Online partition quantitative evaluation and ash removal decision-making method and system for ash deposition of air pre-heater

Abstract

The invention discloses an air pre-heater ash deposition online partition quantitative evaluation and ash removal decision method and system, and relates to the technical field of industrial artificial intelligence and machine learning, and the method comprises the steps: collecting multi-dimensional historical operation data covering all working conditions when an air pre-heater is in a reference clean state, and training and optimizing an initial health prediction model; performing real-time prediction through the initial health prediction model to obtain health reference data; automatically activating the initial health prediction model to autonomously evolve according to a preset triggering condition; and re-determining the health reference data through the health prediction model obtained through autonomous evolution. The problem of model drift is solved, and the long-term effectiveness of the diagnosis method is ensured. According to the invention, the manual maintenance cost is reduced, and the self-maintenance of the model is realized. And a complete intelligent closed loop of diagnosis, ash removal, model optimization and re-diagnosis is formed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial artificial intelligence and machine learning, and more particularly, to an online partition quantitative evaluation and ash removal decision method and system for air preheater ash deposition. BACKGROUND

[0002] Digital twin technology has been tried for state prediction of industrial equipment. However, the core challenge is that the accuracy of the model will decrease as the performance of the equipment itself drifts (such as the performance degradation of the air preheater due to ash deposition). A digital twin model trained based on the initial clean state will have a significantly reduced prediction accuracy after running for a period of time, resulting in diagnostic failure.

[0003] In the prior art, although there are methods of using digital twins for air preheater state prediction, most still use flue gas side pressure difference (ΔP) as the core diagnostic feature. The pressure difference change is a manifestation of the late flow resistance effect of ash deposition, and the diagnosis has serious hysteresis. In addition, these methods lack a core index that can accurately quantify the severity of ash deposition and is directly related to the advanced physical phenomenon (thermal resistance effect), and cannot achieve real advanced warning and long-term effective and accurate diagnosis.

[0004] The accuracy of the digital twin model depends heavily on the state of the equipment represented by its training data. The air preheater is a constantly changing system (such as catalyst aging, heat exchange element corrosion, and ash deposition characteristics change), and its operating characteristics will slowly drift. A model trained based on initial clean state data will have a sharp decrease in prediction accuracy after running for a period of time, resulting in gradual failure of diagnosis based on it. The prior art lacks an automated and systematic process to adapt the model to such changes, usually requiring manual intervention, re-collecting data, and re-training, which is costly and inefficient.

[0005] Therefore, how to propose an online partition quantitative evaluation and ash removal decision method and system for air preheater ash deposition, taking the heat exchange efficiency deviation of the logical control loop as the core index, to realize online, partitioned, and quantitative evaluation of the severity of air preheater ash deposition, and to automatically generate ash removal decision instructions based on the accurate evaluation, while automatically maintaining and continuously optimizing the prediction accuracy of the air preheater health prediction model during use, is a problem that those skilled in the art need to solve. SUMMARY

[0006] In view of this, the present application provides an air preheater ash deposition online partition quantitative evaluation and ash removal decision method and system, a health prediction model step, based on the historical operation data of the air preheater in the benchmark clean state, a model for predicting the ideal heat exchange efficiency benchmark value of each logical control ring is constructed; the health prediction model is not limited to a specific implementation form, and its core is to accurately represent the performance benchmark of the clean air preheater under different working conditions; a real-time benchmark prediction step, receiving real-time working condition parameters, outputting the ideal heat exchange efficiency benchmark value under the current working condition; an online quantitative evaluation step, calculating the efficiency deviation according to the ideal benchmark value and the actual efficiency, and outputting the ash deposition degree coefficient y(i, t) online based on the evaluation model; an ash removal decision step, comparing the ash deposition degree coefficient with the preset threshold, and generating a hierarchical early warning and ash removal decision instruction.

[0007] The present application provides an air preheater ash deposition online partition quantitative evaluation and ash removal decision method and system, which solves the model drift problem of the digital twin model after long-term operation, and ensures the continuous output of high-precision logical control ring ash deposition degree quantitative evaluation results through automatic model self-evolution management, and generates accurate ash removal decision instructions based on this, thereby guaranteeing the long-term effectiveness of the intelligent diagnosis system based on the model. In order to achieve the above-mentioned purpose, the technical scheme of the present application introduces a unique anti-interference feature engineering and a quantitative calibration method based on physical limits, which fundamentally improves the robustness and accuracy of the diagnosis. The present application adopts the following technical scheme: An air preheater ash deposition online partition quantitative evaluation and ash removal decision method, comprising: When the air preheater is in a benchmark clean state, multi-dimensional historical operation data covering all working conditions are collected, and an initial health prediction model is trained and optimized; Real-time prediction is performed through the initial health prediction model to obtain health benchmark data; The initial health prediction model is automatically activated for self-evolution according to a preset trigger condition; The health prediction model obtained through self-evolution is used to re-determine the health benchmark data.

[0008] The present application also provides an air preheater ash deposition online partition quantitative evaluation and ash removal decision system, which serves as the physical carrier and intelligent decision center of the method, comprising: A data acquisition and preprocessing module for collecting historical and real-time operation data of the air preheater, and performing data cleaning, outlier processing, data alignment, standardization and generating derived features to suppress working condition fluctuations; A digital twin model construction and management module for constructing, training, verifying, deploying and updating the initial health prediction model and the health prediction model after self-evolution; a real-time prediction and health benchmarking module configured to load the health prediction model and predict an ideal heat exchange efficiency benchmark value of each logical control loop according to real-time working condition parameters; an online quantitative evaluation module configured to calculate an actual heat exchange efficiency and an efficiency deviation of each logical control loop according to the ideal heat exchange efficiency benchmark value and measured data, and to calculate and output a logical control loop soot accumulation degree coefficient y(i, t) in real time based on a multi-parameter fusion model; a soot cleaning decision module configured to compare the logical control loop soot accumulation degree coefficient y(i, t) with a preset threshold, and generate a hierarchical early warning signal and a soot cleaning control instruction for a target logical control loop; a model self-evolution management module configured to continuously monitor system states, automatically activate a model self-evolution process according to a preset trigger condition, and manage retraining, verification and seamless switching of the model; a communication and control module configured to send the soot cleaning decision instruction to a soot cleaning execution system to drive the soot cleaning execution system to perform precise soot cleaning operation on the specified logical control loop.

[0009] The system realizes online, partitioned and quantitative evaluation of the soot accumulation state of the air preheater through the cooperative work of the above-mentioned modules, and forms a complete intelligent closed loop from diagnosis, decision to execution.

[0010] Optionally, to suppress the interference of working condition fluctuations on the core diagnostic indicators, specific derived features are introduced as inputs in model training and real-time prediction, including but not limited to: specific efficiency (η / Q²), which is used to eliminate the influence of medium flow changes on heat exchange efficiency calculation; and / or unit load efficiency (η / W), which is used to eliminate the influence of unit load changes on heat exchange efficiency calculation.

[0011] Optionally, the 0-100% scale of the logical control loop soot accumulation degree coefficient y(i, t) is defined based on one or more non-crossable physical operating limits, wherein 100% corresponds to any one or a combination of the following critical states: the pressure difference on the flue gas side of the logical control loop reaches the maximum output limit of the induced draft fan; the heat exchange efficiency of the logical control loop decreases, causing the exhaust gas temperature to exceed the environmental upper limit or the hot air temperature to be lower than the safe operation lower limit; the soot cleaning system cannot effectively remove the soot under maximum intensity operation. This kind of calibration method ensures the physical explicitness and engineering guidance value of the coefficient.

[0012] Optionally, the data includes: unit load, flue gas flow, air flow, ambient temperature, rotor speed working condition parameters.

[0013] Optionally, it further includes data cleaning and preprocessing, missing value processing, outlier removal, data alignment and standardization.

[0014] Optionally, the training comprises: collecting current working condition parameters in real time, and performing real-time prediction through the initial health prediction model to obtain ideal heat exchange efficiency reference values of each logical control loop under the current working condition.

[0015] Optionally, the model performance standard requirement is: R 2 >0.98, and MAE<1.5℃.

[0016] Optionally, the automatic activation of the initial health prediction model self-evolution according to the preset trigger condition comprises: confirming that the air preheater is in a clean state again; when the trigger condition is met, the system automatically collects a new round of health reference data, and merges the new data with the original historical health reference data; retraining or incrementally training the initial health prediction model using the merged data; verifying the performance indicators of the new model, and automatically replacing the online old model seamlessly after confirming that the standard is met, to complete a self-evolution cycle.

[0017] Optionally, the clean state comprises: after the unit is subjected to planned maintenance and high-pressure water flushing are completed, manual confirmation by an operation and maintenance personnel; or automatic judgment by the system according to a preset condition.

[0018] Optionally, the automatic judgment by the system according to a preset condition comprises: within M hours, M≥4, simultaneously satisfying: the unit load fluctuation is less than ±N%; the absolute value of the deviation of the actual efficiency of all logical control loops from the model prediction value is continuously lower than a threshold X1%; and the soot accumulation coefficient of all logical control loops is continuously lower than a threshold Y%; within M hours, M≥2, the absolute value of the efficiency deviation of all logical control loops is continuously lower than a threshold X2%.

[0019] Optionally, the retraining or incrementally training the initial health prediction model using the merged data comprises: using an incremental learning or full-amount retraining algorithm, and automatically selecting a strategy according to the amount of new data and the performance drift degree.

[0020] Optionally, the performance drift degree is evaluated by introducing a quantitative indicator based on model prediction uncertainty: during the clean state determination, the confidence interval of the model prediction value of a key output (such as η_ref(i,t)) is calculated synchronously; if the overlap degree of the historical confidence interval and the current confidence interval is continuously lower than a preset threshold, or the prediction uncertainty (such as variance) significantly increases, it is determined that the performance drift is significant, and the full-amount retraining strategy is preferentially triggered.

[0021] Optionally, the model self-evolution process is embedded with an adversarial sample detection module: before real-time prediction data flows into the model, distribution anomaly detection and adversarial disturbance identification are performed; if suspected malicious interference data is detected, the data flow is automatically isolated, a safety alarm is triggered, and the current evolution process is suspended until the data safety is confirmed.

[0022] Through the above technical solutions, compared with the prior art, the application provides an online partition quantitative evaluation and ash removal decision method and system for air preheater ash deposition, which has the following beneficial effects: The application provides an online partition quantitative evaluation and ash removal decision method for air preheater ash deposition, which comprises the following steps: collecting multi-dimensional historical operation data covering all working conditions when the air preheater is in a reference clean state, training and optimizing an initial health prediction model; obtaining health reference data through real-time prediction by the initial health prediction model; automatically activating the initial health prediction model self-evolution according to a preset trigger condition; and re-determining the health reference data by the health prediction model obtained through self-evolution. The application solves the model drift problem and ensures the long-term effectiveness of the diagnosis method. The application reduces the artificial maintenance cost and realizes self-maintenance of the model. A complete intelligent closed loop of diagnosis-ash removal-model optimization-re-diagnosis is formed.

[0023] On this basis, the application further has the following outstanding advantages: 1. Extremely advanced: taking the heat transfer efficiency deviation as the core index, which is a direct manifestation of thermal resistance effect, the early capture of ash deposition germination is realized.

[0024] 2. Precise quantification: the concept of 0-100% logic control ring ash deposition degree coefficient y(i, t) is created, and precise, quantitative and visual diagnosis of the severity of ash deposition is realized.

[0025] 3. Long-term reliable evaluation results: through the model self-evolution process, it is ensured that the digital twin model can be updated adaptively following the performance drift of the equipment, and the accuracy of the quantitative evaluation results in long-term operation is ensured. The evolution process introduces a performance drift evaluation mechanism based on prediction uncertainty, making the selection of evolution strategy more intelligent and forward-looking.

[0026] 4. Strong anti-interference ability: through multi-parameter fusion and derived features (such as η / Q²), the disturbance of working condition fluctuations is effectively weakened or eliminated.

[0027] 5. High system safety: the built-in data security and adversarial sample detection mechanism effectively prevents model mis-evolution or diagnosis failure caused by data anomalies or potential malicious attacks, and improves the robustness and reliability of the industrial AI system.

[0028] 6. Form a complete intelligent decision-making closed loop: the application as an intelligent brain and decision-making core, responsible for online, quantitative evaluation of ash deposition area and severity, and generate a strategy instruction containing target logic control ring and ash removal temperature; the sealing system as its execution trunk, responsible for receiving instructions and completing precise ash removal operation. The two communicate through a data interface, and the ash deposition degree coefficient y(i, t) is one of the most important communication instructions. BRIEF DESCRIPTION OF DRAWINGS

[0029] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, below will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only a part of the embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor based on the provided drawings.

[0030] Figure 1 The model side system structure and data flow chart provided by the present application.

[0031] Figure 2 The off-line training and on-line application flow chart of the digital twin model provided by the present application.

[0032] Figure 3 The self-evolution trigger condition judgment and execution flow chart provided by the present application. DETAILED DESCRIPTION

[0033] The technical solutions in the embodiments of the present application will be described clearly and completely below with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0034] In order to make the purpose, technical solutions and advantages of the present application more clear and obvious, the following first defines several core terms involved in the present application: Annular control area: refers to the concentric annular physical area formed by radial partition in the air preheater duct, which can be independently controlled in terms of air volume and pressure. Each annular control area is provided with independent inlet damper, outlet damper and hot sealing air injection interface.

[0035] Logic control ring: refers to the annular sector on the air preheater rotor corresponding to the annular control area in space. Ash deposition occurs on the continuously rotating logic control ring, while ash removal operation is applied to the static annular control area to affect the logic control ring passing through it.

[0036] The logic control ring dusting degree coefficient y(i, t) is a dimensionless quantitative index, and its value range is 0% to 100%, which is used to represent the dusting severity of the specific logic control ring i at time t. Wherein, 0% represents that the ring is in an absolutely clean state, and 100% represents that the ring dusting has reached one or more insurmountable physical operation limits (such as the maximum output of the induced draft fan, the upper limit of the environmental protection smoke temperature, etc.).

[0037] The partitioned targeted dusting of the application refers to the independent operation of each logic control ring and the corresponding ring-shaped control area.

[0038] The application is based on a brand-new and fundamental inventive concept: in order to realize the accurate prediction and removal of the air preheater dusting, it is necessary to decouple the diagnosis object (logic control ring) of the dusting state and the execution object (ring-shaped control area) of the dusting control in concept, and to perform collaborative mapping and linkage in the system.

[0039] The core of the application is that an intelligent system establishes an accurate mapping and control link from the target logic control ring that needs to be dusted to the corresponding target ring-shaped control area. The system first diagnoses the target logic control ring online and quantitatively, and then drives the corresponding target ring-shaped control area to perform collaborative operation, and realizes accurate and efficient online dusting of the target logic control ring passing through the area by establishing a micro-positive pressure high-temperature environment.

[0040] This inventive concept fundamentally solves the inherent defects of the traditional method, such as intervention lag, insufficient precision and energy waste caused by the confusion of diagnosis and control objects, and lays a brand-new technical architecture for realizing the predictive intelligent maintenance of the air preheater.

[0041] Based on the aforementioned inventive concept, The embodiment of the application discloses an online partitioned quantitative evaluation and dusting decision method for air preheater dusting, and the core process of the method is as shown in the figure. Figure 1 The main steps include the following steps: establishing a health prediction model step, a real-time baseline prediction step, an online quantitative evaluation step, and a dusting decision step, and specifically: The embodiment of the application discloses an online partitioned quantitative evaluation and dusting decision method for air preheater dusting, which comprises: When the air preheater is in a baseline clean state, collect multi-dimensional historical operation data covering all working conditions, train and optimize the initial health prediction model; The reference clean state or the clean state again is not a vague empirical concept, but a technical identification by a set of quantitative criterion system. The system continuously monitors the load fluctuation rate of the unit, the absolute value and persistence of the efficiency deviation of each logic control loop, and the calculated dusting degree coefficient and other physical parameters in multiple dimensions. Only when these parameters simultaneously meet the preset, strict threshold conditions and continuously reach the specified duration, or after physical intervention such as overhaul and water washing, and the state is forcibly identified by the operation and maintenance personnel in the system, can it be confirmed as a clean state. The criterion system itself is an important technical guarantee to ensure the long-term diagnostic accuracy of the present application. Real-time prediction is performed through the initial health prediction model to obtain health reference data; The initial health prediction model is automatically activated according to the preset trigger condition to evolve autonomously; The health prediction model obtained through autonomous evolution is used to re-determine the health reference data.

[0042] Specifically, the partition or specific partition (i) referred to in the present application is a "ring control zone" designed to achieve accurate dust removal. The air preheater is divided into multiple such ring control zones. The ring control zone is formed by radial partitions installed in the air preheater duct and is a plurality of independent concentric ring-shaped physical structures. Each ring control zone is provided with an independent inlet damper, outlet damper and hot air injection interface, and can independently control the air volume, pressure and temperature, and corresponds to a specific ring sector (or logic control loop) on the rotor body. In short, the ring control zone is an independently controllable static physical structure in the duct, while the logic control loop is a dynamic rotating area corresponding to it on the rotor. The present application independently evaluates the dust accumulation of each ring control zone and identifies the target ring control zone that needs intervention. The core purpose of the present application is to provide an online quantitative diagnosis method for the dust accumulation severity of each ring control zone.

[0043] In addition, the health prediction model is not limited to data-driven machine learning models. The core of the present application is to obtain a high-precision ideal efficiency reference, rather than being bound by specific modeling methods. In another embodiment of the present application, a high-precision physical mechanism model can also be established based on the heat transfer mechanism of the air preheater, with operating parameters as input and ideal heat exchange efficiency as output, or a multi-dimensional operating condition-efficiency query table can be constructed through massive historical health data, and the ideal efficiency reference value can be determined by interpolation calculation. These alternative solutions also fall within the scope of the health prediction model described in the present application.

[0044] The high-precision accumulated dust degree coefficient y(i, t) calculated by the application is the core decision input of the precise dust cleaning concept of the logical control loop, and can be directly used to drive the downstream targeted dust cleaning control system. The dust cleaning control system (i.e. the execution system for the annular control zone) generally includes the radial partition plate and the damper, the mechanical plugging device arranged in the rotor compartment for suppressing axial air leakage, and the hot sealing air injection and intelligent control system. Through the cooperative system, a high-temperature sealed micro-positive pressure environment can be constructed for a specific annular control zone of the air preheater, so as to realize precise online dust cleaning. Thus, the application provides the most critical decision basis for the entire predictive intelligent dust cleaning closed loop.

[0045] In the specific embodiment, an online partition quantitative evaluation and dust cleaning decision method for air preheater dust accumulation, the core of which is to realize online partition quantitative evaluation and dust cleaning decision of air preheater dust accumulation, the process of which starts from the construction of a high-precision digital twin model and ends with the issuance of precise action instructions to the dust cleaning execution system, and the specific steps are as follows: S1: initial health prediction model construction When the air preheater is confirmed to be in a baseline clean state, a multi-dimensional historical operation data set covering all operating conditions is collected.

[0046] Specifically, the data set includes operating condition parameters such as unit load, flue gas flow, air flow, ambient temperature, rotor speed, etc. The data set is subjected to data cleaning and preprocessing, including missing value processing, outlier removal, data alignment and standardization. An input feature set is constructed, and η / Q² (specific efficiency, representing the heat exchange efficiency per unit flow, used to effectively suppress the disturbance of flow fluctuations on efficiency calculation) and η / W (unit load efficiency, representing the efficiency performance at the unit load level, used to suppress the influence of load changes) derived features are introduced to suppress the disturbance of operating condition fluctuations. The ideal heat exchange efficiency reference value η_ref(i) of each logical control loop (i) is taken as the output target, and an initial health prediction model for predicting the ideal heat exchange efficiency reference value is obtained by training and optimization using a machine learning algorithm; the model performance needs to meet the requirements of a test set determination coefficient R²>0.98 and an average absolute error MAE<1.5%; the model is a first-level model, and its role is to accurately depict the performance of a clean air preheater under different operating conditions.

[0047] Further, in the construction of the input feature set in S1 and the construction of the fusion model feature vector in S5, in addition to instantaneous values, statistical features (such as mean, standard deviation, slope) of key parameters (such as Δη(i, t), L(t), F_gas(t)) in a recent time window (such as the past 5-15 minutes) can also be introduced to capture the dynamic trend of the dust accumulation process, further improving the robustness and forwardness of the diagnosis.

[0048] S2: Real-time prediction of ideal efficiency reference value

[0049] In the online application stage, the current working condition parameters are collected in real time, and real-time prediction is performed through the digital twin model to obtain the ideal heat exchange efficiency reference value η_ref(i, t) of each logical control ring under the current working condition.

[0050] S3: Real-time calculation of actual heat exchange efficiency

[0051] Real-time collection of temperature sensor data in each ring-shaped control area, and calculation of actual heat exchange efficiency of each logical control ring according to the formula: η_actual(i, t) = [T_air_out(i, t) - T_air_in(i, t)] / [T_gas_in(i, t) - T_air_in(i, t)]

[0052] S4: Efficiency deviation calculation

[0053] Calculate the deviation of real-time heat exchange efficiency of each logical control ring from the model predicted reference value: Δη(i, t) = η_actual(i, t) - η_ref(i, t).

[0054] S5: Online partition quantitative evaluation - soot degree coefficient calculation and output

[0055] The efficiency deviation Δη(i, t) is taken as the core feature, and auxiliary parameters such as the instantaneous value and change trend of unit load L(t) and flue gas flow F_gas(t) are input into a pre-trained multi-parameter fusion regression model to calculate a dimensionless logical control ring soot degree coefficient y(i, t) in the range of 0% (absolute cleanliness) to 100% (severe blockage).

[0056] In the specific implementation, the specific calculation process of the logical control ring soot degree coefficient y(i, t) includes: (1) First-level model (health prediction model): Function: Predict the ideal heat exchange efficiency reference value η_ref(i, t) that an absolutely clean air preheater logical control ring (i) should achieve under the current working condition.

[0057] Objective: Establish a pure performance reference that is not affected by soot.

[0058] (2) Second-level model (multi-parameter fusion diagnosis model): Input: Take the efficiency deviation Δη(i, t) calculated by the first-level model as the core feature, and fuse other working condition parameters.

[0059] Function: Learn the complex nonlinear mapping relationship from features such as efficiency deviation to the severity of fouling, and output the final y(i, t).

[0060] Objective: Convert the physical performance deviation (Δη) into an intuitive and operable index of fouling severity.

[0061] Mathematical details and key formulas

[0062] 1) Calculation of actual heat exchange efficiency (S3)

[0063] The formula is: η_actual(i, t) = [T_air_out(i, t) - T_air_in(i, t)] / [T_gas_in(i, t) - T_air_in(i, t)]; 2) Calculation of efficiency deviation (S4) The formula is: Δη(i, t) = η_actual(i, t) - η_ref(i, t).

[0064] (3) Feature engineering: Derivative features (suppressing working condition fluctuation interference)

[0065] η / Q² (specific efficiency): Represents the heat exchange efficiency under unit flow rate square, used to suppress the interference brought by the fluctuation of flow rate on efficiency calculation.

[0066] η / W (unit load efficiency): Represents the efficiency performance under the unit load level, used to suppress the influence of load change.

[0067] Statistical features: In addition to instantaneous values, the mean, standard deviation, and slope of key parameters (such as Δη(i, t), L(t), F_gas(t)) within a recent time window (such as the past 5-15 minutes) can be introduced to capture the dynamic trend of the fouling process.

[0068] (4) Generation of fouling degree coefficient y(i, t) (S5)

[0069] y(i, t) = SVR_Model([Δη(i, t), L(t), F_gas(t), …]).

[0070] The above is a support vector regression (SVR) model, whose input is a feature vector, which includes: Core features: Efficiency deviation Δη(i, t); Auxiliary working condition features: Unit load L(t), flue gas flow F_gas(t), etc. Dynamic trend features: Recent statistical features of the above parameters; The model outputs a continuous value ranging from 0% to 100%.

[0071] Specifically, the model training process

[0072] Training of the first-level model (GBDT) (S1)

[0073] (1) Data source: Collect 1-3 months of historical operation data covering all working conditions (from the lowest load to the rated load) when the air preheater is confirmed to be in the baseline clean state.

[0074] (2) Input features (X): Working condition parameters such as unit load, flue gas flow, air flow, ambient temperature, rotor speed, and derived features η / Q², η / W.

[0075] (3) Output target (Y): Ideal heat exchange efficiency reference value η_ref (i) of each logical control loop (i).

[0076] (4) Algorithm: Gradient Boosting Decision Tree (GBDT) algorithm is used.

[0077] (5) Hyperparameter optimization: determined by five-fold cross-validation and Bayesian optimization method, the range is usually: n_estimators: 100-500; learning_rate: 0.01-0.1; max_depth: 3-6.

[0078] (6) Performance requirement: The model must meet the following requirements on the test set: Determination coefficient R²>0.98 Mean absolute error MAE<1.5% Only after meeting the requirements can it be solidified as the initial health prediction model for online deployment.

[0079] Training of the second-level model (SVR)

[0080] (1) Data source: Collect operation data of the air preheater under different ash deposition states in the historical period. This data set contains efficiency deviation Δη (i) calculated by the first-level GBDT model.

[0081] (2) Input features (X): Core is Δη (i), combined with auxiliary working condition features such as unit load L, flue gas flow F_gas and its change trend.

[0082] (3) Output target (Y) - Key and difficulty: The training target value y_true of the model is the true ash deposition degree, which is a normalized value (0-100%) estimated based on known ash deposition degree.

[0083] Specifically, the method for obtaining the first training label includes: a) Offline analysis and expert experience calibration: 0% corresponds to the confirmed clean state after major overhaul.

[0084] 100% corresponds to the state where performance has severely degraded before major overhaul, and has approached the operating limit (e.g. the current of the induced draft fan exceeds the limit).

[0085] The intermediate state is linearly or nonlinearly interpolated according to the degree of performance degradation.

[0086] b) Associated physical threshold: 100% corresponds to one or more insurmountable physical operating limits (e.g. maximum output of induced draft fan, upper limit of environmental smoke temperature, maximum capacity of ash removal system).

[0087] (4) Algorithm: Support Vector Regression (SVR) algorithm is used for training and hyperparameter optimization to learn the complex nonlinear mapping relationship from the input feature vector to the soot level coefficient.

[0088] Specifically, the model self-evolution (self-learning) process (S6)

[0089] This is the core of solving the model drift problem and maintaining long-term accuracy.

[0090] 1. Triggering condition (clean state confirmation)

[0091] Any of the following quantitative criteria must be met (continuous monitoring): Criteria A (comprehensive judgment): Within M≥4 consecutive hours, the following conditions are met simultaneously: The unit load fluctuation is less than ±5%; The absolute value of the efficiency deviation of all logical control loops |Δη(i)|<1.5%; The soot level coefficient of all logical control loops y(i,t)<2.0%.

[0092] Criteria B (single index judgment): Within M≥2 consecutive hours, the absolute value of the efficiency deviation of all logical control loops |Δη(i)|<1.0%.

[0093] Criteria C (forced triggering): The operation and maintenance personnel manually confirm the clean state according to the results of major overhaul or water flushing.

[0094] 2. Execution process

[0095] Trigger: After meeting any of the above conditions, automatically trigger the model self-evolution process.

[0096] Data collection: The system automatically collects the new health data set during this period.

[0097] Policy decision: automatically decide to use the training strategy according to whether the new data volume is sufficient and whether the performance drift is significant. The performance drift degree is evaluated by introducing a quantitative indicator based on model prediction uncertainty: during the clean state judgment, the confidence interval of the model's predicted value for the key output (such as η_ref(i, t)) is calculated synchronously; if the overlap degree of the historical confidence interval and the current confidence interval continues to be lower than the preset threshold (such as 30%), or the prediction uncertainty (such as variance) significantly increases (such as more than 50% of the historical baseline), it is determined that the performance drift is significant, and the full-amount retraining strategy is triggered preferentially.

[0098] Security detection: before the data is officially used for model evolution, it flows through the adversarial sample detection module. This module performs data distribution anomaly detection based on Isolation Forest or auto-encoder reconstruction error, and uses a gradient-based adversarial perturbation identification method to screen suspected malicious interference data. If an anomaly is detected, the data flow is automatically isolated, a system security alarm is triggered, and the current evolution process is suspended until the operation and maintenance personnel confirm that the data is safe or the anomaly is resolved.

[0099] Optionally, the adversarial sample and data anomaly detection method is not limited to Isolation Forest or auto-encoder, but can also be selected from, but not limited to: local outlier factor detection, one-class support vector machine, density-based clustering method, or models specifically used for time series anomaly detection.

[0100] Retraining and verification: use the new data set to train the model (GBDT and / or SVR), and the new model must pass the performance verification again (R²>0.98, MAE<1.5%).

[0101] Model hot switching (HotSwap): after verification, through the software architecture of double model parallel running and traffic switching, the real-time prediction task is switched from the old model instance to the new model instance without feeling (within milliseconds), ensuring the continuity of the prediction service. The system also contains version management and rollback mechanism to prevent abnormal new models.

[0102] At this point, the system has completed the online and quantitative evaluation of each partition of the air preheater. The dust deposition coefficient y(i, t) is the core evaluation result of the present application.

[0103] S6: Ash removal decision and instruction generation

[0104] The system compares the real-time dust deposition degree coefficient y(i, t) obtained in step S5 with a preset threshold value, triggers a hierarchical early warning according to the comparison result, and automatically generates a dust cleaning decision instruction for the target logic control loop, wherein the instruction at least includes a target logic control loop identifier and a recommended dust cleaning temperature. The instruction is sent to the dust cleaning execution system through a communication interface to drive the dust cleaning execution system to perform precise dust cleaning operation on the specified logic control loop by establishing a micro-positive pressure and cooperative heating closed environment based on the instruction.

[0105] S7: Model self-evolution triggering and execution

[0106] The model evolution triggering condition is set as the air preheater being confirmed to be in a clean state again; the confirmation of the clean state needs to meet any of the following quantitative judgment criteria: Criterion A (comprehensive judgment): within continuous M hours (M≥4), simultaneously meet: unit load fluctuation is less than ±N% (N=5); the absolute value of the deviation of the actual efficiency of all logic control loops from the model prediction value is continuously lower than threshold X1% (X1=1.5); and the dust deposition degree coefficient y(i, t) of all logic control loops is continuously lower than threshold Y% (Y=2.0).

[0107] Criterion B (single index judgment): within continuous M hours (M≥2), the absolute value of the efficiency deviation of all logic control loops |Δη(i)| is continuously lower than a more stringent threshold X2% (X2=1.0), which can be judged as a clean state and trigger evolution.

[0108] Criterion C (forced triggering): the operation and maintenance personnel can manually confirm that the air preheater is in a clean state according to the maintenance activity results such as overhaul and water washing, and forcibly trigger the model self-evolution process.

[0109] Optionally, the system automatically judges according to preset conditions including but not limited to the criteria A, B and C. The logic judgment of the triggering condition can be realized based on fixed threshold value comparison, can also use dynamic threshold value based on statistical process control (SPC), or introduce a fuzzy logic controller to weight and fuse multiple judgment conditions to trigger the evolution process more smoothly.

[0110] ​Further, the machine learning algorithm in S1 is Gradient Boosting Decision Tree (GBDT), whose hyper-parameters are determined by five-fold cross-validation and Bayesian optimization method, usually in the range of: n_estimators between 100-500, learning_rate between 0.01-0.1, max_depth between 3-6. However, those skilled in the art should understand that the construction of the health prediction model (first-level model) is not limited to the GBDT algorithm. The machine learning algorithm can be selected from, but not limited to: Gradient Boosting Decision Tree (GBDT), Random Forest, Extreme Gradient Boosting (XGBoost), LightGBM, Classification and Regression Tree (CART), Feedforward Neural Network (FNN) or any other supervised learning algorithm capable of high-precision nonlinear regression.

[0111] Further, the multi-parameter fusion regression model in S5 adopts Support Vector Regression (SVR) algorithm. Similarly, the implementation of the second-level diagnosis model is not limited to SVR. The multi-parameter fusion regression model can be selected from, but not limited to: Support Vector Regression (SVR), Gaussian Process Regression (GPR), Relevance Vector Machine (RVM), Feedforward Neural Network (FNN) or an integrated tree model (such as GBDT, XGBoost) same or different from the first-level model.

[0112] In specific embodiments, an online partition quantitative evaluation and cleaning decision system for air preheater ash deposition includes: A data acquisition and preprocessing module for acquiring operation data and clean state confirmation; A digital twin model construction module for training an initial digital twin model and performing feature engineering; A real-time prediction module for real-time calculation of ideal efficiency reference value η_ref(i, t) of each logical control loop; An ash deposition calculation module for calculating actual efficiency, efficiency deviation and final logical control loop ash deposition coefficient y(i, t); A model self-evolution management module for managing model triggering, retraining, verification and seamless switching process.

[0113] In specific embodiments, the system further includes: A data acquisition and preprocessing module for cleaning, aligning, standardizing and generating derived features of original operation data; A model self-evolution management module integrating uncertainty evaluation and safety detection unit for performance drift evaluation and adversarial sample identification; Real-time prediction and soot accumulation degree calculation module, in series with the first level digital twin model and the second level fusion diagnosis model, realizes the complete diagnosis process from working condition input to soot accumulation degree output; Communication and execution interface module, responsible for issuing soot accumulation degree coefficient y(i, t) and cleaning instruction to the logic control loop cleaning execution system, forming an intelligent closed loop.

[0114] In specific embodiments, the model self-evolution management module is integrated with: Uncertainty assessment unit, for calculating model prediction confidence interval and assessing performance drift degree; Security protection unit, built-in anti-samples detection and data security check function, to ensure the integrity and security of the evolution data flow.

[0115] In a specific embodiment, an online partition quantitative evaluation and cleaning decision method for air preheater soot, its model side system composition and data flow are as shown in Figure 1 , including: Subgraph A [data source and input] A1 [DCS / field sensor] A2 [artificially confirmed clean state] end Subgraph B [data acquisition and preprocessing module] B1 [data cleaning] B2 [outlier processing] B3 [data alignment] B4 [standardization] B5 ["generate derived features η / Q², η / W" end A-->B.

[0116] Subgraph C [digital twin model construction module]

[0117] C1 [historical health data set]

[0118] C2 [feature engineering]

[0119] C3 ["model training (GBDT algorithm)"

[0120] C4 ["performance verification R²>0.98, MAE<1.5%"

[0121] C5 [initial health prediction model]

[0122] end

[0123] B-.->| for offline training | C.

[0124] Subgraph D [Real-time prediction module]

[0125] D1 [Load GBDT model]

[0126] D2 [Real-time working condition parameter input]

[0127] D3 ["Calculate ideal efficiency reference value η_ref(i, t)"]

[0128] end

[0129] B-->| Real-time data | D2, C5-->D1.

[0130] Subgraph E [Ash deposition calculation module]

[0131] E1 ["Calculate actual efficiency η_actual(i, t)"]

[0132] E2 ["Calculate efficiency deviation Δη(i, t)"]

[0133] E3 [Load SVR diagnosis model]

[0134] E4 ["Multi-parameter fusion calculation (Δη, L, F_gas,...)"]

[0135] E5 ["Output ash deposition coefficient y(i, t)"]

[0136] end

[0137] B-->| Real-time temperature data | E1, D3-->E2, E2-->E4, E4-->E5.

[0138] Subgraph F [Model self-evolution management module]

[0139] F1 [Monitor clean state trigger condition]

[0140] F2 [Collect new healthy data set]

[0141] F2a [Uncertainty evaluation and safety detection]

[0142] F3 ["Model update strategy decision (incremental / total)"]

[0143] F4 [Model retraining and verification]

[0144] F5 [Model hot switching]

[0145] end

[0146] A2 --> F1, F1 --> |trigger| F2, F2 --> F3, F3 --> F4, F4 --> F5, F5 --> |update| C5, F5 --> |update| E3.

[0147] Subgraph Z [Output and Communication]

[0148] Z1 (["Real-time output y(i,t)"])

[0149] Z2 (["Send to soot cleaning execution system"])

[0150] end

[0151] E5 --> Z1, Z1 --> Z2.

[0152] In the specific embodiment, an online partition quantitative evaluation and soot cleaning decision method for air preheater soot, the digital twin model offline training and online application process is as shown in Figure 2 , including: Subgraph S1 ["First stage: offline model construction (S100)"] A1 ["S101: Health benchmark data collection"] A2 ["S102: Data cleaning and air leakage correction"] A3 ["S103: Feature engineering and data set division"] A4 ["S104: Algorithm training (GBDT)"] A5 {Model performance meets the standard? R²>0.98 and MAE<1.5% A6 ["S105: Model parameter solidification"] end Subgraph S2 ["Second stage: online application (S200-S500)"] B1 ["S200: Real-time data sensing"] B2 ["S210: Real-time air leakage correction"] B3 ["S300: Model real-time prediction Output η_ref_i"] B4 ["S310: Calculate residual Δη_i=η_actual_i-η_ref_i"] B5 ["S320: Multi-parameter fusion diagnosis (Δη_i, wind temperature, pressure difference, oxygen content)"] B6 ["S500: Calculate and output soot degree coefficient y(i,t)"] end Subgraph S3["Third Stage: Model Self-Evolution (S800)"] C1["S800: Trigger Condition Evaluation"] C2["S810: New Health Data Collection"] C2a[S815: Uncertainty Assessment and Security Detection] C3["S820: Incremental Model Training or Retraining"] C4{Does the new model meet performance requirements?} C5["S830: Model Parameter Update"] end A1-->A2, A2-->A3, A3-->A4, A4-->A5, A5--No-->A4, A5--Yes-->A6, B1-->B2, B2-->B3, B3-->B4, B4-->B5, B5-->B6, C1-->C2, C2-->C3, C3-->C4, C4--No-->C3, C4--Yes-->C5, A6--Model Deployment-->B3, C5--Model Iteration Update-->B3.

[0153] In a specific implementation, an online zoned quantitative assessment and dust removal decision-making method for air preheater ash accumulation includes a self-evolving trigger condition judgment and execution process, such as... Figure 3 As shown, it includes: Start (["Start monitoring evolution trigger conditions"]) --> Monitor ["Continuously monitor system status"] Monitor --> ConditionA {"Satisfies criterion A?" Load fluctuation < ±5% |Δη(i)|<;1.5%&; y(i,t) < 2.0% The duration M ≥ 4 hours"}&ConditionB{"Satisfies criterion B?" |Δη(i)|<;1.0% The duration M ≥ 2 hours"}&ConditionC{"satisfies criterion C?" Manual forced trigger"} ConditionA -- Yes ---> Trigger["Trigger model self-evolution process"] ConditionB -- Yes -- Trigger ConditionC -- Yes ---> Trigger Trigger --> CollectData["S810: Collect New Health Dataset"] CollectData --> UncertaintyCheck ["S815: Uncertainty Assessment and Safety Detection"] UncertaintyCheck --> StrategyDecision {"New data volume sufficient and Performance drift significant?"} StrategyDecision -- is --> FullRetrain ["S820: Full Retraining"] StrategyDecision -- no --> IncrementalLearning ["S820: Incremental Learning"] FullRetrain --> ValidateFull {"New model performance validation R²>;0.98&;MAE<;1.5%?"} IncrementalLearning --> ValidateIncremental {"New model performance validation R²>;0.98&;MAE<;1.5%?"} ValidateFull -- no --> FullRetrain ValidateIncremental -- no --> IncrementalLearning ValidateFull -- yes --> HotSwap ["S830: Model Hot Swap"] ValidateIncremental -- yes --> HotSwap HotSwap --> Finish (["Evolution complete, model updated"]) In a specific embodiment, an online partitioned quantitative evaluation and cleaning decision method for air preheater ash accumulation, combined with Figure 1 , the specific principle is as follows: (1) The system data comes from the DCS or field sensors (A1) of the power plant and the clean state signal confirmed by artificial (A2). The data acquisition and preprocessing module (B) receives the original data, performs data cleaning (B1), outlier processing (B2), data alignment (B3), standardization (B4) and generates key derived features η / Q² and η / W (B5) to suppress the interference of operating condition fluctuations.

[0154] (2) The preprocessed data is divided into two paths: one for model construction in the offline stage, and the other for real-time prediction in the online stage.

[0155] (3) The digital twin model construction module (C) utilizes the historical health dataset (C1), performs feature engineering (C2), and adopts machine learning algorithms such as GBDT for model training (C3). The trained model needs to undergo performance verification (C4) to ensure that the determination coefficient R² is greater than 0.98 and the mean absolute error MAE is less than 1.5%. After meeting the standards, the initial health prediction model (C5) is solidified for online deployment.

[0156] (4) The real-time prediction module (D) loads the deployed GBDT model (D1), receives real-time operating condition parameters (D2), calculates and outputs the ideal heat exchange efficiency reference value η_ref(i, t) of each logical control loop under the current operating condition (D3).

[0157] (5) The ash deposition calculation module (E) calculates the actual heat exchange efficiency η_actual(i, t) of each logical control loop according to real-time temperature data in parallel (E1), and calculates the deviation Δη(i, t) from the ideal reference value (E2). Then, input Δη(i, t) as the core feature, fuse other operating condition parameters, and input into the pre-trained SVR diagnostic model (E3) for multi-parameter fusion calculation (E4), finally output a 0-100% dimensionless logical control loop ash deposition degree coefficient y(i, t) (E5) in real time. The coefficient is sent to the downstream ash removal execution system through the communication module (Z2) as the core input of its decision-making.

[0158] (6) The model self-evolution management module (F) continuously monitors the system state, and once any clean state trigger condition (F1) is met, it automatically triggers the evolution process. This module is responsible for collecting new health datasets (F2) and immediately handing them over to the integrated Uncertainty Assessment & Security Detection (F2a) unit for processing. This unit first performs performance drift assessment based on prediction uncertainty to provide quantitative basis for strategy decision-making; at the same time, it performs adversarial sample detection to ensure data security. Subsequently, the module decides to use incremental learning or full retraining strategy (F3) based on the evaluation results, performs model retraining and verification (F4), and finally updates the online running GBDT model (C5) and SVR model (E3) through hot switching technology (F5) without feeling, thereby ensuring the long-term accuracy of model prediction.

[0159] (7) The ash deposition degree coefficient y(i, t) output by the ash deposition diagnosis system will be sent as the core instruction to the ash removal execution system, which is used to trigger the precise ash removal program for the specific ring-shaped control area. The ash removal program establishes and maintains a micro-positive pressure environment (such as 50-150 Pa) relative to adjacent air boxes through the inlet air box of the target ring-shaped control area, and finally forms an aerodynamic seal inside the rotor to achieve efficient ash removal.

[0160] In a specific embodiment, an online partition quantitative evaluation and soot cleaning decision method for air preheater soot, the full life cycle management process is as shown in Figure 2 , mainly including three stages: The first stage (S100) is offline model construction. When the air preheater is in a reference clean state, multi-dimensional historical operation data is collected (S101); the data is cleaned and air leakage is corrected (S102); feature engineering and data set division are performed (S103); an initial health prediction model is trained using the GBDT algorithm (S104); the model performance is verified, and if the determination coefficient R² is greater than 0.98 and the mean absolute error MAE is less than 1.5%, it is qualified, and the model parameters are solidified (S105); if it is not qualified, it is returned to retrain.

[0161] The second stage (S200-S500) is online application. Real-time operation data is collected (S200); real-time air leakage correction is performed (S210); the ideal efficiency reference value η_ref_i of each logical control ring is predicted in real time using the deployed digital twin model (S300); the actual efficiency and the ideal reference value are calculated (S310); multiple parameters (such as Δη_i, wind temperature, pressure difference, oxygen content) are fused for intelligent diagnosis (S320); finally, the logical control ring soot degree coefficient y(i,t) is calculated and output (S500), completing an online diagnosis cycle.

[0162] The third stage (S800) is model self-evolution. When the system meets the clean state confirmation condition (see Figure 3 for details), this process is triggered. It includes trigger condition evaluation (S800), collection of new health data (S810), model incremental training or retraining (S820), new model performance verification (S830), and parameter update. The updated model is iteratively deployed to the online application stage (S300), thereby realizing model self-evolution and long-term accurate prediction.

[0163] In a specific embodiment, an online partition quantitative evaluation and soot cleaning decision method for air preheater soot, the trigger and execution process of the model self-evolution process is as shown in Figure 3 : The system continuously monitors the operating state and simultaneously determines whether any of the three preset trigger criteria is met: Criterion A (comprehensive determination): In the last M (M≥4) hours, the unit load fluctuation is less than ±5%, the absolute value of the efficiency deviation of all logical control rings | Δη(i) | is continuously less than 1.5%, and the soot degree coefficient y(i,t) of all logical control rings is continuously less than 2.0%.

[0164] Criterion B (single-index decision): the absolute value of efficiency deviation of all logical control loops |Δη(i)| continuously remains below a more stringent threshold of 1.0% for M (M≥2) consecutive hours.

[0165] Criterion C (mandatory trigger): the clean state is manually confirmed by the operation and maintenance personnel according to the overhaul results.

[0166] When any of the conditions is met, the model self-evolution process is triggered. The system first collects a new health data set (S810), and then immediately performs the key uncertainty evaluation and safety detection step (S815). This step not only evaluates the degree of performance drift and the intelligent decision training strategy, but more importantly, ensures the integrity and safety of the data used for evolution, preventing potential risks. Subsequently, according to whether the amount of new data is sufficient and whether the performance drift is significant, the system automatically decides to use the full retraining or incremental learning strategy (S820).

[0167] Specifically, the new model must pass performance verification, ensuring that its R² is greater than 0.98 and MAE is less than 1.5%, before it is considered to be up to standard. After meeting the standard, the online model is updated without feeling through model hot switching technology, and the evolution is completed. If the verification is not up to standard, the training is returned to be re-performed.

[0168] In specific embodiments, an online partitioned quantitative evaluation and soot cleaning decision method for air preheaters, which is implemented based on the following scheme: Step one: hardware foundation: To measure the inlet and outlet temperatures of each annular control zone, temperature sensors (such as K-type thermocouples) are arranged on the inlet and outlet air ducts / flues of the primary air, secondary air, and flue gas of each annular control zone (i.e., partition). For example, in this embodiment, the rotor is divided into 24 sectors (i.e., 24 radial measurement positions), and independent inlet and outlet temperature measurement points are configured for each annular control zone corresponding to each radial position (such as the first, second, and third rings). In addition, a data collection system (DCS) of the unit, flow sensors, pressure sensors, and a central industrial control server are also required. The core algorithm software is deployed on the server. The hot seal air is usually introduced from the high-temperature air outlet of the primary air or secondary air.

[0169] Step two: software and model foundation: In this embodiment, the digital twin model is constructed using the Gradient Boosting Decision Tree (GBDT) algorithm and implemented using the Scikit-learn library in Python, which is used to predict the ideal efficiency reference value η_ref(i, t) in real time.

[0170] The multi-parameter fusion diagnosis model for calculating the soot accumulation degree coefficient y(i, t) is a pre-trained support vector regression (SVR) model (i.e., the second-level model). The two models work in series in the system and together complete the intelligent diagnosis process from the working condition input to the soot accumulation degree coefficient output.

[0171] The training process of the multi-parameter fusion diagnosis model (SVR) for calculating the soot accumulation degree coefficient y(i, t) is as follows: Training data source: Collect the operating data of the air preheater under different soot accumulation states in the historical period, and the data set should contain the efficiency deviation Δη(i) calculated by the first-level GBDT model.

[0172] Feature selection and target value: Take the efficiency deviation Δη(i) obtained by the first-level model as the core input feature, and fuse auxiliary working condition features such as unit load L, flue gas flow F_gas and its change trend. The training target value (y_true) of the model is the normalized soot accumulation severity (0-100%) estimated based on the known soot accumulation degree.

[0173] Model training: Use the above historical data set with labels to train and optimize the hyperparameters of the SVR model to learn the complex nonlinear mapping relationship from the input feature vector to the soot accumulation degree coefficient.

[0174] Step three: Implementation details

[0175] (1) Source of SVR diagnosis model training labels: When the system is first deployed and lacks historical soot accumulation data labels, the true soot accumulation degree labels (y_true) used to train the SVR model can be obtained and quantified in the following ways: a) Offline analysis and expert experience calibration: Collect performance data before and after the overhaul. Define the clean state after the overhaul as 0%. Define the state before the overhaul, which has the most serious performance decline and has approached the operating limit (such as over-limit induced draft fan current and abnormal exhaust gas temperature), as 100% or a high value (such as 90%) after expert evaluation. The intermediate state is calibrated according to the performance degradation degree (such as the efficiency decline percentage relative to the 0-100% interval for linear or nonlinear interpolation).

[0176] b) Correlation with physical thresholds: The 100% severe blockage state corresponds to one or more non-crossable physical operating limits (see the next point for details). When first deployed, the design values of the equipment, operating procedures or safety thresholds can be used to deduce the feature values such as efficiency deviation Δη corresponding to these limit states, thereby providing a physical definition and label for the severe soot accumulation state.

[0177] After the first round of deployment, the source of new labels for subsequent model evolution is more explicit: label the data points that are confirmed to be clean before each trigger as 0% automatically; for the dust accumulation process, the label can be interpolated by comparing with these anchor points or estimated based on the physical model, thereby continuously enriching and optimizing the training data set.

[0178] (2) Physical and operational definition of dust accumulation coefficient: The range of dust accumulation coefficient y(i, t) is 0% to 100%, which has a clear physical and operational meaning, and its calibration is based on the safe operation limit of the equipment and the predictive maintenance strategy: a) Physical definition (upper limit of scale): The upper limit of the coefficient 100% is defined as an insurmountable physical operating limit state. It usually corresponds to one or more critical states (the specific threshold is determined according to the design parameters of the unit): Critical flow resistance: The pressure difference ΔP on the flue gas side of the logic control loop reaches the limit value that can be tolerated under the maximum output of the induced draft fan, and if the dust continues to accumulate, the unit will be forced to reduce load or shut down.

[0179] Critical heat loss: The decrease in heat exchange efficiency of the logic control loop has caused the exhaust gas temperature to exceed the upper limit of environmental protection requirements, or the hot air temperature is lower than the minimum requirement for safe operation of the burner.

[0180] Maximum capacity of the dust removal system: Even if the maximum strength of the dust removal operation is performed by the partitioned three-dimensional heat sealing system, its effect has reached the margin and cannot effectively remove dust further.

[0181] Therefore, y=100% is a comprehensive physical boundary that integrates operation safety, equipment limits, and environmental protection requirements, providing an absolute reference for the entire diagnostic scale.

[0182] b) Operational definition (control trigger): In the control system, to implement predictive maintenance, an earlier and more conservative intervention threshold needs to be set. In the control system associated with this embodiment, the following is set: y(i, t)>50% triggers the highest level of severe blockage alarm (level 3 warning) and the strongest automatic dust removal intervention (for example, start the high-temperature dust removal mode, temperature>300℃). The temperature range (220-350℃) is the optional interval for dust removal operation, and the specific target temperature is dynamically set by the intelligent control system according to the severity of dust accumulation.

[0183] This design ensures that the system takes strong measures before the performance of the equipment deteriorates to the physical limit, which is the core embodiment of the early detection and early intervention principle of predictive maintenance. The safety margin between the operation threshold (such as 50%) and the physical limit (100%) can be adjusted according to the specific unit conditions.

[0184] (3) Completeness and duration of initial health data collection: The amount of data and the coverage of working conditions required to build a robust initial digital twin model (first-level model) are crucial. In actual deployment: Recommended collection duration: It is usually necessary to collect at least 1-3 months of historical operation data continuously. This time window usually ensures the capture of the most common load fluctuation periods of the unit (such as daily start-stop, weekend load changes) and different seasonal environmental temperature changes.

[0185] Working condition coverage requirements: Data must cover as much as possible the main operating range of the unit from the minimum stable combustion load to the BECR (boiler rated evaporation load). If certain working conditions (such as very low load) do not appear during the collection period, they can be supplemented through incremental learning in the subsequent model self-evolution process. Actively coordinating with power plant dispatching, allowing the unit to experience necessary load changes under safe conditions, is an effective means to obtain high-quality initial data.

[0186] Data completeness is the fundamental guarantee of the initial accuracy of the model (R²>0.98, MAE<1.5%).

[0187] (4) Technical implementation of model hot switching: To ensure the continuity of real-time prediction services during model updates, the system uses the following strategies in its software architecture to achieve seamless hot switching: a) Dual-model parallel operation and traffic switching: The system maintains two model instances: the online model (Active) and the standby model (Standby). The self-evolution process re-trains and validates on the standby model instance. After validation, a lightweight prediction routing switch is used to instantly switch real-time data flow from the online model instance to the new standby model instance. This switching process is usually completed within milliseconds and is imperceptible to downstream systems.

[0188] b) Version management and rollback mechanism: Before each switch, a snapshot and version archive of the current online model is taken. If the new model has prediction abnormalities (which can be judged by monitoring the reasonableness of its prediction values) in the short term (such as within a few minutes) after going online, the system can automatically and quickly roll back to the previous stable version, ensuring system safety.

[0189] c) State consistency guarantee: For stateful models (such as those involving time series features), necessary state information (such as the cache of recent feature windows) needs to be synchronized from the old instance to the new instance before switching to ensure prediction continuity.

[0190] Through the above design in the software architecture, rather than simple file overwriting, the smoothness, stability, and reliability of hot switching are ensured.

[0191] In specific embodiments, taking the online diagnosis and evolution of 1000 MW ultra-supercritical unit air preheater as an example, the specific implementation is as follows: Taking the rotary air preheater of a certain 1000 MW ultra-supercritical unit as the object, the online diagnosis of the fifth logic control loop of the air preheater and the triggering of the model self-evolution are realized.

[0192] (1) Data acquisition and mapping (corresponding to S1 and S2, S3)

[0193] The system acquires real-time DCS data of the whole plant at a frequency of 1 Hz. Through the rotor angle encoder signal, the data is accurately mapped to the fifth logic control loop currently at the sampling position. The mapped data includes: Fifth loop control zone flue gas inlet temperature T_gas_in(5)=350℃(measured value) Fifth loop control zone air outlet temperature T_air_out(5)=310℃(measured value, performance degradation due to ash deposition) Fifth loop control zone air inlet temperature T_air_in(5)=28℃(measured value) Unit load L=1000 MW(measured value) Total flue gas flow F_gas=1018.04 kg / s(measured value) Ambient temperature 20℃(measured value) (2) Core state index calculation (corresponding to S2, S3) a. Calculate the actual heat exchange efficiency: η_actual(5)=(310-28) / (350-28)=282 / 322≈87.6%; b. Obtain the ideal efficiency reference value: The online deployed GBDT digital twin model receives the above current working condition parameters, and instantaneously predicts that the ideal heat exchange efficiency reference value η_ref(5) of the fifth logic control loop under the clean state should be 90.0%. This value is the prediction given by the model based on massive historical health data, representing the performance level of the equipment in a clean state.

[0194] (3) Intelligent diagnosis of ash deposition degree (corresponding to S4, S5)

[0195] a. Calculate the efficiency deviation: Δη(5)=η_actual(5)-η_ref(5)=87.6%-90.0%=-2.4% This negative deviation clearly indicates that the current heat exchange performance of the logic control loop is lower than its healthy reference.

[0196] b. Multi-parameter fusion and ash deposition degree coefficient generation: The system takes the efficiency deviation Δη (5) = -2.4% as the core feature, and fuses the statistical features such as unit load L, flue gas flow F_gas and its recent trend, and inputs them into the pre-trained SVR deposition degree diagnosis model. The model performs inference calculation according to the non-linear mapping relationship it has learned, and outputs a quantitative diagnosis result in real time: the fifth logical control loop deposition degree coefficient y (5) = 15%.

[0197] (4) Decision and effect verification

[0198] The quantitative diagnosis result y (5) = 15% is sent in real time to the downstream partitioned three-dimensional heat sealing and ash removal system through the data communication interface. After receiving this signal, the ash removal system determines that the area is in a moderate deposition state (second level warning), and automatically starts the targeted and accurate ash removal program (target temperature 230°C). After the ash removal is completed, the monitoring data shows that the heat exchange efficiency of the logical control loop is restored to 89.8% from 87.6% before ash removal, which is significantly close to the healthy benchmark level predicted by the model, which verifies the accuracy of the output of the diagnosis model and the reliability of the ash removal decision basis.

[0199] (5) Model self-evolution (corresponding to S6)

[0200] Suppose that after a period of time, the air preheater is restored to clean after maintenance. The system monitors that the unit load is stable near 1000MW (fluctuation <±5%), the absolute value of the efficiency deviation of all logical control loops is continuously less than 1.5%, and the deposition degree coefficient y (i, t) of all logical control loops is continuously less than 2.0% for 4 hours, automatically determines that the clean state condition of criterion A is met, and triggers the model self-evolution process. The system collects the new healthy data set during this period, first calculates the model prediction confidence interval through the uncertainty evaluation unit, finds that the overlap degree with the historical interval is less than 30%, and determines that the performance drift is significant; at the same time, the safety protection unit does not detect data anomalies. Then, the online GBDT model is retrained in full. The verification performance of the new model is R²=0.992, MAE=1.1%, and the online model is updated without feeling after the verification meets the standard, completing self-evolution and ensuring long-term prediction accuracy.

[0201] In the specific embodiment, an electronic device is also included, which comprises a memory, a processor, and a computer program stored on the memory and executable on the processor, and the processor implements an online partitioned quantitative evaluation and ash removal decision method for air preheater deposition when executing the program. The electronic device is not limited to a specific model of industrial computer or server, and any general-purpose or special-purpose computing device with computing power is included.

[0202] In the specific embodiment, a computer readable storage medium is also included, which stores a computer program. The program is executed by a processor to implement an online partitioned quantitative evaluation and cleaning decision method for ash deposition of an air preheater. The storage medium includes, but is not limited to, a hard disk, a solid state disk, an optical disk, a U disk, an SD card, and various types of non-volatile memories.

[0203] In the specific embodiment, the system further includes a communication module for sending the logic control ring ash deposition coefficient y(i, t) to a cleaning execution system, so that the system performs accurate cleaning operation on the corresponding partition based on the coefficient. The communication mode includes, but is not limited to, OPC UA, Modbus TCP / IP based industrial Ethernet communication, 4G / 5G network based wireless communication, or data interaction through an industrial real-time database.

[0204] In the specific embodiment, the cleaning execution system includes a radial baffle and a damper, a mechanical plugging device, and a hot sealing air injection and intelligent control system, which realizes accurate online cleaning by building a high-temperature sealed micro-positive pressure environment. The execution mechanism for building the micro-positive pressure environment is not limited to this, and any execution system that can independently control the air volume, pressure or temperature of the annular control area to create a local environment conducive to ash removal is equivalent to the cleaning execution system of the patent. The intelligent control logic is not limited to a specific control algorithm, and can be selected from, but not limited to, proportional-integral-derivative control, fuzzy control, model predictive control, or rule-based expert system.

[0205] In the specific embodiment, the ash deposition coefficient y(i, t) is used as a core instruction to directly drive the cleaning execution system to start the cleaning program for a specific annular control area, forming a complete intelligent closed loop of diagnosis-cleaning-model optimization-re-diagnosis.

[0206] In the specific embodiment, the method and system together with the partitioned three-dimensional hot plugging system form an air preheater intelligent maintenance patent wall, covering the complete technology chain from state perception, intelligent diagnosis, decision instruction to accurate execution.

[0207] The embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same and similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the related parts can be referred to the method part.

[0208] The foregoing description of the disclosed embodiments enables a person skilled in the art to make or use the application. Modifications of these embodiments will occur to persons of skill in the art, and that the appended claims are intended to cover all such modifications that do not depart from the true spirit and scope of the application. Therefore, the application is not limited to the embodiments shown but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An online zonal quantitative assessment and dust removal decision method for air preheater ash accumulation, characterized in that, include: Based on historical operating data of the air preheater under baseline clean conditions, one or more health prediction models are constructed to predict the ideal heat exchange efficiency baseline value of each logic control loop under a given operating condition. Receive real-time operating condition parameters and, through the health prediction model, output the ideal heat exchange efficiency benchmark value of each logic control loop under the current operating condition; The efficiency deviation is calculated based on the ideal heat exchange efficiency benchmark value and the actual heat exchange efficiency measured in real time. Based on the efficiency deviation, the grayscale coefficient y(i,t) of each logic control loop is output online and quantitatively through a preset evaluation model. The gray scale coefficient y(i,t) is compared with a preset threshold, and a hierarchical early warning and dust removal decision instruction is generated for the target logic control loop based on the comparison result.

2. The online zoned quantitative assessment and dust removal decision method for air preheater ash accumulation according to claim 1, characterized in that, The operating parameters include one or more of the following: unit load, flue gas flow rate, air flow rate, ambient temperature, and rotor speed.

3. The online zoned quantitative assessment and dust removal decision method for air preheater ash accumulation according to claim 2, characterized in that, When constructing the input features of the health prediction model, derived features are introduced to suppress the disturbance of operating condition fluctuations. The derived features include: specific efficiency η / Q² and / or unit load efficiency η / W, where η is the heat exchange efficiency, Q is the medium flow rate, and W is the unit load.

4. The online zoned quantitative assessment and dust removal decision method for air preheater ash accumulation according to claim 1, characterized in that, The health prediction model and / or the evaluation model are data-driven models based on machine learning, which include one or more of gradient boosting decision trees, random forests, neural networks, or support vector machines.

5. The online zoned quantitative assessment and dust removal decision method for air preheater ash accumulation according to claim 1, characterized in that, The health prediction model and / or the assessment model are mathematical models based on physical mechanisms, and the mathematical models based on physical mechanisms are digital twin models.

6. The online zoned quantitative assessment and dust removal decision method for air preheater ash accumulation according to claim 1, characterized in that, The efficiency deviation, as a core feature, is input into the evaluation model to calculate the grayscale coefficient.

7. The online zoned quantitative assessment and dust removal decision method for air preheater ash accumulation according to claim 1, characterized in that, It also includes a model self-evolution step: based on preset triggering conditions, when it is confirmed that the air preheater is in a clean state, new health baseline data is collected to update the health prediction model and / or the evaluation model.

8. The online zoned quantitative assessment and dust removal decision method for air preheater ash accumulation according to claim 7, characterized in that, The confirmation of the cleanliness status meets any of the following conditions: The maintenance personnel will manually confirm the results based on the inspection findings; or Within a continuous period of M hours (M≥4), the system monitored that the unit load fluctuation was less than ±N%, the absolute value of the efficiency deviation of all logic control loops was consistently lower than the threshold X~1~%, and the grayscale coefficient of all logic control loops was consistently lower than the threshold Y%. or Within a continuous period of M hours (M≥2), the system monitored that the absolute value of the efficiency deviation of all logic control loops remained below a more stringent threshold X~2~.

9. The online zoned quantitative assessment and dust removal decision method for air preheater ash accumulation according to claim 1, characterized in that, The 0-100% scale of the grayscale coefficient y(i,t) is defined based on one or more insurmountable physical operating limits, where 100% corresponds to a critical state of any one or a combination of the following: The logic control loop reaches the maximum output limit of the induced draft fan when the flue gas side pressure difference reaches the limit. A decrease in the heat exchange efficiency of the logic control loop can cause the flue gas temperature to exceed the environmental protection limit or the hot air temperature to fall below the safe operating limit. The dust removal system is no longer able to effectively remove accumulated dust under maximum operating conditions.

10. An online zoned quantitative assessment and dust removal decision system for air preheater ash accumulation, used to implement the online zoned quantitative assessment and dust removal decision method for air preheater ash accumulation as described in any one of claims 1 to 9, characterized in that, include: The data acquisition and preprocessing module is used to collect historical and real-time operating data of the air preheater; The health prediction model building and management module is used to build, train, validate, deploy, and update health prediction models. The real-time prediction and health baseline determination module is used to load the health prediction model and predict the ideal heat exchange efficiency baseline value of each logic control loop based on real-time operating parameters. The online quantitative evaluation module is used to calculate the actual heat exchange efficiency and efficiency deviation of each logic control loop, and calculates and outputs the gray scale coefficient y(i,t) of each logic control loop in real time based on the evaluation model. The dust removal decision module is used to compare the gray scale coefficient y(i,t) of the logic control loop with a preset threshold, and generate a graded early warning signal and a dust removal control command for the target logic control loop. The communication module is used to send the dust removal decision instructions to the dust removal execution system.

Images