Thermal power plant auxiliary control system abnormal state early warning system and method

Abstract

This application discloses an abnormal state early warning system and method for auxiliary control systems in thermal power plants, relating to the field of thermal power plant status early warning. It uses the original control command flow as system excitation, combines it with feedback sensing flow, and utilizes pure time delay measurement technology based on mutual information to construct a synchronous input-output dataset that accurately reflects the control-response timing correlation. To simultaneously resolve the contradiction between false alarms due to equipment aging and missed alarms due to fault adaptation, an online parameter recursive identification algorithm with a forgetting factor is used to construct a time-varying model. The forgetting factor is used to track the slow decay of equipment parameters in real time to adapt to natural wear and tear. Simultaneously, an update enable signal mechanism is introduced, updating model parameters only under normal operating conditions and freezing parameters when potential anomalies occur to prevent the model from learning fault characteristics. Finally, by comparing the theoretical response with the current actual output, an adaptive threshold based on a sliding window is used to determine the residual, achieving highly reliable anomaly early warning.

Description

Technical Field

[0001] This application relates to the field of early warning of the status of thermal power plants, specifically to an early warning system and method for abnormal status of auxiliary control systems in thermal power plants. Background Technology

[0002] The auxiliary control system of a thermal power plant (covering key equipment such as coal mills, blowers, and coal feeders) is a core component ensuring the energy conversion efficiency and operational safety of generating units. With increasing demands for unit flexibility in the power system, thermal power plant equipment operates under complex environments of varying operating conditions and large load fluctuations, and faces natural wear and tear under harsh conditions such as high temperature, high pressure, and dust. Developing an efficient and reliable early warning system for abnormal states in the auxiliary control system can promptly detect potential performance degradation and early failures, which is crucial for reducing the risk of unplanned outages, lowering maintenance costs, and promoting the intelligent construction of thermal power plants.

[0003] However, existing abnormal state early warning technologies for auxiliary control systems in thermal power plants still have significant shortcomings in practical applications, making it difficult to meet the high reliability requirements of industrial-grade systems. First, existing technologies mostly employ static models trained offline (such as fixed neural network weights or memory matrices) for state monitoring. This static modeling approach cannot adapt to the time-varying characteristics of auxiliary control equipment in thermal power plants, such as the slow efficiency decline caused by wear on coal mill rollers or ash accumulation on fan blades. As equipment operates for longer periods, this normal natural aging can be misjudged as abnormal by the static model, leading to a sharp increase in the false alarm rate over time. Second, existing monitoring methods typically focus on autoregressive models that predict future states based on historical sensor data, neglecting the system's external driving forces—control commands, such as inverter commands and blade opening commands. This lack of a control-response causal chain makes it impossible for the model to distinguish between normal response changes caused by changes in control commands and abnormal fluctuations caused by equipment failure when facing large system hysteresis or reverse adjustment conditions. Furthermore, online adaptive algorithms introduced to address aging issues often lack intelligent learning / freeze switching mechanisms. When a device actually fails, the unrecognized adaptive model quickly learns and adapts to the fault state, causing the prediction residuals to disappear and resulting in severe missed fault detection. Therefore, existing technologies cannot avoid missed fault detection while adapting to the natural aging of equipment, and also lack the ability to accurately handle the complex hysteresis relationship between control commands and feedback signals. Summary of the Invention

[0004] To address the aforementioned problems in the existing technology, this application provides a method for early warning of abnormal states in the auxiliary control system of a thermal power plant. The method includes: acquiring the original control command stream and the original feedback sensor stream, wherein the original control command stream includes commands from the coal feeder frequency converter, commands for the blower impeller opening, and the setpoint for the coal mill loading oil pressure; preprocessing the original control command stream and the original feedback sensor stream to obtain a synchronous input / output dataset; determining the pure system time delay based on mutual information in the synchronous input / output dataset to obtain the system pure time delay steps and the regression data vector; performing online parameter recursion identification with a forgetting factor on the regression data vector, the current actual output in the synchronous input / output dataset, and the update enable signal to obtain a system parameter vector; based on the system parameter vector, calculating the theoretical response and dynamic residuals of the regression data vector and the current actual output in the synchronous input / output dataset to obtain a residual error value; and performing an adaptive threshold determination based on the sliding window statistical characteristics on the residual error value to obtain an abnormal state flag.

[0005] This application also provides an abnormal state early warning system for an auxiliary control system of a thermal power plant, comprising: a raw flow information acquisition module for acquiring raw control command flow and raw feedback sensor flow, wherein the raw control command flow includes coal feeder frequency converter command, blower blade opening command, and coal mill loading oil pressure setpoint; a raw flow information preprocessing module for preprocessing the raw control command flow and raw feedback sensor flow to obtain a synchronous input-output dataset; a system time measurement module for measuring the pure system time delay based on mutual information of the synchronous input-output dataset to obtain the system pure time delay step and regression data vector; an online parameter recursion identification module for performing online parameter recursion identification with a forgetting factor on the regression data vector, the current actual output in the synchronous input-output dataset, and the update enable signal to obtain a system parameter vector; a residual error calculation module for performing theoretical response and dynamic residual calculation on the regression data vector and the current actual output in the synchronous input-output dataset based on the system parameter vector to obtain a residual error value; and an abnormal state flag generation module for performing adaptive threshold determination based on sliding window statistical characteristics on the residual error value to obtain an abnormal state flag.

[0006] Compared with existing technologies, this application provides an abnormal state early warning system and method for auxiliary control systems in thermal power plants. Addressing the causal decoupling problem caused by neglecting external driving forces in existing technologies, this scheme uses the original control command flow as the system excitation, combines it with feedback sensing flow, and utilizes a pure time delay measurement technique based on mutual information to construct a synchronous input-output dataset that accurately reflects the control-response timing correlation, thereby solving the problem of dynamic characteristic characterization under large hysteresis conditions. To simultaneously resolve the contradiction between false alarms due to equipment aging and missed alarms due to fault adaptation, an online parameter recursive identification algorithm with a forgetting factor is used to construct a time-varying model. The forgetting factor is used to track the slow decay of equipment parameters in real time to adapt to natural wear. Simultaneously, an update enable signal mechanism is introduced, updating model parameters only under normal operating conditions and freezing parameters when potential anomalies occur to prevent the model from learning fault characteristics. Finally, by comparing the theoretical response with the current actual output, an adaptive threshold based on a sliding window is used to determine the residual, achieving highly reliable anomaly early warning. Attached Figure Description

[0007] The above and other objects, features and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings.

[0008] Figure 1 This is a flowchart illustrating an abnormal state early warning method for an auxiliary control system of a thermal power plant according to an embodiment of this application; Figure 2 This is a schematic diagram of the data flow in the abnormal state early warning method of the auxiliary control system of a thermal power plant according to an embodiment of this application; Figure 3 for Figure 1 Flowchart of step S4; Figure 4 This is a block diagram of an abnormal state early warning system for an auxiliary control system of a thermal power plant according to an embodiment of this application. Detailed Implementation

[0009] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. It should be understood that the drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0010] Based on the shortcomings in the aforementioned technical fields, this application proposes a method for early warning of abnormal states in the auxiliary control system of a thermal power plant. Figure 1 This is a flowchart illustrating an abnormal state early warning method for an auxiliary control system of a thermal power plant according to an embodiment of this application. Figure 2 This is a schematic diagram of the data flow in the abnormal state early warning method of the auxiliary control system of a thermal power plant according to an embodiment of this application. Figure 1 and Figure 2As shown, the abnormal state early warning method for the auxiliary control system of a thermal power plant according to an embodiment of this application includes: Step S1, acquiring the original control command stream and the original feedback sensor stream, wherein the original control command stream includes the coal feeder frequency converter command, the blower blade opening command, and the coal mill loading oil pressure setpoint; Step S2, preprocessing the original control command stream and the original feedback sensor stream to obtain a synchronous input-output dataset; Step S3, performing system pure time delay measurement based on mutual information on the synchronous input-output dataset to obtain the system pure time delay step and the regression data vector; Step S4, performing online parameter recursive identification with a forgetting factor on the regression data vector, the current actual output in the synchronous input-output dataset, and the update enable signal to obtain a system parameter vector; Step S5, based on the system parameter vector, performing theoretical response and dynamic residual calculation on the regression data vector and the current actual output in the synchronous input-output dataset to obtain a residual error value; Step S6, performing adaptive threshold determination based on the sliding window statistical characteristics on the residual error value to obtain an abnormal state flag.

[0011] In step S1, the original control command stream and the original feedback sensor stream are acquired. The original control command stream includes the feeder frequency converter command, the blower impeller opening command, and the coal mill loading oil pressure setpoint. It should be understood that the changes in the operating state of key equipment in the auxiliary control system of a thermal power plant, such as the coal mill, blower, and feeder, are essentially physical response processes driven by commands issued by the distributed control system. Conventional monitoring methods often only focus on the historical trends of sensor feedback data, i.e., using autoregressive models for prediction. This approach severs the causal relationship between control commands and equipment responses, making it difficult to distinguish between normal regulatory responses and abnormal equipment faults under varying loads or large hysteresis conditions, and failing to accurately determine the aging degree of equipment through a single feedback signal. To construct a dynamic model that truly reflects the physical characteristics of the equipment, it is necessary to introduce the command signals driving the equipment's actions and the feedback signals characterizing the actual state of the equipment from the source, establishing a mapping relationship between input and output. Therefore, this application first obtains a complete data foundation including system excitation and system response. By synchronously collecting control command streams and feedback sensor streams, it provides the necessary input variables for the subsequent establishment of dynamic mathematical models based on causal relationships and the determination of pure time delay, thereby solving the problem of early warning failure and false alarms under certain operating conditions caused by simply relying on feedback signals.

[0012] Specifically, an exemplary implementation of step S1 of this application is performed as follows: This process relies on the data interface of the existing plant-wide monitoring information system or distributed control system of the thermal power plant. The main purpose of this step is to extract stimulus-response data pairs that can fully characterize the dynamic characteristics of auxiliary control equipment in a complex industrial environment. The implementation process first involves the construction and configuration of the underlying data acquisition architecture, connecting to the real-time historical database of the thermal power plant through an OPC interface or a dedicated database driver. In the database, each physical device or measuring point corresponds to a unique code, generally following the KKS (Power Plant Identification System) coding rules. The data acquisition process involves batch extraction of time-series data of specified measuring points within a specific time period according to a preset sampling frequency, such as 1 second or 5 seconds, forming a continuous time-series data stream.

[0013] For acquiring the raw control command stream, the system focuses on three core control signals: the coal feeder frequency converter command, the blower blade opening command, and the pulverizer loading oil pressure setpoint. These three signals represent the external actions applied to the pulverizer and pulverizing system, and are the source of system state changes. Specifically, the coal feeder frequency converter command is a value expressed as a percentage or frequency (Hz), which directly controls the speed of the coal feeder motor, thereby determining the mass flow rate of raw coal entering the pulverizer per unit time. This command is set by the unit load control logic or manually by the operator, and in the data stream, it appears as a sequence of values ​​that change over time. For example, during the unit load increase phase, the command value will rise in a step-like or sloping manner. The blower blade opening command is a position control signal issued to the actuator of the primary air fan or blower, used to adjust the angle of the fan blades, thereby changing the air volume and pressure entering the pulverizer. This signal determines the dynamic characteristics of the airflow carrying pulverized coal. The setpoint for the hydraulic loading pressure of the coal mill is a control target value for the hydraulic loading system of a medium-speed coal mill. It determines the grinding pressure of the grinding rollers on the coal bed, directly affecting the fineness of the pulverized coal and the energy consumption of the coal mill. This value usually changes relatively gradually, but adjustments are made when the coal type changes or wear intensifies. In implementation, the storage locations of these three control variables in the historical database are locked based on the corresponding KKS codes, and the numerical sequences are extracted according to a unified timestamp index to form a vector-based raw control command flow.

[0014] For acquiring the original feedback sensing flow, the focus is on the actual physical quantity measurement value after the acquisition equipment responds to the aforementioned control commands. In this application, the feedback flow includes coal feed rate feedback, air volume measurement value, and mill differential pressure. The coal feed rate feedback is the coal quantity data measured in real time by a weighing sensor installed under the coal feeder belt, in tons per hour (t / h). It is a direct response to the commands from the coal feeder frequency converter, but physically there is a lag in the mechanical transmission and weighing process. The air volume measurement value is the data read from a Pitot tube, airfoil anemometer, or thermal flow meter installed in the air duct, measured in cubic meters per hour or kilograms per second, characterizing the actual airflow entering the coal mill under the action of the blade opening command. Mill differential pressure (or coal mill differential pressure) is one of the most critical parameters for monitoring the internal operating status of the coal mill. It is obtained by measuring the pressure difference between the hot air inlet and the pulverized coal outlet of the coal mill using a differential pressure transmitter. This parameter comprehensively reflects the coal storage capacity, ventilation resistance, and grinding efficiency inside the coal mill, and is the result of the combined effect of all the aforementioned control commands. For example, when the coal feed rate increases while the air volume remains constant, the differential pressure in the grinding bowl will rise. The system also extracts measurement data from the database that perfectly matches the time range of the control command flow, based on the measurement point codes corresponding to these physical quantities.

[0015] In a specific implementation scenario, taking the No. E coal mill of a 600MW supercritical coal-fired unit as an example, the sampling period is set to 1 second to extract the operating data of the past 24 hours. First, the corresponding feeder frequency converter command point number (e.g., 10HCE10AA001), blower motor blade opening command point number (e.g., 10HBE10AA002), and loading oil pressure setpoint point number (e.g., 10HDE10AA003) for the No. E coal mill are found by querying the index table. Subsequently, the system does not just capture the value at a single moment, but rather captures a continuous sequence within a time window. For example, at 10:00:00 AM, the feeder command is 45%, and the corresponding coal feed rate feedback (e.g., 10HCE10CF001) is 38 t / h, the air volume measurement value is 90 t / h, and the mill bowl differential pressure is 3.5 kPa. As the unit receives the load increase dispatch, the coal feeder command rises to 50% at 10:00:10. Due to mechanical inertia and transmission delay, the coal feed rate feedback may not start to climb to 42 t / h until 10:00:15, while the differential pressure in the grinding bowl will slowly rise to 4.0 kPa in the following tens of seconds as the coal seam thickens. This temporal misalignment and numerical correlation is the core information upon which subsequent modeling relies.

[0016] In this process, the original control command flow can be viewed as a sequence of input vectors. ,in The frequency converter command for the coal feeder at time t represents the command given at time t. This represents the command for the opening degree of the blower blades. This represents the setpoint for the applied oil pressure. The corresponding raw feedback sensor flow is considered as the output vector sequence. ,in This represents the coal feed rate feedback at time t. Represents the measured air volume. This represents the differential pressure of the grinding bowl. Due to electromagnetic interference, sensor drift, or communication packet loss in the industrial environment, the directly acquired data often contains noise or dead pixels (such as NaN values ​​or spikes exceeding the physical range). To ensure the availability of the data stream, a preliminary validity check is performed at the interface layer to ensure that the extracted data has a quality code of "Good". If data loss occurs due to communication interruption, the acquisition module records the missing time period for subsequent preprocessing steps to fill or remove it.

[0017] In step S2, the raw control command stream and raw feedback sensor stream are preprocessed to obtain a synchronous input-output dataset. Correspondingly, the environment of the auxiliary control field in thermal power plants is filled with various strong electromagnetic interferences and mechanical vibrations, inevitably causing the raw feedback sensor stream obtained from the bottom layer of the distributed control system to be mixed with non-physical high-frequency noise and signal glitches. If these contaminated data are used directly for modeling, random noise is easily misjudged as a sudden change in the actual state of the equipment. At the same time, the actuator has an inherent mechanical dead-zone characteristic when receiving control commands; small command fluctuations often do not cause actual actions, but if used directly as model input without processing, non-causal disturbance information will be introduced, thereby compromising the model's accurate identification of the input-output relationship. Furthermore, the sampling timing of different data sources is inherently asynchronous, requiring preprocessing to remove false signals, correct command dead zones, and align the time base, thereby constructing a standard dataset that accurately reflects the dynamic physical characteristics of the equipment, laying a reliable data foundation for subsequent pure time delay measurements and parameter identification.

[0018] An exemplary implementation of step S2 of this application preprocesses the original control command stream and the original feedback sensor stream to obtain a synchronized input-output dataset, including: step S21, cleaning the feedback signal noise of the original feedback sensor stream to obtain a cleaned feedback sensor stream; step S22, compensating the dead time of the original control command stream based on threshold hysteresis to obtain a compensated control command stream; and step S23, synchronizing the compensated control command stream and the cleaned feedback sensor stream based on zero-order hold multi-source heterogeneous data to obtain a synchronized input-output dataset, wherein the synchronized input-output dataset contains time... Data , For the first The control command value at each sampling time. For the first The sensor feedback value at each sampling time.

[0019] Specifically, step S2 of this application is performed as follows: Step S21 This process aims to eliminate electrical spikes caused by sensor failure or transmission line interference, while retaining the true dynamic characteristics reflecting changes in operating conditions. During implementation, a statistically significant sliding time window needs to be defined first, with a window length of... This length is set according to the sampling frequency and the dynamic characteristics of the process, for example, taking... =30 sampling points, meaning that the statistical characteristics of data from the past 30 seconds are used to evaluate the reasonableness of the current value. As the data stream progresses, the window slides over the raw feedback sensor stream. For the raw sensor data at each sampling time point t, denoted as The calculation window contains Arithmetic mean of historical data points and standard deviation Taking the differential pressure of the grinding bowl mentioned in step S1 as an example, if the average differential pressure of the grinding bowl is among the 30 sampling points before the current moment... The value is 3.50 kPa, reflecting the average level of the data, while the standard deviation is... The value is 0.05 kPa, reflecting the fluctuation range of the data. Based on a normal distribution. The criterion for constructing a confidence interval for valid data is... That is, [3.35, 3.65] kPa. Then, the raw values ​​collected at the current time are... Compare with the confidence interval. If If the data point falls within this range, it indicates that the fluctuation range of the operating conditions is within the normal range and should be retained. If the value exceeds this range, for example, a sudden change to 10.0 kPa is collected, far exceeding the physical range or normal fluctuation range, then this point is determined to be an electrical spike noise or an abnormal bad point. In this case, to maintain the continuity of the signal in the time domain and avoid interruption of the model input, the effective value that has been cleaned in the previous time step is used. Replace the current outlier.

[0020] In this way, the 10.0 kPa outlier in the aforementioned differential pressure data stream will be replaced with the previous normal value, such as 3.51 kPa, thus eliminating spikes in the data. The same processing procedure is performed in parallel for other raw feedback sensor streams, such as coal feed rate feedback and air volume measurement values. Finally, the cleaned feedback sensor stream, after noise reduction and smoothing, is output.

[0021] Step S22 aims to eliminate invalid high-frequency fluctuations in the command signal caused by electronic drift or overly sensitive PID control, transforming it into a stepped, effective signal reflecting the actual mechanical action. This process is performed independently for each control command acquired in step S1, such as the coal feeder inverter command or the blower blade opening command. First, a dead-zone threshold is set. This value needs to be determined based on the physical characteristics of the specific actuator. It should be calibrated by referring to the actuator dead zone parameters provided by the equipment manufacturer or based on the steady-state noise level of historical operating data. For example, for the blower blade opening command, considering the resolution of the hydraulic servo system and the mechanical backlash of the linkage mechanism, it can be... Setting it to 0.5% means that only instruction changes exceeding 0.5% of the current value are considered valid actions. Simultaneously, it initializes the last valid instruction recorded by the system. At the initial moment, this value can be assigned the first raw sample value. Subsequently, for each time step t in the raw control command stream, the current raw control command is read and denoted as... The validity of an action is determined by comparing the relative deviation between the current original value and the valid value from the previous moment.

[0022] The following explanation uses the blower blade opening command obtained in step S1 as an example. For instance, the dead zone threshold... Set to 0.005, or 0.5%, at time t-1, the effective opening command after compensation. Stable at 50.0%. Scenario 1 (Invalid Jitter): At time t, the DCS system is disturbed, causing the original acquired value to... The price fluctuated slightly to 50.1%. At this point, the relative rate of change was |50.1 - 50.0| / 50.0 = 0.002, or 0.2%. Since 0.2% < 0.5%, this condition was not met, and it was determined that this fluctuation could not overcome the mechanical dead zone. The value remained unchanged at 50.0%, thus filtering out this false signal. Scenario 2 (Effective Action): At time t+1, the operator significantly increased the airflow, changing the original value. It jumps to 51.0%. At this point, the relative change rate is |51.0 - 50.0| / 50.0 = 0.02, or 2.0%. Since 2.0% > 0.5%, the condition is met, and it is determined that the actuator has taken a real action; update... The percentage is 51.0%. This transforms the original continuously fluctuating control signal into a series of discrete, stepped-level signals, forming the compensated control command stream. These command data, after dead-time compensation, can more accurately represent the actual input energy acting on the physical equipment.

[0023] Step S23 first requires establishing a unified time scale for the entire system. To this end, a globally discrete sampling time sequence is defined. .in, To standardize the sampling period, this value needs to be set in consideration of both the system's dynamic response speed and computational load. For large inertial objects such as auxiliary control systems of thermal power plants, this application sets... =1 second; For discrete sampling step index; This represents the total number of sampling points. This sequence forms the standard timeline for all subsequent data resampling. Next, a resampling operation based on zero-order hold-ZOH is performed. The physical significance of this method is that at any standard sampling time... The latest data value that the system can observe should be the most recently updated value before that moment, and should remain unchanged until the next update. In specific implementation, this is done separately for the cleaned feedback sensor flow. and the compensated control command flow Iterate through its original irregular timestamp sequence to find the distance from the current standard time. The most recent valid data point that is no later than that moment.

[0024] Taking the blower blade opening command (control flow) processed in step S22 and the grinding wheel differential pressure (feedback flow) processed in step S21 as examples, and using a unified sampling period... =1 second. For the 100th sampling time, i.e., t=100 seconds: in In the stream, the most recent instruction update occurred at 99.8 seconds, with a value of 51.0%. The value is 51.0%. In the flow, the most recent sensor report occurred at 99.5 seconds, with a value of 3.50 kPa. The value is 3.50 kPa. Even if At 100.2 seconds, the change is updated to 52.0%. Following the principle of not being later than the current time, this change will only be adopted at k=101, i.e., at 101 seconds. Finally, vector assembly is performed. The scalar data after the above resampling synchronization is paired one-to-one according to the time index k, forming a standardized data vector. The final synchronized input-output dataset contains a series of ordered data pairs. ,in It not only includes a single command, but a control vector composed of commands from the coal feeder frequency converter, the blower impeller opening, etc. , The feedback vector consists of coal feed rate feedback, grinding bowl differential pressure, etc. This structured dataset eliminates the asynchronicity of the original data, ensuring that at any time k, the control variables... With state variables They are strictly aligned on the timeline.

[0025] In step S3, the system pure time delay is measured based on mutual information on the synchronous input-output dataset to obtain the system pure time delay steps and regression data vector. It is worth noting that the coal feeding, air distribution, and pulverizing processes in the auxiliary control system of a thermal power plant are essentially complex material transport and energy exchange processes, exhibiting significant large inertia and pure time delay characteristics. For example, after an increase in the frequency converter command of the coal feeder, the raw coal needs to go through a series of stages, including belt conveying, coal chute sliding, grinding roller milling, and hot air drying and conveying, before finally causing changes in sensor readings such as mill bowl differential pressure or outlet temperature. This process often takes tens of seconds or even longer. If this objectively existing time delay is ignored, and the control command and feedback signal are directly correlated at the same moment, it will lead to a misalignment in causal logic between the model input and output, failing to capture the true dynamic response characteristics and greatly reducing the accuracy of fault early warning. To this end, we use statistical methods in information theory to measure the pure time delay of the synchronous input-output dataset based on mutual information, and accurately measure the specific number of time steps required for the control command to reach the feedback end. This allows us to strictly align the input and output data causally in time, eliminate the interference of pure time delay on model identification, and construct a regression data vector that conforms to the physical mechanism.

[0026] An exemplary implementation of step S3 of this application involves determining the pure system time delay based on mutual information in a synchronous input-output dataset to obtain the pure system time delay step count and regression data vector, including: step S31, performing local data window partitioning and joint probability distribution estimation on the synchronous input-output dataset to obtain a joint probability distribution set; step S32, performing optimal time delay optimization based on the maximum mutual information criterion on the joint probability distribution set to obtain the pure system time delay step count; and step S33, performing time-series alignment and autoregressive vector construction on the synchronous input-output dataset based on the pure system time delay step count to obtain a regression data vector.

[0027] Specifically, step S3 of this application is performed as follows: Step S31 first defines the data range for statistical analysis. Considering the time-varying nature of the operating state of thermal power plant equipment, it is necessary to focus on the dynamic characteristics near the current moment; therefore, a local window is partitioned. From the synchronous data stream output in step S23, with the current moment k as the endpoint, the most recent N consecutive synchronous data pairs are traced backward to form the local window dataset. Here This represents the length of the observation window, which must be set to cover the system's maximum possible time constant to ensure that the window contains the complete dynamic response process. For example, for a coal mill system, this application uses... =300, meaning 300 seconds of data, can cover the entire process from command issuance to system response stabilization. Subsequently, to find the optimal lag time, it is necessary to construct data distribution models under different lag assumptions. A lag step size search range is then defined. . Set to 0, Estimate based on physical transmission distance, for example, set to 60. For each integer hysteresis step within this range. The sequence of control commands within the window With lag movement Step feedback sequence The data pairs are then re-paired to form a set of displacement data to be analyzed. Next, the joint probability distribution of these paired data is estimated using the histogram method. Specifically, the control commands... value range and feedback signal The value range is divided into several discrete intervals, and the frequency of points falling into each two-dimensional interval is counted. This is based on the blower blade opening command. Differential pressure with grinding bowl For example, consider the current lag steps. =15. System statistics show that in the past 300 sampling points, when the instruction... Feedback is in the range of [50%, 51%] and after 15 seconds. The number of data points occurring within the [3.5 kPa, 3.6 kPa] range. This statistical method is used to calculate the lag. Marginal probability distribution under , and joint probability distribution .in, This indicates the probability of the instruction value occurring. This indicates the probability of the feedback value occurring. This represents the joint probability of both occurring simultaneously under a specific lag relationship. This process will be performed on each [item / group] within the search range. Value, such as Repeated execution eventually generates a joint probability distribution set containing all possible lag cases. .

[0028] Step S32 targets the preset search range For example, each candidate lag step within 0 to 60 seconds Calculate mutual information using Shannon's information theory formula. Mutual information content It essentially describes known control commands. In the case of delayed feedback How much has the uncertainty been reduced? After traversing and calculating all candidates... After obtaining mutual information about the values, a global optimization operation is performed. Compare this series... Find the lag step size corresponding to the maximum value, and use it as the final identified number of pure lag steps in the system. : Taking coal mill feed control as an example, such as in The mutual information calculated at 10 seconds is 0.2 bits. =0.5 bit at 20 seconds, while It reached a peak of 1.8 bits at 35 seconds, and then... The value dropped to 0.3 bits after 50 seconds. This indicates that although the system continued to run after the command was issued, the feedback signal after 35 seconds was most correlated with the current command, accurately reflecting the physical transport time required for raw coal to fall from the end of the feeder belt into the grinding bowl and be ground into powder. Therefore, the pure time delay of the system was determined. =35.

[0029] Step S33 mainly includes two core steps: alignment shifting and regression vector assembly. First, alignment shifting is performed, based on the system's pure time delay determined in step S32 using the mutual information maximization criterion. This involves logically shifting the control command flow in the synchronous input / output dataset over time. Specifically, for each time k, the feedback output... The actual input that takes effect is not the current one. It is not the past moment. Therefore, a new logical correspondence is constructed, which will include the data from the original dataset. The mapping is to the valid input at the current time k. This operation is equivalent to artificially inserting a delay in the input data channel, so that the shifted instruction sequence and the feedback sequence are physically causally synchronized. For example, if the pure time delay of the coal feeding process is measured to be 35 seconds, then when constructing the model input for predicting the differential pressure of the grinding bowl at t=100 seconds, the coal feeding instruction at t=100 seconds is no longer used, but the instruction data at t=65 seconds is captured as the driving source. Then, the regression data vector is assembled. In order to fully capture the dynamic inertial characteristics of the system, it is necessary not only to use the aligned input, but also to introduce historical input and output information. In this application, the autoregressive ergodic (ARX) model structure is used as the basic framework, which assumes that the output at the current time depends on the past. The output value at each moment and the past The input values ​​at each time point (after hysteresis correction). Among them, Let the order be the autoregressive order. The moving average order is used as the parameter. These two parameters are preset based on the system complexity, for example, taking... =2, =2 means that the current state is predicted using the historical states of the past two moments. Therefore, for each moment... Construct a regression data vector in column vector form. .

[0030] Taking a specific numerical value as an example, such as the current time... =200, system pure time delay =35, model order is set to =2, =1. The system extracts the feedback value at time 199 from the historical database. =3.60, feedback value at time 198 =3.58; simultaneously extract the instruction value at time 200-35=165. =50.0, and the instruction value at time 200-35-1=164. =49.8. Stacking these four values ​​sequentially generates a regression data vector. This vector It encapsulates all explanations of the current output. Required historical information.

[0031] In step S4, online parameter recursion identification with a forgetting factor is performed on the regression data vector, the current actual output in the synchronous input-output dataset, and the update enable signal to obtain the system parameter vector. It should be understood that the physical parameters (such as friction coefficient and heat exchange efficiency) of auxiliary control equipment in thermal power plants exhibit slow time-varying characteristics due to long-term operation under harsh conditions such as high temperature wear, ash accumulation, and mechanical fatigue. Traditional offline modeling methods obtain a fixed set of static parameters, which cannot track this dynamic evolution, leading to a decrease in model prediction accuracy over time. To solve this problem, the monitoring system needs to be endowed with online learning capabilities, enabling it to adjust internal model parameters in real time based on the latest observation data, thereby adapting to the natural aging of the equipment. However, simple online updates face the risk of mistaking sudden failures for parameter drift. Therefore, a mechanism is needed to accurately calculate the correction effect of new data on the model and quantify the deviation between the current model prediction and actual observation. Therefore, this application introduces this step to utilize the core intermediate variables of the recursive least squares algorithm with a forgetting factor—Kalman gain and prior error—to dynamically evaluate the reliability of new data and the degree of model bias.

[0032] Figure 3 for Figure 1 The flowchart for step S4. (See attached flowchart.) Figure 3As shown, an exemplary implementation of step S4 of this application involves online parameter recursion identification with a forgetting factor on the regression data vector, the current actual output in the synchronous input-output dataset, and the update enable signal to obtain the system parameter vector. This includes: step S41, calculating the Kalman gain and estimating the prior error on the regression data vector and the current actual output to obtain the Kalman gain vector and the prior prediction error; and step S42, updating the model parameter conditions of the system parameter vector at the previous time step based on the update enable signal, the Kalman gain vector, and the prior prediction error to obtain the system parameter vector.

[0033] Specifically, step S4 of this application is performed as follows: Step S41 is based on the recursive least squares (RLS) algorithm framework with a forgetting factor. This process mainly utilizes the covariance matrix of the previous time step. The regression data vector constructed at the current moment and the actual observed output values The correction factor and error signal used to update the parameters are calculated. First, the forgetting factor needs to be set. This parameter determines the length of historical data the algorithm remembers, and its value ranges from (0,1). For the slow-changing natural aging process of parameters in the auxiliary control system of a thermal power plant, this parameter is set... Close to 1, for example =0.995. This means that when the algorithm updates parameters, the weights assigned to older data decay exponentially, making the model pay more attention to recent data and enabling it to track time-varying characteristics. Next, we calculate the Kalman gain vector, a key intermediate variable. Kalman gain In a physical sense, it represents the magnitude of the correction that the new observation data should produce to the system parameter vector. To ensure the stability of the numerical calculation, the calculation process is divided into two steps. First, the scalar denominator in the gain formula is calculated. The formula is .here, It is the regression data vector output by step S33, which contains historical input and output information. It is the covariance matrix left over from the previous recursive update, which approximately reflects the uncertainty of the current parameter estimate. This measure evaluates the length of the projection of the new data vector into the current uncertainty space. Subsequently, the Kalman gain vector is generated. .

[0034] The parameters have not yet been used at the current time. Before updating the data, first use the model parameter estimates from the previous time step. To predict the current system output. Prior estimation. The calculation formula is: .here, It is a parameter vector containing system dynamic coefficients (such as autoregressive coefficients and control input coefficients). Based on this, the actual observed values ​​are calculated. The deviation between the value and the prior estimate is called the prior prediction error. : , The magnitude of this value directly reflects the interpretability of the old model parameters for new data. If... The small value indicates that the existing model can describe the current system behavior well; if A large value indicates a possible change in the system state or a need for significant correction of the model parameters. For example, in a coal mill pressure difference model, the covariance matrix at the previous time step... The forgetting factor is 0.1 times the size of the identity matrix. =0.99. The regression vector constructed at the current time step. The parameter vector at the previous time step. First, calculate the predicted value: =3.60×0.8+3.58×0.1+50.0×0.05+49.8×0.02=6.734. This is the actual sensor reading at that time. =6.750. Therefore, the prior prediction error is... =6.750 - 6.734 = 0.016. Simultaneously calculate... scalar in denominator After the calculations are complete, combined with the molecular vector, a final result is obtained. Gain vectors of the same dimension ,For example .

[0035] Step S42 is a crucial step in achieving a balance between model adaptation and anti-interference. First, it's necessary to define and obtain the update enable signal. This signal is a Boolean logic control flag, its core function being to indicate whether the data sample at the current sampling time is eligible to update the model parameters. In the specific implementation architecture, this signal is generated in real-time by a pre-processed error evaluation module. This module monitors the prior prediction error calculated in step S41. This is combined with the statistical distribution characteristics of historical errors, such as the mean and standard deviation of errors within the sliding window, to make a judgment. If the current... If the deviation is within the preset statistical confidence interval and does not exceed 3 times the standard deviation, it indicates that the current deviation is a normal model error or a small parameter drift, and the signal is set to True; if An unusual step jump or large deviation indicates that the system may have suffered a sudden disturbance or malfunction. To prevent the model from being contaminated by erroneous data, the signal is immediately set to False. This is done after obtaining the current update enable signal and Kalman gain vector. and prior prediction error Afterward, the system enters the model parameter condition update phase. At this point, the algorithm logic branches. Case A: Normal adaptation mode. When the update enable signal is True, the system determines that it is currently in normal operation or natural aging phase, allowing the model to self-evolve. The system uses the observation bias to correct the parameter vector from the previous time step based on the recursive least squares principle. The correction amount is determined by the gain vector. With scalar error The product of these factors determines the parameter adjustment step size, meaning that the larger the error (within acceptable limits) or the greater the gain, the larger the step size. Case B: Failure Freeze Mode. When the update enable signal is False, the system determines that there is a potential failure risk. To preserve the model's memory of its healthy state, the system forcibly abandons the current round of parameter updates and directly updates the parameter vector from the previous time step. As the parameter vector at the current moment Output. This freezing mechanism ensures that the model always evaluates the system with normal physical logic during the fault duration, thereby guaranteeing that fault characteristics (i.e., residuals) can persist and be detected. An exemplary implementation of step S42 of this application, based on the update enable signal, Kalman gain vector, and prior prediction error, updates the system parameter vector at the previous time step using model parameter conditions to obtain the system parameter vector, including: updating the system parameter vector at the previous time step using the following formula:

[0036] in, This is the system parameter vector from the previous time step. The Kalman gain vector. For prior prediction error, To update the enable signal, This is the system parameter vector. Immediately following the parameter update, to ensure the recursive algorithm works in the next time step ( To continue functioning normally, the system also needs to process the covariance matrix, which represents the uncertainty in parameter estimation. Corrections are then made. This step is primarily performed in normal mode (or synchronously frozen in freeze mode), and its calculation relies on the inverse matrix lemma of the Kalman filter, introducing the core parameter forgetting factor. The formula is as follows: , in the formula, Is with Identity matrices with the same matrix dimensions This is the current regression data vector. (Item) This demonstrates the role of introducing new information in reducing uncertainty; that is, the more data available, the more accurate the estimation. The value decreases. However, to prevent... When the matrix converges to zero, the algorithm loses its ability to track new changes. This can be addressed by dividing by a forgetting factor less than 1. (e.g., 0.99), the matrix is ​​artificially expanded. This dynamic balance between expansion and contraction ensures that the covariance matrix maintains a certain numerical level, thus guaranteeing that the current Kalman gain will not decay to zero, giving the algorithm the ability to continuously adapt to time-varying characteristics online. Continuing with the numerical example in step S41, if the system identifies a dynamic model of a coal mill, the parameters at the previous time step... The prior prediction error calculated at this moment. =0.016, Kalman gain Scenario 1 (Normal Wear): The system's judgment error of 0.016 is somewhat off, but within a reasonable range (e.g., the historical error standard deviation is 0.01), so the update enable signal is set to True. At this time, the update is executed: Correction item.

[0037] New parameter vector This minute change may reflect a slight increase in resistance due to wear on the grinding rollers. Scenario 2 (Sudden Anomaly): Sensor malfunction causes a sharp increase in actual readings, exceeding prior prediction errors. The error suddenly increased to 5.0. The system judged that this error was far beyond what was expected. The threshold is set to False, thus setting the update enable signal to False. At this point, the correction term is not calculated. Forced to remain as The algorithm refuses to allow the model to adapt to this large deviation of 5.0, thus this deviation will be fully retained as an alarm criterion in subsequent steps. Finally, based on the updated (or retained) state, a new covariance matrix is ​​calculated. Stored in memory for use in the next time step k+1. The call completes a full online iterative loop.

[0038] In step S5, based on the system parameter vector, the theoretical response and dynamic residual are calculated on the regression data vector and the current actual output in the synchronous input-output dataset to obtain the residual error value. Correspondingly, in abnormal state monitoring, directly observing the physical readings of sensors often makes it difficult to determine the health status of the equipment, because changes in readings can be normal responses caused by external control commands or abnormal fluctuations caused by equipment malfunctions, and are also affected by operating point drift. Simply setting high and low alarm thresholds can easily generate false alarms during changing operating conditions. Therefore, a benchmark value that can represent the expected performance of the equipment under the current operating conditions is needed. This benchmark value needs to be based on a deep understanding and real-time monitoring of the equipment's physical characteristics. To this end, this application utilizes online identified and updated system parameters, combined with the current control excitation and historical states, to calculate the theoretical response value of the equipment under normal conditions through mathematical model inversion, and compares it with the actual observed value, thereby extracting the dynamic residual signal purely caused by equipment abnormalities, providing high signal-to-noise ratio feature data for the final anomaly identification.

[0039] An exemplary implementation of step S5 of this application, based on the system parameter vector, performs theoretical response and dynamic residual calculation on the regression data vector and the current actual output in the synchronous input-output dataset to obtain a residual error value, including: step S51, based on the system parameter vector, performing theoretical model prediction on the regression data vector based on instantaneous parameters to obtain a theoretical response value; step S52, performing observation-prediction difference and residual extraction on the theoretical response value and the current actual output in the synchronous input-output dataset to obtain a residual error value.

[0040] Specifically, step S5 of this application is performed as follows: Step S51 is the core deduction link in the anomaly detection logic chain. At this point, the time step has been completed. Data acquisition, synchronous preprocessing, and online updating (or freezing) of model parameters. Utilizing the latest system parameter vector in memory. With the constructed regression data vector Perform the calculation. Among them, This is the column vector output from step S42, which contains autoregressive coefficients and control gain coefficients describing the current dynamic characteristics of the system, for example... These coefficients reflect the system's inertia, damping, and sensitivity to input signals. It is the column vector assembled in step S33, which contains historical input and output data after hysteresis correction, for example

[0041] The specific operation involves performing a vector dot product. The system performs an inner product of the two vectors in algebraic space, calculated using the following formula: The formula represents that After transpose and Multiplication involves multiplying corresponding elements and then summing the results. The physical meaning of this calculation is: based on the system's past output inertia (caused by...). Equal terms and sum (determined by coefficients) and the driving effect of external control commands (by...) Equal terms and sum (Determined by coefficients), to calculate the system's current... The theoretical output value that should be generated at time 1 This value represents the ideal performance assuming the equipment is perfectly healthy and follows the physical laws described by the currently identified parameters. Specifically, the identified parameters of the coal mill differential pressure model. Updated to The current regression data vector for Perform the dot product operation: =3.60×0.80+3.58×0.10+50.0×0.05+49.8×0.02=6.734 kPa, and the calculated 6.734 kPa is the theoretical response value.

[0042] Step S52 is a crucial step in achieving dynamic decoupling. The current time step is read from the synchronization dataset output in step S23. Actual sensor feedback value after cleaning Simultaneously, the theoretical response value calculated using the current model parameters is obtained from step S51. Next, perform the difference calculation. Subtract the theoretical prediction from the actual observation value, using the following formula: ,in, This is the extracted residual error value. This residual value represents the portion of actual observations that cannot be explained by the current normal physical model. Because Based on control commands such as coal feed rate and air volume In addition to calculating the response that should occur during normal adjustment processes, the differential operation effectively filters out all fluctuation components caused by normal load increases / decreases and command changes in the time domain. Taking coal mill differential pressure monitoring as an example, if the unit is rapidly increasing its load at a rate of 10 MW / min, the coal feed and air volume will increase significantly. Driven by this, the actual differential pressure of the coal mill... The pressure rapidly increased from 3.50 kPa to 6.75 kPa. In a traditional fixed-threshold alarm (with an upper limit of 6.0 kPa), this would already be a false alarm. However, in this solution, the theoretical response value is calculated based on real-time commands. It also rose to 6.734 kPa. At this point, residual extraction was performed: =6.750-6.734=0.016 kPa. Although the absolute pressure is very high and varies drastically, the extracted residual... The actual pressure rise is only 0.016 kPa, very close to zero. This indicates that the rise in actual pressure is completely controlled and expected, and therefore will not trigger an alarm. Conversely, if the load is stable and the command remains unchanged, the theoretical value... Maintained at 3.50 kPa, while the actual differential pressure Due to internal blockage, it slowly drifted to 4.00 kPa, at which point the residual... =4.00-3.50=0.50kPa. This significant non-zero residual will directly expose the abnormal state of the equipment.

[0043] In particular, while the aforementioned theoretical model prediction mechanism utilizes efficient vector dot product operations to generate theoretical response values, its inherent limitations stem from a fundamental assumption: that the system strictly adheres to the principle of linear superposition. However, in the complex industrial scenario of auxiliary control systems in thermal power plants, the effectiveness of this assumption has significant weaknesses. The physical processes of equipment, such as coal mills, fans, or feedwater pumps, do not exhibit simple linear combinations in their dynamic responses; rather, they display strong nonlinearity and coupling characteristics. Specifically, the linear model neglects two crucial special physical relationships: first, state-dependent gain, meaning that the equipment's response sensitivity strongly depends on its current operating point; for example, the adjustment gain of the blower blades varies drastically at different opening degrees. Second, the bilinear effect of fluid-structure interaction, where the effect of the control variable has a significant product coupling relationship with the current state variable; for example, the impact of the coal feed rate on the mill bowl differential pressure amplifies nonlinearly with the increase of pulverized coal inventory in the mill. Therefore, when the unit undergoes rapid load adjustments or operates in high-load areas, prediction methods relying solely on linear superposition will produce structural model biases, which are easily misjudged as equipment malfunctions. Implementing an enhanced prediction step based on state-control bilinear coupling aims to introduce physically meaningful nonlinear correction terms, enabling the prediction model to adaptively adapt to changing operating conditions. This significantly improves the fitting accuracy of the actual physical response of equipment under complex operating conditions while maintaining computational efficiency, and reduces false alarms caused by model mismatch. Based on this, an exemplary preferred embodiment of step S51 of this application, based on the system parameter vector, performs theoretical model prediction on the regression data vector based on instantaneous parameters to obtain the theoretical response value, including: A basic linear manifold projection is performed on the system parameter vector and regression data vector to obtain a linear baseline response. Although the system exhibits nonlinearity, linearity still dominates near a small range of steady-state operating points. Through linear manifold projection, a robust anchor point describing the system's basic dynamic behavior near the current operating point can be established, ensuring that the model predictions have a solid physical basis. In practice, the system utilizes the system parameter vector updated online in step S42. The regression data vector constructed in step S33 Perform the vector dot product operation. This operation maps historical input-output information to the current linear response space, and the calculation formula is as follows: ,in, The linear reference response, i.e., the theoretical response value obtained in the first embodiment of step S51, represents the theoretical output value when the system only considers linear characteristics at time k. Taking the differential pressure prediction of a coal mill as an example, the updated parameter vector... (Corresponding to two autoregressive coefficients and two control coefficients), regression data vector (correspond , , , The linear reference response is then calculated as follows: =6.734 kPa.

[0044] A bilinear correction term is calculated from the latest control input and the most recent state feedback in the regression data vector to obtain the bilinear correction term. This step is introduced to explicitly model the product effect between the control input and the system state, compensating for the coupling information lost by linear models that treat them as independent variables. This bilinear structure can quantitatively simulate physical phenomena such as enhanced regulation under high loads. In practice, the system calculates the bilinear correction term from the regression data vector. Extract the currently active valid control commands. and the system status feedback from the previous moment Multiply these two together, and then multiply by a preset nonlinear coupling strength coefficient. Thus, the correction amount is obtained. The calculation formula is as follows: ,in, This is a bilinear correction term; The effective control command after pure time delay correction, i.e., the regression vector. The first item in the control section; For a moment The actual system state feedback value, i.e., the regression vector. The first term of the state part; This is the nonlinear coupling strength coefficient, set based on prior knowledge of the equipment's physics. For example, for objects with strong fluid-structure interaction, such as coal mills, it can be set to... =0.001, used to quantify the strength of the interaction. Continuing with the example above, extract... =50.0, =3.60. Calculate the bilinear correction term: =0.001×(50.0×3.60)=0.001×180=0.18. This value of 0.18 reflects that due to the current high load (large amount of coal, high pressure), the impact of control commands on pressure is amplified by physical mechanisms, which is a part not captured by the linear model.

[0045] Nonlinear augmented synthesis prediction is performed on the linear reference response and the bilinear correction term to obtain the theoretical response value. Finally, nonlinear augmented synthesis prediction is completed to combine the linear and nonlinear components in a physically reasonable and numerically stable manner. The execution process involves combining the bilinear correction term calculated in the previous step. First, use a saturating activation function (such as the hyperbolic tangent function). The data is processed and then added to the linear reference response. Above. Specifically, The function perfectly simulates the soft saturation characteristics of physical devices—for example, the adjustment capability of an actuator has an upper limit, and its effect does not increase linearly indefinitely. This smooth fusion of linear and nonlinear information ultimately generates an enhanced theoretical prediction that highly approximates the actual physical response.

[0046] Continuing the above calculations, )≈0.178. Final theoretical response value =6.734 + 0.178 = 6.912 kPa. Compared to the 6.734 kPa obtained using only the linear model, the enhanced predicted value of 6.912 kPa is closer to the actual physical response under high load conditions. If the actual observed value is 6.912 kPa, the residual of the linear model is 6.920 - 6.734 = 0.186 (which may trigger a false alarm), while the residual of the enhanced model is only 6.920 - 6.912 = 0.008 (which is considered normal), thus significantly improving the robustness of the early warning system.

[0047] In step S6, an adaptive threshold determination based on the statistical characteristics of a sliding window is performed on the residual error value to obtain an abnormal state indicator. That is, simply obtaining the residual signal after filtering out normal operating fluctuations through the model is insufficient to directly trigger an alarm. This is because in the complex electromagnetic and mechanical environment of a thermal power plant, even if the equipment is operating perfectly normally, the calculated residual error value is by no means strictly zero due to factors such as sensor measurement accuracy, signal transmission interference, and minor model mismatches. Instead, it manifests as Gaussian white noise or colored noise that jumps instantaneously around zero. Simply setting a fixed absolute threshold for discrimination easily leads to a dilemma: if the threshold is set too low, false alarms will frequently occur when the model error increases instantaneously due to drastic changes in unit load; if the threshold is set too high, early minor faults cannot be identified during stable low-load operation. Furthermore, the inherent noise levels of different equipment vary and change dynamically with operating time. To this end, this application introduces an adaptive discrimination mechanism based on statistical principles. By learning the distribution pattern of the current residual sequence in real time, a floating threshold closely following the noise envelope is dynamically constructed, thereby effectively suppressing random interference while keenly capturing abnormal states that truly violate statistical laws.

[0048] Specifically, an exemplary implementation of step S6 of this application is as follows: This process aims to transform a hard judgment based on single-point values ​​into a flexible evaluation based on the statistical characteristics of time series. The first step in implementation is to establish a residual statistical buffer mechanism. A fixed-length buffer is allocated in memory... A first-in, first-out (FIFO) queue is used as a sliding window to store residual data calculated at historical time points. Window length. The settings need to balance the stability and sensitivity of the statistics. For an auxiliary control system with a sampling period of 1 second, this application sets... =60, meaning the current signal state is evaluated using historical residual data from the past 60 seconds. Each time step S5 outputs a new residual value... This value is then pushed to the end of the queue, while the oldest data at the head of the queue is... Removed, ensuring the window always contains the most recent content. One sample. In the initial stage, if the queue is not full, preset empirical parameters can be used temporarily or the discrimination logic can be started after the queue is full.

[0049] Then, dynamic threshold calculation is performed. For each sampling time... Read all residual samples within the current sliding window. And calculate its statistical characteristics: mean with standard deviation Mean This reflects the central location of the residual distribution; under ideal fault-free conditions, this value should approach zero; standard deviation. This quantifies the dispersion or fluctuation amplitude of the residual signal, directly characterizing the background noise level under the current operating conditions. Based on the normal distribution... Based on statistical principles, which state that under normal random fluctuations, the probability of a value falling within the range of the mean plus or minus three standard deviations is 99.7%, the algorithm dynamically generates the current upper and lower limit thresholds. The calculation logic follows the formula below: ; ,in, and These are the dynamic upper threshold and the dynamic lower threshold, respectively. This is the sensitivity coefficient, set to 3, which can be fine-tuned based on the tolerance for false alarm rates. This calculation process gives the algorithm strong adaptability to various operating conditions: when the unit is operating in steady state, signal fluctuations are small. The thresholds are also very narrow, allowing for the detection of extremely weak abnormal drifts; however, when the unit undergoes significant load adjustments, the residual fluctuations naturally increase due to the influence of nonlinear model errors. As the threshold increases, the upper and lower limits automatically widen, thus forming a safety envelope that adjusts to the noise level, effectively avoiding false alarms.

[0050] Finally, anomaly detection and status flag generation are performed. The residual error value at the current time is then used. With the dynamic threshold just calculated A comparison is performed. The judgment logic is divided into two levels. The first is an instantaneous comparison: if... It falls within this closed interval, that is If the current residual fluctuation is within the normal range of statistical noise, the equipment is considered to be in a healthy state, and the output abnormal state flag is set to normal or logic 0. Exceeding this range, i.e. or If the data point is not found to be an anomaly candidate, it is considered an anomaly candidate. However, to prevent false alarms caused by single-point glitches due to electromagnetic interference, a de-jittering counter logic is added in the implementation. A real physical anomaly is only confirmed when the residuals of C consecutive sampling times (e.g., C=5) all exceed the dynamic threshold range of their respective times. In this case, the output anomaly status flag is set to an anomaly or logic 1, which can trigger an audible and visual alarm or push a notification to the host computer screen of the maintenance personnel.

[0051] Specifically, under normal operating conditions of the coal mill, due to the white noise of the sensor itself, the residual output in step S5... It fluctuates slightly around 0. Of the past 60 points recorded by the sliding window, most are between -0.02 and 0.02 kPa. The calculated mean within the window... ≈0.001 kPa, standard deviation ≈0.01 kPa. Set the sensitivity coefficient. =3. At this point, the dynamic upper limit threshold... =0.001 + 3 × 0.01 = 0.031 kPa; Dynamic lower limit threshold =0.001-3×0.01=-0.029 kPa. For the load increase scenario mentioned in the first embodiment of step S5, the normally calculated residual... =0.016 kPa. Comparison revealed -0.029 < 0.016 < 0.031, classifying it as normal, with the abnormal state flag set to False. This indicates that although there are fluctuations, they are entirely within the statistically permissible range. Conversely, if an internal blockage occurs in the coal mill, a genuine physical anomaly would lead to residual errors. The value mutated to 0.50 kPa. Comparison revealed that 0.50 > 0.031, a point significantly exceeding the acceptable range. The statistical boundary is marked as an anomaly candidate. If the residual remains at a high level such as 0.48 or 0.52 in the following seconds, although the dynamic threshold calculated each time will be slightly increased due to the entry of outliers (such as a slow increase in standard deviation), due to the inertia of the window, the threshold will still be much smaller than the magnitude of the outlier in the short term. Therefore, the continuous over-limit counter accumulates to meet the condition and finally outputs an anomaly alarm in a robust manner.

[0052] In summary, the abnormal state early warning method for auxiliary control systems of thermal power plants based on the embodiments of this application has been clarified. Addressing the causal decoupling problem caused by neglecting external driving forces in existing technologies, this scheme uses the original control command flow as the system excitation, combines it with feedback sensing flow, and utilizes a pure time delay measurement technique based on mutual information to construct a synchronous input-output dataset that accurately reflects the control-response timing correlation, thereby solving the problem of dynamic characteristic characterization under large hysteresis conditions. To simultaneously resolve the contradiction between false alarms due to equipment aging and missed alarms due to fault adaptation, an online parameter recursive identification algorithm with a forgetting factor is used to construct a time-varying model. The forgetting factor is used to track the slow decay of equipment parameters in real time to adapt to natural wear and tear. Simultaneously, an update enable signal mechanism is introduced, updating model parameters only under normal operating conditions and freezing parameters when potential anomalies occur to prevent the model from learning fault characteristics. Finally, by comparing the theoretical response with the current actual output, an adaptive threshold based on a sliding window is used to determine the residual, achieving highly reliable anomaly early warning.

[0053] Figure 4 This is a block diagram of an abnormal state early warning system for an auxiliary control system of a thermal power plant according to an embodiment of this application. Figure 4As shown, the abnormal state early warning system 100 of the auxiliary control system of a thermal power plant according to an embodiment of this application includes: a raw flow information acquisition module 110, used to acquire raw control command streams and raw feedback sensor streams, wherein the raw control command streams include coal feeder frequency converter commands, blower blade opening commands, and coal mill loading oil pressure setpoints; a raw flow information preprocessing module 120, used to preprocess the raw control command streams and raw feedback sensor streams to obtain a synchronous input / output dataset; and a system time measurement module 130, used to perform system pure time delay measurement based on mutual information on the synchronous input / output datasets to obtain the system pure time delay. The system includes a step count and a regression data vector; an online parameter recursion identification module 140, used to perform online parameter recursion identification with a forgetting factor on the regression data vector, the current actual output in the synchronous input-output dataset, and the update enable signal to obtain the system parameter vector; a residual error calculation module 150, used to perform theoretical response and dynamic residual calculation on the regression data vector and the current actual output in the synchronous input-output dataset based on the system parameter vector to obtain the residual error value; and an abnormal state flag generation module 160, used to perform adaptive threshold determination based on the sliding window statistical characteristics of the residual error value to obtain an abnormal state flag.

[0054] Here, those skilled in the art will understand that the specific operations of each step in the above-mentioned abnormal state early warning system of the auxiliary control system of thermal power plants have been referenced above. Figures 1 to 3 The abnormal state early warning method of the auxiliary control system of thermal power plant has been described in detail, and therefore, its repeated description will be omitted.

[0055] It is understood that the above embodiments are merely exemplary embodiments used to illustrate the principles of this disclosure, and this disclosure is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and substance of this disclosure, and these modifications and improvements are also considered to be within the scope of protection of this disclosure.

Claims

1. A method for early warning of abnormal state of auxiliary control system of thermal power plant, characterized in that, include: Acquire the raw control command stream and the raw feedback sensor stream. The raw control command stream includes the coal feeder frequency converter command, the blower blade opening command, and the coal mill loading oil pressure setpoint. The raw control command stream and raw feedback sensor stream are preprocessed to obtain a synchronous input-output dataset; The system pure time delay is determined based on mutual information on the synchronous input-output dataset to obtain the system pure time delay steps and regression data vector; Online parameter recursion identification with a forgetting factor is performed on the regression data vector, the current actual output in the synchronous input-output dataset, and the update enable signal to obtain the system parameter vector; Based on the system parameter vector, theoretical response and dynamic residual calculation are performed on the regression data vector and the current actual output in the synchronous input-output dataset to obtain the residual error value; An adaptive threshold determination based on the statistical characteristics of a sliding window is performed on the residual error value to obtain an abnormal state indicator.

2. The method for early warning of abnormal state of auxiliary control system of thermal power plant according to claim 1, characterized in that, The raw control command stream and raw feedback sensor stream are preprocessed to obtain a synchronous input / output dataset, including: The original feedback sensing flow is cleaned of feedback signal noise to obtain a cleaned feedback sensing flow. The original control command stream is compensated for dead time based on threshold hysteresis to obtain the compensated control command stream. The compensated control command stream and the cleaned feedback sensor stream are synchronized using zero-order hold multi-source heterogeneous data to obtain a synchronized input-output dataset. The synchronized input-output dataset contains time... Data , For the first The control command value at each sampling time. For the first The sensor feedback value at each sampling time.

3. The method of claim 1, wherein the method further comprises: The system's pure time delay is determined based on mutual information on the synchronous input-output dataset to obtain the system's pure time delay steps and regression data vector, including: Local data window partitioning and joint probability distribution estimation are performed on the synchronous input / output dataset to obtain the joint probability distribution set; The system's pure time delay is obtained by performing optimal time delay search based on the maximum mutual information criterion on the joint probability distribution set. Based on the system's pure time delay steps, time-series alignment and autoregressive vector construction are performed on the synchronous input-output dataset to obtain the regression data vector.

4. The method of claim 1, wherein, Online parameter recursion identification with a forgetting factor is performed on the regression data vector, the current actual output in the synchronous input-output dataset, and the update enable signal to obtain the system parameter vector, including: Kalman gain calculation and prior error estimation are performed on the regression data vector and the current actual output to obtain the Kalman gain vector and prior prediction error; Based on the update enable signal, Kalman gain vector, and prior prediction error, the system parameter vector at the previous time step is updated with model parameter conditions to obtain the system parameter vector.

5. The method of claim 4, wherein the method further comprises: Based on the updated enable signal, Kalman gain vector, and prior prediction error, the system parameter vector from the previous time step is updated using model parameter conditions to obtain the new system parameter vector. This includes updating the system parameter vector from the previous time step using the following formula: in, This is the system parameter vector from the previous time step. The Kalman gain vector. For prior prediction error, To update the enable signal, This is the system parameter vector.

6. The method of claim 1, wherein, Based on the system parameter vector, theoretical response and dynamic residual calculations are performed on the regression data vector and the current actual output in the synchronous input-output dataset to obtain the residual error value, including: Based on the system parameter vector, a theoretical model based on instantaneous parameters is used to predict the regression data vector to obtain the theoretical response value; The observation-prediction difference and residual extraction are performed on the theoretical response value and the current actual output in the synchronous input-output dataset to obtain the residual error value.

7. The method of claim 6, wherein the method further comprises: Based on the system parameter vector, a theoretical model based on instantaneous parameters is used to predict the theoretical response value of the regression data vector, including: A basic linear manifold projection is performed on the system parameter vector and the regression data vector to obtain the linear benchmark response; Bilinear correction terms are calculated by applying bilinear correction terms to the latest control input and the most recent state feedback in the regression data vector; Nonlinear enhanced synthetic prediction is performed on the linear baseline response and the bilinear correction term to obtain the theoretical response value.

8. A power plant auxiliary control system abnormal state early warning system, characterized in that, include: The raw flow information acquisition module is used to acquire the raw control command flow and the raw feedback sensor flow. The raw control command flow includes the coal feeder frequency converter command, the blower blade opening command, and the coal mill loading oil pressure setpoint. The raw flow information preprocessing module is used to preprocess the raw control command flow and raw feedback sensor flow to obtain a synchronous input-output dataset; The system time measurement module is used to measure the pure system time delay based on mutual information of the synchronous input-output dataset to obtain the system pure time delay step count and regression data vector. The online parameter recursive identification module is used to perform online parameter recursive identification with a forgetting factor on the regression data vector, the current actual output in the synchronous input-output dataset, and the update enable signal to obtain the system parameter vector. The residual error calculation module is used to calculate the theoretical response and dynamic residuals of the regression data vector and the current actual output in the synchronous input-output dataset based on the system parameter vector to obtain the residual error value. The abnormal state flag generation module is used to perform adaptive threshold determination on the residual error value based on the statistical characteristics of the sliding window to obtain the abnormal state flag.

Images