. New sample data is obtained based on newly acquired flame images and is periodically incrementally injected into the model (sample database). Through iterative training, the judgment model is fine-tuned online. When the working condition recognition accuracy is lower than a preset threshold (such as 90%), full retraining is triggered, and the working condition feature parameter threshold library is updated (when appropriate).

[0014] The fully automated labeled combustion condition identification system based on intelligent flame analysis, employing any of the fully automated labeled combustion condition identification methods based on intelligent flame analysis disclosed in this invention, performs combustion condition identification (or judgment), including:

[0015] Image acquisition devices (e.g., furnace flame cameras) are used to continuously (periodically) acquire flame images;

[0016] The image preprocessing module is used to preprocess the flame image data from the image acquisition device, including noise reduction, enhancement and combustion area segmentation, to form preprocessed flame image data.

[0017] The automatic annotation module is used to extract feature parameters from preprocessed flame image data to obtain basic feature parameters and working condition-specific feature parameters of the flame image (normalization processing is performed when necessary). The working condition-specific feature parameters of the working condition of interest are matched with the working condition feature parameter threshold library (e.g., intelligent matching) to obtain the corresponding working condition status (normal or abnormal). The flame image (digital flame image), the basic feature parameters (parameter group) of the flame image, or the working condition-specific feature parameters (parameter group) of the flame image are used as the model input data, and the working condition status of the working condition of interest is used as the model output data to construct sample data, and the sample data can be stored in the sample database.

[0018] The machine learning module (or model building and maintenance module, or model training module) is used to build and maintain the working condition judgment model using sample data. After the judgment model is established, the model is iterated or rebuilt based on newly added sample data.

[0019] The prediction module is used to judge the working condition of the target working condition based on real-time model input data and output the judgment result.

[0020] The display module is used to display the judgment results.

[0021] Preferably, when the judgment result output by the prediction module includes one or more abnormal operating conditions, an alarm signal is generated and sent to the display along with the judgment result. When appropriate, the judgment result and other necessary data are also transmitted to the expert system, which proposes countermeasures (e.g., adjustment plans) and returns them to the prediction module. The prediction module then sends the countermeasures to the display device, where they are displayed together with the prediction result and the alarm signal.

[0022] The present invention has the following beneficial effects:

[0023] 1) Deep integration with waste incineration operations to achieve accurate operational condition identification. This invention employs image data annotation that goes beyond mere technical aspects. It uses actual waste incineration operational conditions as the core calibration basis, and deeply binds flame features with operational condition labels through cross-validation of flame feature parameters and DCS process parameters. This ensures that the machine learning model and operational condition identification always closely match actual on-site operations, completely solving the problem of low identification accuracy caused by the traditional technology's "heavy emphasis on images, light emphasis on operations".

[0024] 2) Significantly reduces manual labeling work, lowers costs, and facilitates project implementation. This invention abandons the traditional model of "manually labeling hundreds of thousands / millions of flame images one by one," achieving end-to-end fully automated labeling of flame conditions from data collection to label generation, without any manual intervention. This significantly reduces the workload and cost of manual labeling, lowers the threshold for project implementation, and allows flame-based intelligent analysis-based condition recognition technology to quickly adapt to the on-site needs of different waste-to-energy plants.

[0025] 3) Supports long-term automatic machine learning, adapting to dynamic changes in waste. This invention can capture changes in flame characteristics and combustion conditions in real time as factors such as waste composition, calorific value, and fermentation time change, continuously outputting standardized new data samples for machine learning; enabling machine learning models to complete long-term automatic iteration and optimization without human intervention, ensuring that the condition judgment model always adapts to the on-site combustion conditions, and solving the industry problem of model failure caused by changes in waste characteristics.

[0026] This invention is mainly used for identifying the operating conditions of waste incinerators, but it can also be used in other suitable applications. Attached Figure Description

[0027] Figure 1 This invention relates to a basic flowchart of the process from image acquisition to working condition recognition;

[0028] Figure 2 This invention relates to an iterative flowchart of automatic annotation, machine learning, and working condition identification.

[0029] Figure 3 This invention relates to a flowchart of image analysis, annotation, and model training. Detailed Implementation

[0030] See Figures 1 to 3The system acquires images of the flame inside the furnace (or flame images, or furnace images) using a high-temperature data camera (flame image detector, or furnace fire-viewing camera, or furnace flame camera) for flame imaging. Flame state information is obtained through image analysis, and the flame state signal is converted into data information such as flame size, position, color, brightness, and saturation (which can be called flame data or quantified flame characteristic parameters) and stored in a database. Multi-parameter cross-judgment is performed with reference to the parameters of the relevant DCS (Distributed Control System). For example, a longer fire line results in a higher air chamber pressure difference, a lower grate temperature, and a higher flue gas temperature above the grate. This allows for the determination of the corresponding operating conditions under the flame image (such as a long fire line, oxygen deficiency, uneven material distribution, thick or thin material, dust, etc., depending on actual needs). A correlation is established between flame data and operating condition information, enabling the labeling of operating condition information for image samples or corresponding flame parameter samples. Based on the automatically labeled samples, an operating condition judgment model is constructed. Automatic operating condition identification is performed based on the operating condition judgment model to obtain identification results (operating condition information), which are used to remind and guide operators to adjust operating parameters.

[0031] The main contents include:

[0032] 1. Provide high-quality, standardized, and large-scale training sample datasets for machine learning.

[0033] The flame images are transformed into standardized samples consisting of "quantified flame feature parameters + precise operating condition labels + timestamps + DCS linkage parameters", which solves problems such as unlabeled machine learning data, inconsistent labels, and data disconnect from business.

[0034] By automatically and continuously labeling, a large number of samples can be accumulated quickly to improve the efficiency of model training and its fit with the actual situation.

[0035] 2. Provide continuous data input for long-term automatic iteration of machine learning models.

[0036] The system captures flame characteristics and changes in operating conditions in real time as the characteristics of the waste change, continuously generating new standardized sample input models. This allows the model to complete long-term automatic iterations without human intervention, always adapting to on-site combustion conditions and ensuring long-term accuracy in operating condition identification.

[0037] This allows for a deep integration of flame characteristics with operational conditions, ensuring that the technology aligns with reality.

[0038] 3. Combining the business logic of waste incineration, and using actual combustion parameters as the calibration basis, flame characteristics are deeply bound to business condition labels such as "fire line length / short", "material layer thickness / thin", "clumping", and "ash emission", so that machine learning can establish a correlation between "flame characteristics and business conditions", which can significantly improve the actual accuracy of condition identification.

[0039] This operating condition identification method includes the following steps:

[0040] Step 1: Real-time acquisition of flame status and quantization extraction of features.

[0041] High-temperature digital cameras capture raw flame images (flame images), which are then preprocessed (denoising / enhancing / combustion region segmentation) to transform them into quantifiable and computable core flame feature parameters, which are then temporarily stored in a flame database. These parameters are categorized into two types:

[0042] Basic feature parameters: morphological (size, position, centroid coordinates, etc.), visual (color HSV value, brightness, saturation, etc.), and texture (variance, number of corner points, flame intensity, etc.).

[0043] Operating condition-specific characteristic parameters: dust emission (number of corner points, occlusion rate), fire line / material deviation (centroid movement trend, intensity distribution), material layer thickness (temperature color field, fire line area), clumping (variance, distribution of characteristic points in the drying section).

[0044] Step 2: Cross-calibrate multi-source data to ensure business accuracy.

[0045] By linking with the DCS system through an external data interface, the flame characteristic parameters and DCS process parameters (air chamber pressure difference, grate temperature, oxygen content, furnace temperature, etc.) are synchronized and cross-verified in real time. With flame parameters as the core and DCS parameters as the calibration basis, the quantitative threshold range of flame characteristic parameters under different operating conditions is determined to ensure that the labeling is consistent with the actual business conditions, realize the long-term automatic operation of the machine, and adapt to fully automatic labeling under different waste combustion conditions.

[0046] Step 3: Intelligent matching of working condition thresholds to automatically determine the working condition type.

[0047] The system has a pre-set threshold library of operating condition characteristic parameters derived from "industry business experience + cross-calibration of multi-source data," and intelligently matches real-time flame parameters with this threshold library; it supports simultaneous matching of complex operating conditions. For special operating conditions involving ash emission, when the number of corner points reaches a threshold, it automatically determines and locks the valid data from the previous moment to avoid abnormal data interfering with the accuracy of labeling. At the same time, it achieves intelligent matching of operating condition thresholds and automatically determines the type of operating condition. This enables the machine to operate automatically for a long time and adapt to fully automatic labeling under different waste combustion conditions.

[0048] Step 4: Fully automated injection of standard samples into the library to generate standardized machine learning samples.

[0049] After the operating condition is determined, the system automatically completes multi-dimensional standardized annotation and permanently stores it in the flame database, forming a complete data sample that can be directly used for machine learning. The annotation content includes: core labels (operating condition type, warning level), related parameters (flame characteristic parameters + DCS process parameters), and timestamp (collection / determination time); finally forming a complete data chain of "flame characteristics - operating condition labels - DCS parameters - timestamp".

[0050] Step 5: Use labeled data to automate machine learning and model iteration.

[0051] Standardized samples are input into the machine learning module in real time to complete two core actions: initial training: building a basic model for judging working conditions based on flame features; automatic iteration: completing incremental training and parameter optimization of the model with new labeled data, without human intervention throughout the process.

[0052] Construct a flame condition judgment model that is deeply integrated with business operations and adapts to dynamic changes:

[0053] Based on fully automated, labeled, standardized, and large-scale operational data, a working condition judgment model is formed through machine learning training, with original flame characteristics as its core. It possesses two core characteristics:

[0054] Deeply aligned with business needs: The model learns the correlation between "flame features and actual combustion conditions" rather than simply image features, solving the problems of traditional models being disconnected from business needs and having low recognition accuracy.

[0055] Long-term automatic self-optimization: Fully automatic annotation continuously generates new data, driving long-term iteration of the model without human intervention, adapting to dynamic changes in garbage composition and heat value, without the need for manual retraining.

[0056] Achieve real-time intelligent identification of flame conditions and output direct control / operation basis:

[0057] The operating condition judgment model reads the flame feature information collected by the high-temperature camera in real time, quickly completes fully automatic operating condition recognition, and outputs the results directly to the field application terminal.

[0058] Provide reference for operators: Present the working conditions in a concise and visual form to the operator's workstation, and provide standardized adjustment suggestions in conjunction with the AI ​​platform to reduce the operator's reliance on personal experience, achieve operation standardization, and overcome differences in operator skill levels.

[0059] This forms an application cycle: fully automatic annotation → machine learning → real-time working condition recognition → intelligent control / manual operation reference → fully automatic annotation (iteration).

[0060] The main operating conditions include: dust generation; long fire line, short fire line, and material deviation; thin material layer, thick material layer; and clumping.

[0061] Dust Pollution Assessment and Handling: Dust generated during combustion can sometimes obstruct the view from the flame camera, preventing operators from assessing the flame's condition via central control video. To address this, the flame analysis system uses built-in corner point detection to identify abnormal dust pollution conditions. When dust obstructs the view, it retains the previous image data to prevent abnormal data from affecting subsequent control, ensuring the accuracy of data transmission from the flame analysis system and alerting the operator to this situation.

[0062] Determining Long, Short, and Off-center Flame Conditions: During waste incineration, changes in waste composition and the lag in process adjustments mean that by the time DCS data changes, operator adjustments are often too late, leading to abnormal conditions such as excessively long or short flames, and consequently, abnormal combustion. To address this issue, the flame analysis system, considering process lag, traces back to the source flame data and uses flame intensity to initially determine the combustion location. However, due to the irregular shape of the flame, relying solely on flame intensity has limitations. Therefore, a "centroid" parameter is introduced, using the center of the entire combustion area as the centroid. By analyzing the overall flame movement trend, conditions such as excessively long or short flames, or off-center flames are identified. Simultaneously, cross-validation is performed by incorporating changes in the combustion temperature field to ensure the accuracy of the condition determination.

[0063] Determining the Thickness of the Feed Layer: Accurately determining the thickness of the feed layer during waste-to-energy combustion is crucial for achieving a proper air-feed ratio. An improper air-feed ratio can easily lead to abnormal conditions such as increased loss on ignition and feed layer burnout, affecting combustion stability. To address this issue, the flame analysis system acquires data on flame intensity and temperature color field changes under different combustion conditions, converting this data into indicators that can be used to determine feed layer thickness. This data is then input into a dedicated functional module for analysis. The system can also be linked to parameters in the DCS (Distributed Control System) such as air chamber pressure, grate temperature, and oxygen content through a data interface, enabling multi-source data fusion and judgment. This allows for accurate identification of abnormal feed layer conditions, providing operators with appropriate operational guidance.

[0064] The following section provides a further detailed implementation of the entire data processing workflow, from flame images to model building / iteration:

[0065] Step 1: Flame Image Acquisition and Preprocessing

[0066] Data source: Real-time acquisition of incinerator flame images via a high-temperature digital camera inside the furnace.

[0067] Preprocessing operations: (1) Noise removal: Gaussian filtering + median filtering combined algorithm is used to remove salt and pepper noise and Gaussian noise caused by dust and flue gas interference in the furnace; (2) Enhancement: Histogram equalization is used to improve the contrast of the flame area and highlight the flame edge and brightness layering; (3) Combustion area segmentation: Divide the combustion section into combustion section one, combustion section two, and burnout section (divided according to the waste-to-energy industry process and on-site grate conditions).

[0068] Step 2: Quantitative Extraction of Flame Features

[0069] Quantifiable metrics for extracting fundamental feature parameters from segmented flame images include:

[0070] 1) Geometric features, including: (1) Centroid X,Y coordinates. Source: Geometric center of the flame burning area; Acquisition method: Calculate the weighted average of the coordinates of all pixels in the flame area for the segmented binary image of the flame; (2) Burning area and total area. Source: Pixel statistics of the flame area; Acquisition method: Burning area = total number of pixels in the flame foreground area, total area = total number of pixels in the entire image (for normalization); (3) Perimeter of the contour line. Source: Flame outer contour; Acquisition method: Extract the closed contour of the flame through Canny / Sobel edge detection or contour tracking algorithm, and calculate the perimeter according to the pixel chain code length. (4) Roundness of the closed area. Source: Regularity of the flame contour shape; Acquisition method: Roundness = ,

[0071] The closer to 1, the closer to a circle; (5) Fire line position. Source: Determined by a combination of grayscale, minimum fire line width, and flame density; Acquisition method: Input grayscale, minimum fire line width, flame density features, etc. into a preset threshold; Output is the specific position of the fire line.

[0072] 2) Color channel features (RGB), including: (1) Red channel mean R, green channel mean G, blue channel mean B. Source: pixel intensity of each channel in the flame area; acquisition method: calculate the average value of R, G, and B channels for all pixels in the flame area; (2) Hue mean H, saturation mean S, brightness mean V. Source: RGB to HSV color space; acquisition method: convert each pixel in the flame area from RGB to HSV, and calculate the mean values ​​of H, S, and V respectively; (3) Gray level mean and gray level variance. Source: gray level image of the flame; acquisition method: gray level = 0.299R + 0.587G + 0.114B, gray level mean = average gray level of the flame area, gray level variance = dispersion of gray level values ​​in the flame area.

[0073] 3) Texture / quality features, including: (1) Roughness, sharpness, and smoothness. Source: Flame image texture and gradient; Sharpness acquisition method: Gradient amplitude mean / Laplacian variance; Roughness acquisition method: Gray-level difference statistics, local variance; Smoothness acquisition method: Evaluation index based on neighborhood gray-level change rate. All of the above are calculated within the flame area.

[0074] 4) Combustion physical quantity inversion: (1) Equivalent heat of combustion. Source: Comprehensive inversion of flame color, area, and brightness; Acquisition method: Based on the flame radiation characteristics, a mapping model is constructed through combustion area, average brightness, and color temperature / temperature characteristics to calculate the equivalent heat of combustion. (2) Average temperature. Source: Flame color temperature and radiation intensity; Acquisition method: Temperature is measured using the dual-color / monochromatic method. A calibration model is established using the correspondence between the R / G channel ratio, HSV color characteristics, and flame temperature to invert the average temperature of the flame area.

[0075] 5) Flame level / state, including: (1) Level 0 flame, Level 1 flame (combustion level). Source: judged by the above grayscale and area features; acquisition method: input grayscale, area, brightness, temperature, color features, etc. into the preset threshold / classification model, Level 0 flame: obvious combustion; Level 1 flame: incomplete combustion.

[0076] Step 3: Cross-calibration of multi-source data

[0077] Data Association: Using timestamps as unique keys, basic flame characteristic parameters are aligned with actual field conditions to form a "flame characteristic - actual situation" time-series sample. This includes:

[0078] Operating condition threshold matching: The operating condition feature parameter threshold library is called, and the basic feature parameters extracted in real time are compared with the preset thresholds to automatically determine the current operating condition type (such as dust, fire line material deviation, material layer thickness, clumping, etc.).

[0079] Label generation: The matched working condition type is used as a label and bound to the corresponding "flame feature - actual situation" sample to complete the automatic labeling and generate standardized machine learning samples.

[0080] Step 4: Model Building and Training

[0081] Model input: Working condition feature parameters (shape / brightness / color / texture) obtained from flames + professional experience and factory site working conditions + confirmation based on actual working conditions to set preliminary thresholds.

[0082] Model output results: dust conditions (conditions, the same below), clumping conditions, thin material conditions, thick material conditions, and left / right material deviation conditions.

[0083] Model selection: A lightweight CNN+LSTM hybrid model (CNN-LSTM hybrid neural network structure) is adopted to take into account both spatial feature extraction of flame images and temporal data dependency modeling; or a gradient boosting tree model (model structure) such as XGBoost is adopted to achieve efficient real-time inference.

[0084] The training process is based on the corresponding neural network structure or model structure. For example, it may include:

[0085] 1) Model training: Using labeled samples as supervision signals, optimize model parameters and minimize the error in identifying working conditions;

[0086] 2) Validation and evaluation: Evaluate model performance through accuracy, recall, and F1 score, and iteratively optimize features and thresholds.

[0087] Step 5: Automatic Model Iteration

[0088] After the newly acquired flame images are verified against the actual situation on site and new samples are generated through fully automated annotation, they are periodically incrementally injected into the model for online fine-tuning.

[0089] When the accuracy of working condition identification is lower than the preset threshold (such as 90%), full retraining is triggered to automatically update the threshold library of working condition feature parameters to adapt to changes in waste characteristics.

[0090] During the automated labeling process, the system uses preset threshold rules, working condition judgment logic, and AI algorithms to automatically determine the working condition type and automatically label (ash / material imbalance / thickness imbalance / clumping / normal) of massive flame images and operating data. The entire process does not require manual viewing of each image and labeling frame by frame, achieving fully automated, large-scale, and standardized labeling of sample data.

[0091] Automated annotation includes:

[0092] 1) Establishment of basic rules

[0093] Based on the initial working condition thresholds (characteristic parameter thresholds for dust, material imbalance, excessive thickness, and clumping) calibrated by industry experts, the core judgment rules for automated annotation are established.

[0094] 2) Multi-source data synchronous access

[0095] Real-time acquisition of key data includes morphological data (size, position, centroid coordinates, etc.), visual data (color HSV value, brightness, saturation, etc.), texture data (variance, number of corner points, flame intensity, etc.), ash data (number of corner points, occlusion rate), fire line / material deviation data (centroid movement trend, intensity distribution), material layer thickness data (temperature color field, fire line area), and agglomeration data (variance, distribution of characteristic points in the drying section).

[0096] 3) Automatic feature calculation and matching

[0097] The system compares the real-time calculated flame characteristic parameters with a preset / dynamically updated threshold library to automatically determine the current operating condition.

[0098] The specific labeling method for operating conditions is as follows:

[0099] If the threshold for dust-raising conditions is met, the system will be automatically marked as dust-raising.

[0100] If the material deviation threshold is met, it will be automatically marked as a hot-line material deviation.

[0101] If the material thickness threshold is met, the dynamic label indicates that the material layer is too thick.

[0102] Meeting the clumping threshold; automatically marked as clumped;

[0103] If none of the above thresholds are met, the system is automatically marked as operating normally.

[0104] 4) Threshold dynamic calibration (labeled and self-optimized)

[0105] By combining clustering and statistical analysis, the newly generated labeled samples are continuously learned and the thresholds are automatically optimized, making the labeling rules increasingly consistent with the current characteristics of waste and furnace conditions, thus continuously improving the labeling accuracy.

[0106] 5) Manual bottom-line calibration

[0107] Under extreme working conditions and special waste conditions, engineers can correct the annotation results with one click, forming a closed-loop annotation system of "automatic as the main method and manual as the auxiliary method".

[0108] 6) Labeled samples are stored in the database and directly used for model training.

[0109] The automatically labeled samples are directly entered into the model training library for real-time iteration of the working condition judgment model and optimization of inference accuracy.

[0110] Automated annotation has the following characteristics:

[0111] 1) Eliminate reliance on manual labor: No need for manual frame-by-frame annotation of flame images, improving annotation efficiency by over 90%;

[0112] 2) Standardized approach: Objective labeling is achieved based on a threshold library, eliminating differences in subjective experience among different operators;

[0113] 3) Strong real-time performance: It can complete the annotation of a single frame image within 100ms, which meets the real-time requirements of industry. In actual operation, operators often rely on on-site fire observation or DCS parameter adjustments, which are subject to lag. However, the operating condition results are output based on the source of DCS parameters (flame data), thus achieving early warning.

[0114] 4) Improved accuracy: By integrating flame visual features and process data, misjudgments based on a single DCS parameter are avoided, and the accuracy rate of operating condition identification is ≥95%;

[0115] 5) Supporting iteration: Continuously generating high-quality labeled samples to provide a data foundation for model self-learning.

[0116] The content and preset method of the threshold library for operating condition characteristic parameters are as follows:

[0117] The database stores the threshold values ​​of the characteristic parameters corresponding to each operating condition, for example:

[0118] Dust control: Utilizing corner points, grayscale, roughness, variance, sharpness, etc.

[0119] Fire line length: Utilizes factors such as center of mass, fire line position, and flame intensity;

[0120] Thin / thick material layer: average flame brightness, flame intensity, fire area, etc.;

[0121] Clustering: Flame texture complexity, flame variance, etc.

[0122] The threshold is set as follows:

[0123] 1) Initial threshold: determined by experts in the waste incineration industry based on historical stable operating data;

[0124] 2) Dynamic update: During model iteration, the threshold is automatically optimized through clustering and statistical analysis based on the feature distribution of newly labeled samples. For example, in judging dust conditions, the number of corner points was used as the judgment criterion in the early stage. Later, roughness, sharpness, and variance were added to re-label the samples to realize the dynamic update and iteration of the dust condition model.

[0125] 3) Manual calibration: Allows engineers to manually fine-tune the threshold under special working conditions to ensure compatibility.

[0126] This invention focuses on four (abnormal) operating conditions: ash emission, material deviation, material thickness, and agglomeration. These conditions represent the comprehensive operational status of solid waste within the incinerator grate area during feeding, pushing, and combustion. They fundamentally reflect material distribution, combustion front morphology, and physical fluidization characteristics, directly impacting furnace temperature field distribution, combustion stability, flue gas composition, and equipment safety. They serve as crucial data for DCS regulation and combustion optimization control of the incinerator. When the condition-specific characteristic parameters of a particular condition exceed a set threshold, it is marked as that (abnormal) condition. When the condition-specific characteristic parameters of multiple conditions all exceed their respective set thresholds, they are simultaneously marked as multiple (abnormal) conditions. When the condition-specific characteristic parameters of all conditions are within the set threshold range, it is marked as normal (normal operating condition). As needed, key operating conditions can be added, deleted, or replaced.

[0127] A working condition identification system can be built based on an existing DCS or other form of control system. The software for running the working condition identification method of the present invention is loaded into the DCS or other control system as the working condition identification system of the present invention, and the working condition identification method of the present invention is used to implement working condition identification (including automatic annotation, model building and iteration, etc.).

[0128] Unless otherwise specified, the preferred and optional technical means disclosed in this invention can be arbitrarily combined to form several different specific embodiments when one preferred or optional technical means is a further limitation of another technical means.

Claims

1. A fully automated method for identifying combustion conditions based on intelligent flame analysis, characterized in that... Based on flame image processing, the specific feature parameters of the working condition of interest are obtained. The specific feature parameters of the working condition of interest are matched with the threshold library of working condition feature parameters to obtain the corresponding working condition status. The threshold library of working condition feature parameters records the threshold of the specific feature parameters of the corresponding working condition. If the specific feature parameters of the working condition are within the corresponding threshold range, it is a normal working condition; if the specific feature parameters of the working condition exceed the corresponding threshold range, it is an abnormal working condition.

2. The fully automatic combustion condition identification method as described in claim 1, characterized in that... The operating conditions of interest include long fire line, short fire line, and off-center fuel conditions. The specific characteristic parameters for these operating conditions include the centroid of the combustion zone, the position of the fire line, and the flame intensity.

3. The fully automatic combustion condition identification method as described in claim 1, characterized in that... The operating conditions of interest include thin and thick material layers. The specific characteristic parameters for thin and thick material layer conditions include average flame brightness, flame intensity, and fire line area.

4. The fully automatic combustion condition identification method as described in claim 1, characterized in that... The operating conditions of interest include clumping conditions, and the specific characteristic parameters for clumping conditions include flame texture complexity and flame variance.

5. The fully automatic combustion condition identification method as described in claim 1, characterized in that... The operating conditions we are concerned with include dust-generating conditions, and the specific characteristic parameters for dust-generating conditions include the number of corner points and the dust obstruction rate.

6. The fully automatic method for identifying combustion conditions as described in any one of claims 1-5, characterized in that... Based on historical data, sample data suitable for model training is constructed. The model input data in the sample data consists of flame images, basic feature parameters of flame images, or working condition-specific feature parameters of flame images. The model output data is the working condition of interest. A working condition judgment model is constructed based on the sample data.

7. The fully automatic combustion condition identification method as described in claim 6, characterized in that... The working condition judgment model is a lightweight CNN+LSTM hybrid model trained with a CNN-LSTM hybrid neural network structure, or an XGBoost judgment model trained with an XGBoost model structure.

8. The fully automatic combustion condition identification method as described in claim 6, characterized in that... The judgment model is automatically iterated. New sample data is obtained by processing newly acquired flame images and injected into the model incrementally at regular intervals. Through iterative training, the judgment model is fine-tuned online. When the accuracy of the working condition recognition is lower than the preset threshold, full retraining is triggered.

9. A fully automated combustion condition identification system based on intelligent flame analysis, characterized in that... The combustion condition identification method based on intelligent flame analysis, as described in any one of claims 1-8, includes: Image acquisition device for continuously acquiring images of flames; The image preprocessing module is used to preprocess the flame image data from the image acquisition device, including noise reduction, enhancement and combustion area segmentation, to form preprocessed flame image data. The automatic annotation module is used to extract feature parameters from the preprocessed flame image data to obtain the basic feature parameters and working condition-specific feature parameters of the flame image. The working condition-specific feature parameters of the working condition of interest are matched with the working condition feature parameter threshold library to obtain the corresponding working condition state. The model input data is the flame image, the basic feature parameters of the flame image or the working condition-specific feature parameters of the flame image, and the model output data is the working condition state of the working condition of interest. The sample data can be stored in the sample database. The machine learning module is used to build and maintain a working condition judgment model using sample data. After the judgment model is established, it is used to iterate or rebuild the model based on newly added sample data. The prediction module is used to judge the working condition of the target working condition based on real-time model input data and output the judgment result. The display module is used to display the judgment results.

10. The fully automatic combustion condition identification system as described in claim 9, characterized in that... When the judgment result output by the prediction module includes an abnormal operating condition, an alarm signal is generated and sent to the display along with the judgment result. The judgment result and other necessary data are also transmitted to the expert system, which proposes countermeasures and returns them to the prediction module. The prediction module then sends the countermeasures to the display device for display along with the prediction result and the alarm signal.

Images