Boiler combustion early warning method based on improved deep forest algorithm

Abstract

The invention discloses a boiler combustion early warning method based on an improved deep forest algorithm. The method comprises the following steps: obtaining an initial quantitative index of a combustion state of combustion equipment; determining a dynamic fluctuation mode of the combustion process of the combustion equipment under the multi-parameter coupling condition; constructing a multi-level feature evaluation structure by adopting an improved deep forest algorithm to obtain a credibility score of the combustion features; if the credibility score of the combustion feature is lower than a preset threshold value, obtaining an adjusted intermediate layer risk feature; judging a triggering condition of a potential coking risk; the source position of abnormal combustion of the combustion equipment is determined; and generating a real-time early warning signal according to the source position of the abnormal combustion of the combustion equipment. The problems that a traditional scheme is insufficient in combustion abnormity early warning capacity and lagged in response under the multi-source complex working condition are effectively solved.

Description

Technical Field

[0001] This invention relates to the field of boiler technology, and in particular to an early warning method for boiler combustion based on an improved deep forest algorithm. Background Technology

[0002] During the operation of industrial boilers, real-time monitoring and early warning of combustion status are crucial for ensuring equipment safety and production efficiency.

[0003] Boiler combustion involves the complex interaction of multiple parameters. Abnormalities can lead to equipment damage or even safety accidents. Therefore, researching how to accurately identify combustion anomalies and trace their root causes through intelligent means has become a key issue that urgently needs to be addressed in the industrial sector.

[0004] Many current monitoring methods struggle to accurately quantify the reliability of their assessments when faced with complex changes in combustion status. Existing technologies rely heavily on single-level data analysis, lacking the ability to dynamically evaluate information interactions at different levels. This is especially true in situations involving multi-parameter coupling, where it's difficult to effectively distinguish between normal fluctuations and potential risks. This limitation often leads to misjudgments or omissions when dealing with ambiguous states, failing to provide maintenance personnel with clear decision-making support.

[0005] Meanwhile, the assessment of combustion status needs to be progressively deepened from bottom-level sensor data to higher-level risk characteristics, and the reliability of the judgment at each layer is affected by the quality of information in the previous layer. If the reliability assessment of the bottom-level data is inaccurate, it will lead to biases in the analysis of subsequent layers, thereby affecting the overall reliability of the early warning. Summary of the Invention

[0006] One objective of this invention is to propose an early warning method for boiler combustion based on an improved deep forest algorithm. This invention effectively solves the problems of insufficient early warning capability and delayed response of traditional solutions for combustion anomalies under complex multi-source operating conditions.

[0007] An early warning method for boiler combustion based on an improved deep forest algorithm according to an embodiment of the present invention, the method comprising:

[0008] Collect temperature data, oxygen data, and flame image data inside the combustion chamber of the combustion equipment, extract the low-level feature vectors of the temperature data, oxygen data, and flame image data, and obtain the initial quantitative indicators of the combustion state of the combustion equipment.

[0009] A time series analysis model was used to process the changing trends of the initial quantitative indicators and to determine the dynamic fluctuation pattern of the combustion process of the combustion equipment under multi-parameter coupling conditions.

[0010] An improved deep forest algorithm was used to construct a multi-level feature evaluation structure. The combustion features of each layer were analyzed by initial quantification indicators to obtain the confidence score of the combustion features.

[0011] If the confidence score of the combustion feature is lower than the preset threshold, the influencing factors are calibrated based on the weights of the underlying feature vector through a weighted fusion mechanism to obtain the adjusted intermediate layer risk features.

[0012] Based on the adjusted intermediate layer risk characteristics, the probability distribution of high-level anomalies in the combustion process of the combustion equipment is obtained to determine the triggering conditions for potential coking risks.

[0013] By using the triggering conditions of the potential coking risk, a causal analysis model is employed to trace the interaction paths between multiple parameters and determine the root cause of combustion abnormalities in the combustion equipment.

[0014] A real-time early warning signal is generated based on the root cause of the combustion abnormality in the combustion equipment.

[0015] Optionally, the extraction of the low-level feature vectors from the temperature data, oxygen data, and flame image data includes:

[0016] Temperature and oxygen levels inside the furnace are collected in real time using temperature and oxygen sensors, while flame image data is acquired using an image sensor.

[0017] The collected temperature data, oxygen data, and flame image data are processed by feature extraction to generate corresponding feature vectors and determine the underlying characterization information of the combustion state.

[0018] Based on the generated feature vector, the changing trends of temperature and oxygen data are analyzed. If the changing trend exceeds the preset threshold range, an abnormal state indicator is triggered to determine whether there is a potential combustion instability.

[0019] By performing pattern comparison on the feature vectors extracted from flame image data and combining them with a pre-established flame morphology database, the morphology category that best matches the current combustion state is obtained, and the visual representation result of the combustion process is determined.

[0020] A comprehensive analysis is performed on the underlying characterization information and visual characterization results of the combustion state. If the correlation between temperature data and oxygen data is lower than a preset threshold, the feature vector is weighted and adjusted to obtain the combustion state assessment result.

[0021] Based on the combustion status assessment results and combined with historical data of equipment status, the operational stability inside the furnace is analyzed, the status change characteristics under long-term operating trends are obtained, and a complete set of initial quantitative indicators are constructed.

[0022] Optionally, determining the dynamic fluctuation mode of the combustion process of the combustion device under multi-parameter coupling conditions includes:

[0023] The initial data is segmented based on the combustion status assessment results of the combustion equipment, and the changing trends within different time windows are extracted to obtain the stage fluctuation information of the combustion process.

[0024] Based on the phased fluctuation information, and combined with the coupled scenario of multiple parameter groups, a correlation analysis was conducted on multiple variables in the combustion process to determine the main driving factors of dynamic fluctuations.

[0025] By extracting the main driving factors of dynamic fluctuations, a framework for identifying the regularity of the combustion process is constructed to obtain the response patterns of the combustion state under different coupled scenarios.

[0026] If the response mode is within the preset threshold range, the dynamic fluctuations of the combustion process will be continuously tracked to determine whether there are any abnormal deviations.

[0027] If the data exceeds the threshold range, the fluctuation data of the current time series is recorded to obtain the anomaly labeling result;

[0028] Based on the anomaly marking results, a localized analysis is performed on the combustion status assessment results of the combustion equipment. The current dynamic fluctuations are compared with a pre-established reference database to identify potential sources of fluctuations.

[0029] By analyzing the potential sources of fluctuations and combining the changing trends of the time series, a dynamic adjustment basis for the combustion process is generated, and optimization reference information for multi-parameter coupling scenarios is obtained.

[0030] Based on the optimized reference information, the operating data of the combustion equipment is continuously collected and compared to determine whether the combustion state tends to be stable, thus obtaining the final dynamic fluctuation mode.

[0031] Optionally, the construction of a multi-level feature evaluation structure using an improved deep forest algorithm includes:

[0032] The initial quantitative indicators of the combustion state of the combustion equipment are obtained, and the dynamic fluctuations are collected in real time. Various parameters during the combustion process are recorded by sensors to obtain the raw fluctuation data set.

[0033] The original fluctuation data set is denoised and standardized, and the basic form of the fluctuation pattern is determined by extracting dynamic fluctuation segments in segments.

[0034] Based on the basic morphology of the wave pattern, a multi-layer feature extraction structure is constructed, and the deep forest algorithm is used to decompose and represent the combustion features of each layer to obtain a hierarchical feature set.

[0035] For the hierarchical feature set, a layer-by-layer analysis process is implemented. If the significance of a feature in a certain layer is lower than a preset threshold, its weight is adjusted to determine the preliminary importance ranking of each layer of features.

[0036] Based on the initial importance ranking, the confidence score of each combustion feature is calculated, and the scores of each feature are integrated by weighted accumulation to determine the final confidence score.

[0037] If the final credibility score deviates from the preset range, a feature re-evaluation mechanism is triggered. This mechanism performs local corrections by backtracking the hierarchical feature set, resulting in an adjusted credibility result.

[0038] Optionally, the adjusted intermediate layer risk characteristics include:

[0039] If the confidence score of the combustion feature is lower than the preset threshold, the preliminary structure of the weight distribution is determined by analyzing the weights of the underlying feature vectors one by one.

[0040] If the initial structure of the weight distribution deviates from the preset threshold, the vector weights are calibrated through a weighted fusion mechanism to obtain the calibrated weight combination.

[0041] Based on the calibrated weight combination, the risk characteristics of the intermediate level are calculated, their correlation with combustion characteristics is analyzed, and a quantitative representation of the risk characteristics is obtained.

[0042] By quantifying risk characteristics, a mapping logic for feature adjustment is constructed to determine whether the adjusted features meet expectations and obtain the final correction result.

[0043] Based on the final correction results, intermediate layer risk characteristics of combustion features are generated to identify potential risk points in the combustion process.

[0044] Optionally, the step of obtaining the high-level anomaly probability distribution of the combustion process of the combustion equipment and determining the triggering conditions for potential coking risk includes:

[0045] Based on the adjusted risk characteristics of the intermediate layer, the probability distribution of high-rise anomalies is obtained by real-time data collection during the operation of the combustion equipment, thereby determining the range of possible anomalies.

[0046] Based on the probability distribution of high-level anomalies, a preset threshold is used for screening. If the probability distribution exceeds the threshold range, the abnormal data is stratified to obtain the key attention interval for high-level anomalies.

[0047] For key areas of concern, extract risk feature data related to intermediate layer risk characteristics, determine the degree of correlation between potential coking and high-level anomalies, and obtain the priority ranking of risk features;

[0048] By prioritizing and analyzing the triggering conditions for potential coking risks, and combining this with real-time changes in equipment status, the specific manifestations of the triggering conditions are determined.

[0049] Optionally, the step of using a causal analysis model to trace the interaction paths between multiple parameters and determine the root cause of combustion abnormalities in the combustion equipment includes:

[0050] By analyzing the triggering conditions of the potential coking risk, the logical relationships between parameter changes are sorted out using causal analysis methods, and the interaction paths between multiple parameters are obtained.

[0051] Based on the interaction paths between multiple parameters, the direct correlation between combustion anomalies and changes in each parameter is analyzed to determine the preliminary root cause of the combustion anomalies.

[0052] Based on the initial root cause location, real-time monitoring data of the equipment status is obtained, and combined with dynamic information on parameter changes, the key influencing factors for anomaly determination are identified.

[0053] By identifying key influencing factors, a framework for assessing coking risk is constructed. If parameter changes exceed a preset threshold range, the anomaly detection is prioritized to determine the key areas for risk assessment.

[0054] Based on the key directions of risk assessment, specific data on the triggering conditions of potential coking risks are extracted, the abnormal probability of combustion equipment under different operating scenarios is analyzed, the distribution range of the triggering conditions of potential coking risks is determined, and the root cause of combustion abnormalities in combustion equipment is identified.

[0055] Optionally, generating a real-time early warning signal based on the root cause of the combustion abnormality in the combustion device includes:

[0056] To pinpoint the root cause of combustion abnormalities in combustion equipment, historical operating records related to the root cause of the combustion abnormality are obtained. By comparing the deviation between current data and historical data, the specific category of the root cause of the abnormality is determined.

[0057] From the specific categories of the anomaly root cause, obtain the corresponding impact range data, classify the impact range, and determine the priority of real-time early warning and the signal generation method;

[0058] Based on the priority of real-time warnings, corresponding warning signals are generated and transmitted to the control module through the internal system channel to obtain the signal-triggered response mechanism.

[0059] Regarding the response mechanism, adjustment instructions for the combustion process are generated by combining optimization methods, and then sent to the combustion equipment through the instruction execution module to complete process monitoring and status updates.

[0060] The beneficial effects of this invention are:

[0061] (1) This invention extracts the low-level feature vectors of multi-source heterogeneous information and models the dynamic fluctuation mode under multi-parameter coupling by combining time series analysis model. It can perform in-depth modeling of various dynamic changes and multi-variable coupling behaviors in the boiler combustion process. Through hierarchical quantitative indicators and time series analysis, it can reflect the abnormal combustion trend and potential instability signal in the furnace in real time and in detail. It effectively solves the problem of insufficient early warning capability and delayed response of traditional schemes for combustion anomalies under multi-source complex working conditions.

[0062] (2) This invention introduces the deep forest algorithm into the field of boiler combustion anomaly early warning, and makes structural improvements on this basis to construct a multi-level feature evaluation framework: hierarchical modeling and weighted fusion of dynamic fluctuation features, risk features and high-level anomaly probability distribution, which can automatically identify the importance of each layer of features and dynamically adjust the weight distribution, effectively suppressing false alarms and false alarms caused by single feature anomalies. Through layer-by-layer credibility analysis and local correction mechanism, stronger anomaly detection robustness is achieved, and at the same time, it has high model interpretability. Attached Figure Description

[0063] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0064] Figure 1 This is a flowchart of an early warning method for boiler combustion based on an improved deep forest algorithm proposed in this invention. Detailed Implementation

[0065] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0066] refer to Figure 1 As shown in Example 1: An early warning method for boiler combustion based on an improved deep forest algorithm, the method includes:

[0067] Collect temperature data, oxygen data, and flame image data inside the combustion chamber of the combustion equipment, extract the low-level feature vectors of the temperature data, oxygen data, and flame image data, and obtain the initial quantitative indicators of the combustion state of the combustion equipment.

[0068] A time series analysis model was used to process the changing trends of the initial quantitative indicators and to determine the dynamic fluctuation pattern of the combustion process of the combustion equipment under multi-parameter coupling conditions.

[0069] An improved deep forest algorithm was used to construct a multi-level feature evaluation structure. The combustion features of each layer were analyzed by initial quantification indicators to obtain the confidence score of the combustion features.

[0070] If the confidence score of the combustion feature is lower than the preset threshold, the influencing factors are calibrated based on the weights of the underlying feature vector through a weighted fusion mechanism to obtain the adjusted intermediate layer risk features.

[0071] Based on the adjusted intermediate layer risk characteristics, the probability distribution of high-level anomalies in the combustion process of the combustion equipment is obtained to determine the triggering conditions for potential coking risks.

[0072] By using the triggering conditions of the potential coking risk, a causal analysis model is employed to trace the interaction paths between multiple parameters and determine the root cause of combustion abnormalities in the combustion equipment.

[0073] A real-time early warning signal is generated based on the root cause of the combustion abnormality in the combustion equipment.

[0074] In industrial boiler systems, data is collected in real time using thermocouples and zirconia probes installed at key locations in the furnace.

[0075] The sensor collects temperature and oxygen concentration data once per second. In Example 1, the temperature range is 800-1200 degrees Celsius and the oxygen content is between 2-5%. At the same time, a CCD camera is used to capture flame images to form a multi-dimensional data basis. The preliminary acquisition results include time series data, temperature curves and oxygen fluctuation graphs, as well as RGB image sequences of the flame.

[0076] By comparing the feature vector of the flame image with a pre-established flame morphology database, the database in Example 1 includes categories such as normal conical flame, deflected flame, and flashing flame. The matching degree is calculated using cosine similarity to obtain the most matching morphology in the current state, thereby determining the visual representation result of the combustion process.

[0077] A comprehensive analysis of the underlying characterization information and visual characterization results is performed. If the Pearson correlation coefficient between temperature and oxygen data is lower than the preset threshold of 0.8, the weight of the feature vector is adjusted, and the weight of the oxygen vector is increased to 0.6 to obtain a more accurate combustion state assessment result. In Example 1, the assessment score is increased from 0.7 to 0.9, indicating more stable combustion.

[0078] Based on the evaluation results and the analysis of historical equipment data, the stability of furnace operation was determined, and long-term trend characteristics were obtained. The gradual increase in the average temperature indicated a carbon buildup problem, and optimization directions such as adjusting fuel input were identified.

[0079] In this embodiment, the extraction of the low-level feature vectors from temperature data, oxygen content data, and flame image data includes:

[0080] Temperature and oxygen levels inside the furnace are collected in real time using temperature and oxygen sensors, while flame image data is acquired using an image sensor.

[0081] The collected temperature data, oxygen data, and flame image data are processed by feature extraction to generate corresponding feature vectors and determine the underlying characterization information of the combustion state.

[0082] Based on the generated feature vector, the changing trends of temperature and oxygen data are analyzed. If the changing trend exceeds the preset threshold range, an abnormal state indicator is triggered to determine whether there is a potential combustion instability.

[0083] By performing pattern comparison on the feature vectors extracted from flame image data and combining them with a pre-established flame morphology database, the morphology category that best matches the current combustion state is obtained, and the visual representation result of the combustion process is determined.

[0084] A comprehensive analysis is performed on the underlying characterization information and visual characterization results of the combustion state. If the correlation between temperature data and oxygen data is lower than a preset threshold, the feature vector is weighted and adjusted to obtain the combustion state assessment result.

[0085] Based on the combustion status assessment results and combined with historical data of equipment status, the operational stability inside the furnace is analyzed, the status change characteristics under long-term operating trends are obtained, and a complete set of initial quantitative indicators are constructed.

[0086] In this embodiment, determining the dynamic fluctuation mode of the combustion process of the combustion device under multi-parameter coupling conditions includes:

[0087] The initial data is segmented based on the combustion status assessment results of the combustion equipment, and the changing trends within different time windows are extracted to obtain the stage fluctuation information of the combustion process.

[0088] Based on the phased fluctuation information, and combined with the coupled scenario of multiple parameter groups, a correlation analysis was conducted on multiple variables in the combustion process to determine the main driving factors of dynamic fluctuations.

[0089] By extracting the main driving factors of dynamic fluctuations, a framework for identifying the regularity of the combustion process is constructed to obtain the response patterns of the combustion state under different coupled scenarios.

[0090] The pattern recognition framework includes a rule-based system that defines expected patterns under various coupled scenarios. Under high-load scenarios, temperature should be positively correlated with oxygen content. By comparing actual data with the patterns, the response patterns of combustion state under different scenarios can be obtained.

[0091] If the response mode is within the preset threshold range, the dynamic fluctuations of the combustion process will be continuously tracked to determine whether there are any abnormal deviations.

[0092] If the data exceeds the threshold range, the fluctuation data of the current time series is recorded to obtain the anomaly labeling result;

[0093] Based on the anomaly marking results, a localized analysis is performed on the combustion status assessment results of the combustion equipment. The current dynamic fluctuations are compared with a pre-established reference database to identify potential sources of fluctuations.

[0094] By analyzing the potential sources of fluctuations and combining the changing trends of the time series, a dynamic adjustment basis for the combustion process is generated, and optimization reference information for multi-parameter set coupling scenarios is obtained.

[0095] Based on the optimized reference information, the operating data of the combustion equipment is continuously collected and compared to determine whether the combustion state tends to be stable, thus obtaining the final dynamic fluctuation mode.

[0096] In one possible implementation, if the response mode is within a preset threshold range, such as a temperature fluctuation of less than 2%, the dynamic fluctuations of the combustion process are continuously tracked, and the presence of abnormal deviations is determined by real-time monitoring of the data stream. If the response mode exceeds the threshold range, the fluctuation data of the current time series is recorded. In Example 1, the peak data at a specific time point is marked to obtain the abnormal marking result. This can provide timely warnings of potential faults in industrial boiler monitoring.

[0097] In this embodiment, the construction of a multi-level feature evaluation structure using an improved deep forest algorithm includes:

[0098] The initial quantitative indicators of the combustion state of the combustion equipment are obtained, and the dynamic fluctuations are collected in real time. Various parameters during the combustion process are recorded by sensors to obtain the raw fluctuation data set.

[0099] The original fluctuation data set is denoised and standardized, and the basic form of the fluctuation pattern is determined by extracting dynamic fluctuation segments in segments.

[0100] Based on the basic morphology of the wave pattern, a multi-layer feature extraction structure is constructed, and the deep forest algorithm is used to decompose and represent the combustion features of each layer to obtain a hierarchical feature set.

[0101] For the hierarchical feature set, a layer-by-layer analysis process is implemented. If the significance of a feature in a certain layer is lower than a preset threshold, its weight is adjusted to determine the preliminary importance ranking of each layer of features.

[0102] In Example 1, when analyzing the combustion characteristics of the boiler, the third layer characteristics of the deep forest were examined. If the significance of the oxygen concentration correlation was only 0.3, its weight coefficient was increased to 1.2, and the temperature characteristics were ranked as the most important.

[0103] The process of calculating the confidence score of each combustion feature layer adopts a weighted accumulation method. In Example 1, the weights for each layer are assigned as follows: 0.2 for the first layer, 0.3 for the second layer, and 0.5 for the third layer. Then, the scores of each layer are accumulated to obtain the final value. In the boiler scenario, if the temperature layer score is 0.8, the oxygen layer score is 0.7, and the fuel layer score is 0.6, the final confidence score after weighting is 0.71, which reflects the overall reliability of the combustion features and provides data support for optimizing combustion control, thereby improving fuel utilization and reducing emission risks in business operations.

[0104] When the final credibility score deviates from the preset range of 0.6 to 0.9, a feature re-evaluation mechanism is triggered, which locally corrects the abnormal layer by backtracking the hierarchical feature set.

[0105] In the case of low-oxygen combustion in the boiler, if the score drops to 0.5, the second layer of oxygen characteristics is traced back and its representation vector is corrected to obtain an adjusted result of 0.75. This ensures the accuracy of the analysis. When generating a dynamic fluctuation mode description based on the adjusted confidence result, it can be mapped to the operating status of the combustion equipment, with high confidence scores corresponding to stable modes and low scores corresponding to potential fault trends.

[0106] In boiler maintenance, if the confidence level is 0.8, it is described as "the temperature fluctuation pattern tends to be stable and the oxygen supply is sufficient", indicating that efficient combustion may be maintained in the next hour. In actual business, this can guide timely adjustment of parameters to achieve preventive maintenance.

[0107] Based on the initial importance ranking, the confidence score of each combustion feature is calculated, and the scores of each feature are integrated by weighted accumulation to determine the final confidence score.

[0108] If the final credibility score deviates from the preset range, a feature re-evaluation mechanism is triggered. This mechanism performs local corrections by backtracking the hierarchical feature set, resulting in an adjusted credibility result.

[0109] Based on the business content and related attributes of the combustion equipment, the following solutions are generated:

[0110] Combustion process data is acquired during the operation of the combustion equipment, and real-time acquisition is performed to monitor dynamic fluctuations. Sensors record various parameters during the combustion process, resulting in raw fluctuation data sets. These raw data sets are then preprocessed for noise reduction and standardization. Dynamic fluctuation segments are extracted in segments to determine the basic form of the fluctuation pattern. Based on this basic pattern, a multi-layer feature extraction structure is constructed. A deep forest algorithm is used to decompose and represent each layer of combustion features, resulting in a hierarchical feature set. For each hierarchical feature set, a layer-by-layer analysis process is implemented. If the significance of a feature at a certain layer is lower than a preset threshold, its weight is adjusted to determine the preliminary importance ranking of each layer's features. Based on this preliminary importance ranking, a confidence score for each layer's combustion features is calculated. The scores of each layer's features are then integrated using a weighted accumulation method to determine the final confidence score. If the final confidence score deviates from a preset range, a feature re-evaluation mechanism is triggered. This involves backtracking through the hierarchical feature set for local correction, resulting in an adjusted confidence score. Based on the adjusted confidence score, a dynamic fluctuation pattern description of the combustion features is generated. By mapping this description to the operating status of the combustion equipment, the potential trend of the fluctuation pattern is determined.

[0111] In this embodiment, the adjusted intermediate layer risk characteristics include:

[0112] If the confidence score of the combustion feature is lower than the preset threshold, the preliminary structure of the weight distribution is determined by analyzing the weights of the underlying feature vectors one by one.

[0113] If the initial structure of the weight distribution deviates from the preset threshold, the vector weights are calibrated through a weighted fusion mechanism to obtain the calibrated weight combination.

[0114] Based on the calibrated weight combination, the risk characteristics of the intermediate level are calculated, their correlation with combustion characteristics is analyzed, and a quantitative representation of the risk characteristics is obtained.

[0115] By quantifying risk characteristics, a mapping logic for feature adjustment is constructed to determine whether the adjusted features meet expectations and obtain the final correction result.

[0116] Based on the final correction results, intermediate layer risk characteristics of combustion features are generated to identify potential risk points in the combustion process.

[0117] In this embodiment, the step of obtaining the high-level anomaly probability distribution of the combustion process of the combustion equipment and determining the triggering conditions for potential coking risk includes:

[0118] Based on the adjusted risk characteristics of the intermediate layer, the probability distribution of high-rise anomalies is obtained by real-time data collection during the operation of the combustion equipment, thereby determining the range of possible anomalies.

[0119] Based on the probability distribution of high-level anomalies, a preset threshold is used for screening. If the probability distribution exceeds the threshold range, the abnormal data is stratified to obtain the key attention interval for high-level anomalies.

[0120] For key areas of concern, extract risk feature data related to intermediate layer risk characteristics, determine the degree of correlation between potential coking and high-level anomalies, and obtain the priority ranking of risk features;

[0121] By prioritizing and analyzing the triggering conditions for potential coking risks, and combining this with real-time changes in equipment status, the specific manifestations of the triggering conditions are determined.

[0122] In Example 1, if the probability of the temperature exceeding the normal range is 0.3, the preliminary distribution shows that the probability of an anomaly is in the range of 20% to 50%, thereby determining the range of possible occurrence of the anomaly, which helps to identify potential problems early and ensure the stability of the combustion process.

[0123] Based on the preliminary probability distribution, a preset threshold such as 0.4 is used for screening. If the distribution exceeds this range, the abnormal data is processed in layers. During boiler operation, if the probability distribution shows that the probability of temperature anomaly reaches 0.6, the data is divided into a high-frequency anomaly layer and a low-frequency anomaly layer. The data is grouped using K-means to obtain the key attention intervals for high-level anomalies, such as the temperature peak interval within a specific time period. When extracting risk feature data related to the intermediate level for the key attention intervals, coking-related features such as carbon deposition rate and heat transfer efficiency are extracted from the data to determine the degree of correlation between potential coking and high-level anomalies.

[0124] Using the Pearson correlation coefficient, if the correlation coefficient between coking characteristics and anomaly probability is 0.7, the priority ranking of risk characteristics is obtained, with carbon deposition rate ranked first. Correlation judgment helps to reveal the root cause of anomalies and improve the accuracy of detection.

[0125] By prioritizing and analyzing the triggering conditions for potential coking, and combining this with real-time changes in equipment status such as fluctuations in fuel input rate, the specific manifestations of the triggering conditions can be determined.

[0126] In a boiler system, if the carbon deposition rate has a high priority and the fuel input suddenly increases, the triggering condition is a continuous state where the gas flow rate exceeds 2 m / s. This is confirmed through time series analysis to ensure the accuracy of the condition description. When obtaining operational risk data during operation based on the specific manifestation of the triggering condition, equipment vibration and pressure data can be collected to determine whether the equipment status is close to the pressure threshold of 5 bar for abnormal detection, thus obtaining a dynamic assessment result of the operational risk.

[0127] If vibration data shows a trend approaching the critical point, the assessment result is quantified as a high-risk level. Dynamic assessment can bring about timely intervention in actual operation, reducing equipment failure rate.

[0128] In this embodiment, the step of using a causal analysis model to trace the interaction paths between multiple parameters and determine the root cause of combustion abnormalities in the combustion equipment includes:

[0129] By analyzing the triggering conditions of the potential coking risk, the logical relationships between parameter changes are sorted out using causal analysis methods, and the interaction paths between multiple parameters are obtained.

[0130] Causal analysis methods represent the dependencies between variables by constructing causal graphs. In Example 1, an increase in temperature may lead to an increase in pressure, thereby affecting the oxygen content. For instance, in a real-world operating scenario of a pulverized coal boiler, a sudden increase in fuel flow can trigger a sharp rise in temperature. Pearson correlation coefficient analysis is used to quantify the strength of the correlation, and then an interaction path graph from fuel flow to temperature and then to pressure is plotted to help identify potential chain reactions.

[0131] Based on the interaction paths between multiple parameters, the direct correlation between combustion anomalies and changes in each parameter is analyzed to determine the preliminary root cause of the combustion anomalies.

[0132] Based on the initial root cause location, real-time monitoring data of the equipment status is obtained, and combined with dynamic information on parameter changes, the key influencing factors for anomaly determination are identified.

[0133] By identifying key influencing factors, a framework for assessing coking risk is constructed. If parameter changes exceed a preset threshold range, the anomaly detection is prioritized to determine the key areas for risk assessment.

[0134] Based on the key directions of risk assessment, specific data on the triggering conditions of potential coking risks are extracted, the abnormal probability of combustion equipment under different operating scenarios is analyzed, the distribution range of the triggering conditions of potential coking risks is determined, and the root cause of combustion abnormalities in combustion equipment is identified.

[0135] In Example 1, if the interaction path in the boiler system shows that a decrease in oxygen content directly leads to flame instability, the root cause may be located in the oxygen supply system. By comparing historical data and real-time path diagrams, it can be confirmed that the anomaly originates from a faulty oxygen supply valve. Based on this preliminary root cause, real-time monitoring data of the equipment status is obtained, and combined with dynamic information of parameter changes such as the fluctuation curve of oxygen content, the key influencing factors for anomaly determination are identified.

[0136] Specifically, during operation, if the oxygen content drops from 21% to 18% in a short period of time, combined with valve opening data, it can be determined that the key factor is valve response delay, which helps to understand how the anomaly spreads from a local problem to the overall combustion process.

[0137] In one embodiment, a coking risk assessment framework is constructed through key influencing factors. In Embodiment 1, the framework includes a multi-layered index system, with temperature and ash content as core indicators. If parameter changes, such as temperature exceeding a preset threshold range of 800 degrees Celsius, are prioritized to determine the key areas for risk assessment. For example, priority is given to the possibility of coking in high-temperature zones, which can effectively improve the accuracy of risk identification and bring the benefit of reducing downtime in actual boiler maintenance.

[0138] In this embodiment, generating a real-time early warning signal based on the root cause of the combustion abnormality in the combustion device includes:

[0139] To pinpoint the root cause of combustion abnormalities in combustion equipment, historical operating records related to the root cause of the combustion abnormality are obtained. By comparing the deviation between current data and historical data, the specific category of the root cause of the abnormality is determined.

[0140] From the specific categories of the anomaly root cause, obtain the corresponding impact range data, classify the impact range, and determine the priority of real-time early warning and the signal generation method;

[0141] Based on the priority of real-time warnings, corresponding warning signals are generated and transmitted to the control module through the internal system channel to obtain the signal-triggered response mechanism.

[0142] Regarding the response mechanism, adjustment instructions for the combustion process are generated by combining optimization methods, and then sent to the combustion equipment through the instruction execution module to complete process monitoring and status updates.

[0143] Example 2: During the monitoring of a large industrial boiler's normal operation, the boiler monitoring platform automatically collected temperature data from different spatial locations at 5-minute intervals. , , , oxygen content , , The data includes multi-channel image data from three flame imaging probes. After being transmitted to the analysis platform, the data undergoes synchronization and noise reduction processing. Taking data from a specific continuous time period as an example... The regional temperature gradually rose from 1170℃ to 1216℃. The oxygen content slowly decreased from 2.25% to 1.66%. At the same time, the image feature code captured by the flame probe jumped from 3 to 6 in a short period of time. Image analysis showed that the flame was elongated and the brightness distribution was uneven, with abnormal bright spot peaks in some areas.

[0144] Based on the aforementioned raw data, the system extracts low-level features for each time window, including the average temperature at each point, the maximum rate of temperature rise, the extreme value of oxygen content, the fluctuation amplitude, the edge contour parameters of the flame morphology, and the principal component features of brightness. For example, during the period from 06:15 to 06:30, The average temperature was 1197.3℃, and the maximum rate of change was 3.2℃ / min. The mean value was 1.81%, the minimum value was 1.66%, the mean edge length of the flame shape increased from 76 pixels to 94 pixels, and the peak value of the principal component of brightness increased by 8.7%. This series of low-level features were integrated into a feature vector and input into the next analysis module.

[0145] The analysis platform employs time series modeling to perform collaborative trend mining on three features: temperature, oxygen content, and flame morphology. Through short-time Fourier transform and sliding window statistics, the system identifies the time period from 06:20 to 06:30. and The correlation coefficient between the parameters decreased from -0.31 to -0.57, indicating a decline in fluctuation consistency, which is a typical precursor to multi-parameter coupling instability. During this period, the rate of change of flame image feature encoding also increased significantly, which the model automatically identified as a high-risk fluctuation phase. Further, through a trend segmentation algorithm, the system detected... The oxygen content remained below 2.0% at a rate of -0.045% / min, a characteristic that was categorized as a "high-risk zone for coking" by historical samples.

[0146] After all feature data is fed into the improved deep forest model, the system first performs a hierarchical feature importance ranking. The base layer initially scores the importance of temperature, oxygen content, and flame imagery, with temperature having a weight of 0.43, oxygen content a weight of 0.39, and flame morphology a weight of 0.18. The second layer adjusts the weight distribution based on dynamic trends and fluctuations. The synergy between temperature change and flame elongation was given greater attention, with weights adjusted to 0.53 and 0.24 respectively. Ultimately, the multi-parameter confidence score for the current time period dropped to 0.65, below the safety threshold of 0.75, triggering the system's automatic feature weight calibration mechanism.

[0147] During feature weight calibration, the system performs weighted correction on the degree of coordinated fluctuation of underlying features. Detected The coefficient of variation increased to 0.035. The coefficient of variation reached 0.022, both higher than the normal average levels of 0.012 and 0.009. The correlation between flame morphology and low oxygen levels increased to 0.61. After weighted fusion, the calibrated intermediate layer risk feature score rose to 0.78, and the system marked it as a "key monitoring area". Simultaneously, the model jointly weighted abnormal flame brightness and excessive temperature, establishing a feature adjustment mapping to determine that the overall risk level at this point was high.

[0148] The system initiates high-level anomaly probability distribution inference and outputs anomaly probability distribution curves for the current feature window. In this example, the probability of coking risk increases from 0.10 in the previous period to 0.23, the probability of flame elongation and shift increases to 0.19, and the probability of low oxygen content triggering anomalies reaches 0.21. The system automatically stratifies and partitions these three indicators, classifying the current time period of 06:25-06:30 into the anomaly priority attention zone. At this time, based on flame morphology similarity analysis, the similarity score with historical coking precursor cases reaches 88%. In the anomaly triggering condition identification module, the system outputs: "Temperature continuously above 1200℃, oxygen content below 1.7%, flame elongation code of 6, brightness principal component peak increase of more than 8%", which are typical coking anomaly precursors.

[0149] The system employs a causal inference algorithm to analyze the interaction path between temperature, oxygen content, and flame images. Model training shows that in 143 historical actual coking events, 96% exhibited a sequential change of rapid temperature rise – oxygen content drop – abrupt flame morphology change. Through Bayesian network structure learning, the system determined the causal path probability strength of the current monitoring data to be 0.74, far exceeding the upper limit of the safe range of 0.55. The model automatically identified "temperature rise leading to oxygen content drop, and flame anomalies caused by the superposition of local hypoxia and temperature rise," initially tracing the cause as local fluctuations in pulverized coal supply and uneven distribution of air volume.

[0150] During the real-time root cause localization phase, the system compared all high-risk cases in the local database, extracted common parameters, and found that the particle size distribution of pulverized coal supply device B deviated from the normal value during this anomaly. Its average particle size suddenly increased from 78μm to 91μm, the standard deviation increased to 17.6μm, and the proportion of fine pulverized coal particles decreased by 9%. After considering all risk factors, the system output an anomaly report: "High temperature + low oxygen + flame elongation + pulverized coal coarsening = high probability of coking," and automatically suggested inspecting the pulverized coal supply device, appropriately increasing the air supply volume, and adjusting the distribution ratio.

[0151] Meanwhile, within 16 seconds of detecting the anomaly, the system automatically generated a Level II risk warning signal and directly issued a draft adjustment command through the control module: "Main blower airflow +5%, coal powder B supply reduced by 2%, recalibrate particle screening." Upon receiving the alarm, the inspection personnel responded quickly and discovered a slight blockage in the coal powder silo screen. After timely clearing, the monitoring data showed a significant drop. Five minutes later... The regional temperature dropped to 1194℃. The oxygen level recovered to 1.84%, the flame pattern code returned to 4, the brightness distribution became more uniform, and the system automatically deactivated the warning.

[0152] To fully verify the practical effectiveness of this invention, both the traditional alarm algorithm and the algorithm of this invention were run concurrently, and their early warning performance was compared using real-world data from 30 consecutive days. The traditional method uses temperature >1200℃ or oxygen content <1.7% as separate alarms, resulting in a false alarm rate of 14.2%, a false alarm rate of 16.9%, and an average early warning lead time of only 2.3 minutes. Most coking events only triggered alarms after the anomaly had already occurred or was about to occur, leading to delayed inspection responses; coking actually occurred 10 times.

[0153] The method of this invention, through multi-parameter fusion and causal tracing, reduces the false alarm rate to 3.6%, the false alarm rate to 4.7%, and the average early warning lead time to 7.6 minutes. Only one coking event actually occurred, and a sufficient early warning was given. Typical comparison samples are as follows:

[0154] During the morning shift on the 11th day, The temperature rose from 1178℃ to 1212℃ between 06:20 and 06:35. The percentage of coal dust decreased from 2.12% to 1.66%, and the flame code changed from 3 to 6. Using the traditional method, an alarm was only issued at 06:33, with a warning lead time of only 1 minute. This meant that the coal powder supply could not be adjusted in time, leading to furnace coking and requiring a temporary load reduction. Using the method of this invention, all abnormal indicators were identified and triggered by 06:27, and timely inspections addressed the issues. No coking occurred in the furnace, and the equipment operated smoothly.

[0155] During the 23rd night shift, Oxygen levels remained between 2.08% and 1.71%. The temperature fluctuated between 1174℃ and 1195℃, the flame morphology was normal, and all characteristics were within the normal fluctuation range. Traditional methods, due to one... A sudden drop to 1.69% triggered a false alarm, leading to repeated inspections by on-site personnel that found no abnormalities. When using this invention, the system comprehensively considers multiple parameters and trends to determine that the fluctuation is normal, resulting in no alarm, effectively avoiding false alarms and improving operational efficiency.

[0156] By comparing and statistically analyzing data from all 52 high-risk monitoring cases, the method of this invention outperforms traditional solutions in key indicators such as automatic alarm, root cause analysis, response timeliness, and control of false alarms and missed alarms. Each anomaly is accompanied by detailed feature sequence analysis, probability inference, and causal link tracing. Training samples are shown below:

[0157] Sample A: =1199℃, =1.83%, flame code 5, coal powder particle size 88μm, alarm by traditional method, alarm by this invention, no actual coking;

[0158] Sample B: =1215℃, =1.62%, flame code 6, coal powder particle size 92μm, traditional alarm not reported, this invention alarm, actual coking;

[0159] Sample C: =1170℃, =2.17%, flame code 4, coal powder particle size 78μm, neither method triggered an alarm, actually normal;

[0160] Sample D: =1203℃, =1.74%, flame code 6, coal powder particle size 87μm, traditional alarm not reported, this invention alarm, actual coking;

[0161] Sample E: =1159℃, =2.08%, flame code 4, coal powder particle size 75μm, neither method triggered an alarm, indicating normal operation.

[0162] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A boiler combustion early warning method based on an improved deep forest algorithm, characterized in that, include: Collect temperature data, oxygen data, and flame image data inside the combustion chamber of the combustion equipment, extract the low-level feature vectors of the temperature data, oxygen data, and flame image data, and obtain the initial quantitative indicators of the combustion state of the combustion equipment. A time series analysis model was used to process the changing trends of the initial quantitative indicators and to determine the dynamic fluctuation pattern of the combustion process of the combustion equipment under multi-parameter coupling conditions. An improved deep forest algorithm was used to construct a multi-level feature evaluation structure. The combustion features of each layer were analyzed by initial quantification indicators to obtain the confidence score of the combustion features. If the confidence score of the combustion feature is lower than the preset threshold, the influencing factors are calibrated based on the weights of the underlying feature vector through a weighted fusion mechanism to obtain the adjusted intermediate layer risk features. Based on the adjusted intermediate layer risk characteristics, the probability distribution of high-level anomalies in the combustion process of the combustion equipment is obtained to determine the triggering conditions for potential coking risks. By using the triggering conditions of the potential coking risk, a causal analysis model is employed to trace the interaction paths between multiple parameters and determine the root cause of combustion abnormalities in the combustion equipment. A real-time early warning signal is generated based on the root cause of the combustion abnormality in the combustion equipment.

2. The boiler combustion early warning method based on the improved deep forest algorithm according to claim 1, characterized in that, The extracted low-level feature vectors of temperature data, oxygen data, and flame image data include: Temperature and oxygen levels inside the furnace are collected in real time using temperature and oxygen sensors, while flame image data is acquired using an image sensor. The collected temperature data, oxygen data, and flame image data are processed by feature extraction to generate corresponding feature vectors and determine the underlying characterization information of the combustion state. Based on the generated feature vector, the changing trends of temperature and oxygen data are analyzed. If the changing trend exceeds the preset threshold range, an abnormal state indicator is triggered to determine whether there is a potential combustion instability. By performing pattern comparison on the feature vectors extracted from flame image data and combining them with a pre-established flame morphology database, the morphology category that best matches the current combustion state is obtained, and the visual representation result of the combustion process is determined. A comprehensive analysis is performed on the underlying characterization information and visual characterization results of the combustion state. If the correlation between temperature data and oxygen data is lower than a preset threshold, the feature vector is weighted and adjusted to obtain the combustion state assessment result. Based on the combustion status assessment results and combined with historical data of equipment status, the operational stability inside the furnace is analyzed, the status change characteristics under long-term operating trends are obtained, and a complete set of initial quantitative indicators are constructed.

3. The boiler combustion early warning method based on the improved deep forest algorithm according to claim 1, characterized in that, The determination of the dynamic fluctuation mode of the combustion process of the combustion device under multi-parameter coupling conditions includes: The initial data is segmented based on the combustion status assessment results of the combustion equipment, and the changing trends within different time windows are extracted to obtain the stage fluctuation information of the combustion process. Based on the phased fluctuation information, and combined with the coupled scenario of multiple parameter groups, a correlation analysis was conducted on multiple variables in the combustion process to determine the main driving factors of dynamic fluctuations. By extracting the main driving factors of dynamic fluctuations, a framework for identifying the regularity of the combustion process is constructed to obtain the response patterns of the combustion state under different coupled scenarios. If the response mode is within the preset threshold range, the dynamic fluctuations of the combustion process will be continuously tracked to determine whether there are any abnormal deviations. If the data exceeds the threshold range, the fluctuation data of the current time series is recorded to obtain the anomaly labeling result; Based on the anomaly marking results, a localized analysis is performed on the combustion status assessment results of the combustion equipment. The current dynamic fluctuations are compared with a pre-established reference database to identify potential sources of fluctuations. By analyzing the potential sources of fluctuations and combining the changing trends of the time series, a dynamic adjustment basis for the combustion process is generated, and optimization reference information for multi-parameter coupling scenarios is obtained. Based on the optimized reference information, the operating data of the combustion equipment is continuously collected and compared to determine whether the combustion state tends to be stable, thus obtaining the final dynamic fluctuation mode.

4. The boiler combustion early warning method based on the improved deep forest algorithm according to claim 1, characterized in that, The method of constructing a multi-level feature evaluation structure using an improved deep forest algorithm includes: The initial quantitative indicators of the combustion state of the combustion equipment are obtained, and the dynamic fluctuations are collected in real time. Various parameters during the combustion process are recorded by sensors to obtain the raw fluctuation data set. The original fluctuation data set is denoised and standardized, and the basic form of the fluctuation pattern is determined by extracting dynamic fluctuation segments in segments. Based on the basic morphology of the wave pattern, a multi-layer feature extraction structure is constructed, and the deep forest algorithm is used to decompose and represent the combustion features of each layer to obtain a hierarchical feature set. For the hierarchical feature set, a layer-by-layer analysis process is implemented. If the significance of a feature in a certain layer is lower than a preset threshold, its weight is adjusted to determine the preliminary importance ranking of each layer of features. Based on the initial importance ranking, the confidence score of each combustion feature is calculated, and the scores of each feature are integrated by weighted accumulation to determine the final confidence score. If the final credibility score deviates from the preset range, a feature re-evaluation mechanism is triggered. This mechanism performs local corrections by backtracking the hierarchical feature set, resulting in an adjusted credibility result.

5. The boiler combustion early warning method based on the improved deep forest algorithm according to claim 1, characterized in that, The adjusted intermediate layer risk characteristics include: If the confidence score of the combustion feature is lower than the preset threshold, the preliminary structure of the weight distribution is determined by analyzing the weights of the underlying feature vectors one by one. If the initial structure of the weight distribution deviates from the preset threshold, the vector weights are calibrated through a weighted fusion mechanism to obtain the calibrated weight combination. Based on the calibrated weight combination, the risk characteristics of the intermediate level are calculated, their correlation with combustion characteristics is analyzed, and a quantitative representation of the risk characteristics is obtained. By quantifying risk characteristics, a mapping logic for feature adjustment is constructed to determine whether the adjusted features meet expectations and obtain the final correction result. Based on the final correction results, intermediate layer risk characteristics of combustion features are generated to identify potential risk points in the combustion process.

6. The boiler combustion early warning method based on the improved deep forest algorithm according to claim 1, characterized in that, The method of obtaining the high-level anomaly probability distribution of the combustion process of the combustion equipment and determining the triggering conditions for potential coking risks includes: Based on the adjusted risk characteristics of the intermediate layer, the probability distribution of high-rise anomalies is obtained by real-time data collection during the operation of the combustion equipment, thereby determining the range of possible anomalies. Based on the probability distribution of high-level anomalies, a preset threshold is used for screening. If the probability distribution exceeds the threshold range, the abnormal data is stratified to obtain the key attention interval for high-level anomalies. For key areas of concern, extract risk feature data related to intermediate layer risk characteristics, determine the degree of correlation between potential coking and high-level anomalies, and obtain the priority ranking of risk features; By prioritizing and analyzing the triggering conditions for potential coking risks, and combining this with real-time changes in equipment status, the specific manifestations of the triggering conditions are determined.

7. The boiler combustion early warning method based on the improved deep forest algorithm according to claim 1, characterized in that, The method of using a causal analysis model to trace the interaction paths between multiple parameters and determine the root cause of combustion abnormalities in the combustion equipment includes: By analyzing the triggering conditions of the potential coking risk, the logical relationships between parameter changes are sorted out using causal analysis methods, and the interaction paths between multiple parameters are obtained. Based on the interaction paths between multiple parameters, the direct correlation between combustion anomalies and changes in each parameter is analyzed to determine the preliminary root cause of the combustion anomalies. Based on the initial root cause location, real-time monitoring data of the equipment status is obtained, and combined with dynamic information on parameter changes, the key influencing factors for anomaly determination are identified. By identifying key influencing factors, a framework for assessing coking risk is constructed. If parameter changes exceed a preset threshold range, the anomaly detection is prioritized to determine the key areas for risk assessment. Based on the key directions of risk assessment, specific data on the triggering conditions of potential coking risks are extracted, the abnormal probability of combustion equipment under different operating scenarios is analyzed, the distribution range of the triggering conditions of potential coking risks is determined, and the root cause of combustion abnormalities in combustion equipment is identified.

8. The boiler combustion early warning method based on the improved deep forest algorithm according to claim 1, characterized in that, The step of generating a real-time early warning signal based on the root cause of the combustion abnormality in the combustion device includes: To pinpoint the root cause of combustion abnormalities in combustion equipment, historical operating records related to the root cause of the combustion abnormality are obtained. By comparing the deviation between current data and historical data, the specific category of the root cause of the abnormality is determined. From the specific categories of the anomaly root cause, obtain the corresponding impact range data, classify the impact range, and determine the priority of real-time early warning and the signal generation method; Based on the priority of real-time warnings, corresponding warning signals are generated and transmitted to the control module through the internal system channel to obtain the signal-triggered response mechanism. Regarding the response mechanism, adjustment instructions for the combustion process are generated by combining optimization methods, and then sent to the combustion equipment through the instruction execution module to complete process monitoring and status updates.

Images