Intelligent water level control system for heavy gas generator set waste heat boiler

Abstract

The present application relates to the technical field of automatic control of waste heat boiler of gas generator set, in particular to a water level intelligent control system for waste heat boiler of heavy gas generator set; comprising: an edge computing module receiving first time series data and second time series data, performing differential calculation on the first time series data, and processing the second time series data through a preset digital band-pass filter to extract transient characteristic scalar; a communication bus module receiving the transient characteristic scalar and asynchronously transmitting scheduling instructions containing the scalar; a central processing module receiving the scheduling instructions and third time series data, running a preset state observer, dynamically modulating the observation noise covariance matrix based on the transient characteristic scalar, calculating the state estimation result, generating the target control instruction and outputting it to the controlled equipment for execution; the present application realizes the suppression of false water level, the reduction of the central processing module computing power burden and the improvement of the stability of the feedwater control.

Description

Technical Field

[0001] This invention relates to the field of automatic control technology for waste heat boilers in gas-fired generator sets, specifically to an intelligent water level control system for waste heat boilers in heavy-duty gas-fired generator sets. Background Technology

[0002] Existing waste heat boiler drum water level control systems typically require the collection of time-series data such as pressure, load commands, and liquid level, with a central processing module handling state estimation and feedwater regulation.

[0003] In related technologies, to suppress the influence of false water levels caused by rapid load changes, all high-frequency measurement data from the field can be uploaded to a central processing module for centralized calculation; alternatively, the liquid level measurement data can be filtered and smoothed before generating control commands. However, when the unit experiences rapid load increases and decreases and multiple time-series data are sampled at different frequencies, the former method incurs a significant communication and computational burden, while the latter method is prone to introducing response lag and reducing the accuracy of identifying true water level changes. Therefore, the steam drum water level control methods in related technologies struggle to simultaneously achieve transient disturbance identification capability, real-time control, and accurate state estimation. Summary of the Invention

[0004] To solve the above-mentioned technical problems, the present invention provides an intelligent control system for the water level of waste heat boilers in heavy-duty gas generator sets. Specifically, the technical solution of the present invention includes:

[0005] The edge computing module is used to receive first time-series data as pressure measurement data and second time-series data as load command data, perform differential calculation on the first time-series data to obtain the data change rate, and combine the data change rate with the second time-series data through a preset digital bandpass filter to extract transient feature scalars.

[0006] The communication bus module is connected to the edge computing module and is used to receive transient feature scalars and asynchronously transmit scheduling instructions containing transient feature scalars.

[0007] The central processing module is connected to the communication bus module. It is used to receive scheduling instructions and third-time data as liquid level measurement data, run the preset state observer, and dynamically modulate the observation noise covariance matrix based on the transient characteristic scalar in the scheduling instructions to calculate the state estimation result. Based on the state estimation result, the target control instruction is generated and output to the water supply regulating actuator for execution.

[0008] Preferably, the edge computing module includes: a differential computing unit for performing first-order differential processing on the first time-series data to obtain the data change rate; and a feature extraction unit for inputting the data change rate and the second time-series data into a preset digital bandpass filter with a preset forgetting factor to extract transient feature scalars.

[0009] Preferably, the central processing module includes: a mapping establishment unit for constructing a mapping function between transient feature scalars and the observation noise covariance matrix; and a matrix modulation unit for dynamically modulating the observation noise covariance matrix according to the mapping function, wherein, when the transient feature scalar is greater than or equal to a preset scalar threshold, the noise variance value corresponding to the third time series data in the observation noise covariance matrix is ​​increased to be greater than a preset benchmark variance value, and integral calculation is performed based on the mass conservation prediction model built into the preset state observer; when the transient feature scalar is less than the preset scalar threshold, the noise variance value is reduced to the preset benchmark variance value.

[0010] Preferably, the preset state observer is a Kalman filter; the central processing module further includes: a gain calculation unit, used to calculate the Kalman gain matrix based on the modulated observation noise covariance matrix; and a state update unit, used to update the state of the preset mass conservation prediction model using the Kalman gain matrix and third time series data, so as to output the state estimation result.

[0011] Preferably, the edge computing module operates at a first sampling frequency, and the central processing module operates at a second sampling frequency, wherein the first sampling frequency is greater than the second sampling frequency. The communication bus module uses a dual-rate asynchronous covariance scheduling mechanism to send transient feature scalars to the central processing module. The dual-rate asynchronous covariance scheduling mechanism specifically includes: buffering transient feature scalars according to the first sampling frequency, and extracting the buffered transient feature scalars according to the second sampling frequency, so as to send the transient feature scalars to the central processing module.

[0012] Preferably, the edge computing module further includes: a data denoising unit, used to denoise the first time-series data based on a fast Fourier transform algorithm or a wavelet denoising algorithm before performing first-order difference processing on the first time-series data, so as to remove power frequency interference data and high-frequency white noise data, and retain time-series data of a preset frequency band to reflect the drop trend of the first time-series data.

[0013] Preferably, the feature extraction unit is further configured to, after extracting the transient feature scalar, normalize the transient feature scalar to a preset numerical range, and send the normalized transient feature scalar to the communication bus module.

[0014] Preferably, the preset state observer is a dimensionality-reduced state observer.

[0015] Compared with the prior art, the present invention has the following beneficial effects:

[0016] 1. By adopting the above technical solution, the present invention deploys the high-frequency transient disturbance identification task on the edge computing module side. The edge computing module performs differential calculation on the pressure measurement data, and combines the load command data with the transient feature scalar extracted by the preset digital bandpass filter. Then, it is sent to the central processing module asynchronously through the communication bus module, which effectively avoids the problem of increased communication and computing burden caused by uploading all the original high-frequency time series data in a centralized manner.

[0017] 2. This invention receives scheduling instructions and liquid level measurement data through a central processing module, and dynamically modulates the noise variance corresponding to the liquid level measurement channel in the observation noise covariance matrix based on transient characteristic scalars. This enables the state observer to reduce its dependence on the apparent liquid level during strong transient phases and rely more on the mass conservation prediction model for state estimation. After the disturbance weakens, it resumes the use of the measured liquid level value, ensuring that the false water level surge phase will not mislead the water supply control decision.

[0018] 3. This invention effectively overcomes the shortcomings of centralized high-frequency computing in related technologies and the easy introduction of response lag in simple smoothing of liquid level signals by the synergistic effect of edge-side feature extraction, bus-side lightweight scheduling and central-side covariance adaptive modulation. Ultimately, it can take into account transient disturbance recognition capability, real-time control and state estimation accuracy, thereby improving the stability and reliability of waste heat boiler drum water level control. Attached Figure Description

[0019] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0020] Figure 1 This is a schematic diagram of the module of the intelligent control system for waste heat boiler water level of heavy-duty gas generator set provided in the embodiments of this application. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0022] The intelligent control system for waste heat boiler water level in heavy-duty gas generator sets includes:

[0023] The edge computing module is used to receive first time-series data as pressure measurement data and second time-series data as load command data, perform differential calculation on the first time-series data to obtain the data change rate, and combine the data change rate with the second time-series data through a preset digital bandpass filter to extract transient feature scalars.

[0024] The communication bus module is connected to the edge computing module and is used to receive transient feature scalars and asynchronously transmit scheduling instructions containing transient feature scalars.

[0025] The central processing module is connected to the communication bus module. It is used to receive scheduling instructions and third-time data as liquid level measurement data, run the preset state observer, and dynamically modulate the observation noise covariance matrix based on the transient characteristic scalar in the scheduling instructions to calculate the state estimation result. Based on the state estimation result, the target control instruction is generated and output to the water supply regulating actuator for execution.

[0026] This embodiment provides a distributed time-series data intelligent control mechanism, such as... Figure 1 As shown; specifically, the application scenario is set as the steam drum water level control process of the waste heat boiler of a heavy-duty gas turbine generator set; in this scenario, the gas turbine load may increase from 70% to 100% in a short period of time, and the steam drum pressure, apparent liquid level and feedwater regulating valve opening will change synchronously at different sampling frequencies; the system does not directly send all the raw high-frequency data into the central processing module for unified solution, but first the edge computing module near the field input / output module extracts the transient features that can characterize the strength of false water levels, and then the central processing module adjusts the degree of trust in the liquid level measurement value according to the features, thereby generating stable feedwater control commands;

[0027] Specifically, the edge computing module receives first time-series data and second time-series data. In this application scenario, the first time-series data is specifically configured as a steam drum pressure measurement sequence characterizing the change in boiler internal pressure, and the second time-series data is specifically configured as a gas turbine load command sequence characterizing the expected change in operating conditions. The edge computing module acquires these two data streams at a first preset sampling frequency, for example, once every 5 milliseconds. For the pressure sequence, the edge computing module first performs differential operations to obtain the pressure change rate.

[0028] For ease of explanation, assume the pressure values ​​at four consecutive sampling times are 10.00 MPa, 9.98 MPa, 9.94 MPa, and 9.93 MPa, respectively. The adjacent difference results can be approximated as -0.02, -0.04, and -0.01. If the load command increases from 80 to 92, and then to 100, it can be input into a digital bandpass filter along with the aforementioned difference sequence. This filter does not output a complete waveform, but rather a single-valued transient characteristic scalar, such as 0.15, 0.72, or 0.68. The larger this scalar, the closer the current operating condition is to the false water level stage caused by a sudden pressure change.

[0029] The communication bus module receives the transient feature scalar and encapsulates it into a scheduling instruction, which is then asynchronously transmitted to the central processing module. This scheduling instruction can be understood as a lightweight data message, including at least the feature value, generation timestamp, and corresponding sampling window number. For example, within a 50-millisecond master control cycle, if the edge computing module has generated 10 transient feature scalars, the communication bus does not need to directly transmit all the original pressure points; instead, it only needs to send the valid transient feature scalars within this cycle to the central processing module. This avoids high-frequency raw data accumulating on the bus, reducing the computational load on the central processing module for direct fusion of large-scale multi-source matrices.

[0030] The central processing module runs a preset state observer at a second preset sampling frequency, for example, performing a state update every 50 milliseconds; its input includes at least a scheduling instruction from the communication bus and third time-series data; the third time-series data can be the steam drum liquid level measurement value; the central processing module reads the transient characteristic scalar in the scheduling instruction, and then dynamically modulates the noise weight of the liquid level measurement channel in the observation noise covariance matrix accordingly.

[0031] To illustrate this process, assume that at a certain moment, the state observer initially sets the noise variance of the liquid level channel to 1.0. When a transient characteristic scalar of 0.72 is received, the central processing module determines this value to be in a high-disturbance phase and temporarily increases the noise variance of the liquid level channel to 4.0. At this time, the state observer will reduce its dependence on the apparent liquid level and rely more on the internal mass conservation model to predict the actual water volume change. If the transient characteristic scalar received in the next cycle drops to 0.18, the noise variance of the liquid level channel can fall back to 1.2, and the system will once again enhance its utilization of the measured liquid level value.

[0032] Based on the updated state estimation results, the central processing module generates target control commands and outputs them to the controlled equipment for execution. The controlled equipment can be a feedwater regulating valve actuator, a variable frequency feedwater pump, or a digital controller linked to it. For example, when the estimation results show that the actual water level is 2% lower than the target value, but the current apparent liquid level is falsely displayed as too high due to false expansion, the central processing module will not mistakenly judge it as reducing the feedwater. Instead, it will maintain or increase the feedwater flow rate according to the state estimation results, thereby avoiding a true low water level in the steam drum during the pressure recovery phase.

[0033] As a fault-tolerant processing mechanism, if the edge computing module fails to generate a valid transient feature scalar within a certain sampling window, such as short-term packet loss of pressure data, disordered timestamps, or invalid differential results, the communication bus can send a default scheduling instruction. The central processing module maintains the observation noise covariance matrix of the previous cycle within this cycle, or regresses the baseline value with a preset conservative strategy and a preset step size to avoid sudden changes in the observer parameters exceeding the preset limit due to feature loss. If the third time series data itself is abnormal, such as the liquid level sensor exceeding its range, remaining fixed, or suddenly jumping beyond the physical upper limit, the central processing module can directly set the liquid level channel to a low confidence state, prioritize short-term transition control based on internal state prediction, and wait for subsequent effective measurement recovery.

[0034] For example, within the first 200 milliseconds after the gas turbine executes the 30% load increase command, the steam drum pressure drops rapidly, but the liquid level measurement value shows an upward surge due to the expansion of steam bubbles; the edge computing module extracts a transient feature scalar of 0.76 within the first high-frequency window, and the communication bus sends it to the central processing module with a timestamp.

[0035] After receiving the data, the central processing module temporarily amplifies the noise variance of the liquid level channel, and then mainly updates the actual water level based on the mass conservation relationship corresponding to the water supply and steam volume. The generated control command does not follow the apparent liquid level surge and erroneously close the valve, but instead maintains a smooth transition of the water supply valve opening. After the pressure fluctuation weakens and the characteristic value drops below 0.20, the system gradually restores normal trust in the liquid level sensor.

[0036] The purpose of this step is to decouple the task of identifying high-frequency transient disturbances from the task of estimating low-frequency states in time and space, thereby suppressing false water levels, reducing the amount of data processed by the central processing module, and improving the stability of water supply control.

[0037] Furthermore, the edge computing module includes: a differential computing unit for performing first-order differential processing on the first time-series data to obtain the data change rate; and a feature extraction unit for inputting the data change rate and the second time-series data into a preset digital bandpass filter with a preset forgetting factor to extract transient feature scalars.

[0038] This embodiment provides a refined mechanism for extracting transient features on the edge side. Specifically, in the aforementioned gas turbine load increase scenario, simply stating that extracting transient features from edge nodes is insufficient to support engineering implementation. This is because if the formation method of the data change rate and the method of the filter retaining historical information are not clearly defined, the feature values ​​may overshoot or lag when rapid load increases and decreases occur alternately, affecting the adjustment accuracy of the central side's reliability in liquid level measurement. Therefore, this embodiment further introduces a differential calculation unit and a feature extraction unit with a forgetting factor.

[0039] Specifically, the differential calculation unit performs first-order differential processing on the first time-series data. To avoid slow response at the edges caused by directly using a long window, one implementation method is to use differential sampling between adjacent sampling points; another implementation method is to use sliding differential sampling with an interval of two points, so as to balance noise suppression and sensitivity. Taking pressure measurement as an example, if the pressures corresponding to sampling times t1, t2, t3, and t4 are 10.00, 9.98, 9.94, and 9.93 respectively, then the first-order differential sequence can be recorded as [-0.02, -0.04, -0.01]. In engineering, it can be further divided by the sampling period to obtain an approximate pressure change rate sequence. This change rate is not directly used as a control quantity, but as an intermediate quantity for feature extraction.

[0040] The feature extraction unit sends the aforementioned rate of change and the second time series data together into a digital bandpass filter with a preset forgetting factor. In this process, the rate of change of data is used to reflect the strength of pressure jitter, and the second time series data, i.e. the load command, is used as a multiplicative gain to remove spurious disturbances triggered by non-operating conditions.

[0041] Specifically, the feature extraction unit multiplies the data change rate at each sampling time with the corresponding second time series data element by element to obtain a weighted change rate sequence after gain adjustment, and inputs the weighted change rate sequence into a preset digital bandpass filter; the forgetting factor here is used to reduce the influence of old historical samples on the current feature determination;

[0042] In terms of algorithm structure, let the current sampling window number be... The fundamental characteristic response parameter generated from the weighted rate of change sequence within this window after bandpass filtering is denoted as... The forgetting factor is denoted as The currently extracted transient feature scalar Calculate according to the following formula:

[0043]

[0044] in, This represents the historical transient feature results extracted from the previous sampling window; for example, the base weight of the most recent sampling window can be set to 1.0, and the weight of the previous sampling window can be retained as 0.7, i.e. The residual weight after natural decay of the earlier sampling window is about 0.49. For ease of deduction, assuming that the average absolute values ​​of the pressure change rate in the current three high-frequency windows are 0.04, 0.03, and 0.01, respectively, and the corresponding load command jump variables are 20, 10, and 0, respectively, the feature extraction unit does not simply add them together during calculation, but makes the current window have a greater impact on the output.

[0045] After bandpass processing, the system retains the mid-frequency components related to the false water level, suppresses changes below the preset lower frequency limit and noise changes above the preset upper frequency limit, and finally outputs a transient characteristic scalar. For example, the current window output is 0.78, the residual contribution of the previous window after forgetting is 0.21, the residual contribution of the window before that is 0.04, and the total is 0.83. In this way, even if the load command stops changing at a certain moment, the system can still retain the perception of the pressure transient that just occurred for a short time. However, this retention will gradually decay according to the set factor and will not produce long-term lag effects.

[0046] It is important to emphasize that the reason for introducing the forgetting factor is that if only a memoryless filtering method is used, the transient characteristics may be prematurely judged as ended when the load command has just ended and the two-phase flow effect inside the steam drum has not yet fully subsided, and the central side will prematurely restore the high confidence level of the liquid level measurement. Conversely, if a full history equal weight superposition is used, a transient that has already ended may continue to affect the characteristic value for a period of time exceeding the preset duration, causing the liquid level channel to remain in a confidence state below the preset lower limit for a long time. Through the forgetting mechanism with limited memory, a more stable engineering balance can be achieved between these two types of defects.

[0047] As an anomaly response strategy, if the pressure data at several consecutive sampling points remains constant, resulting in all first-order difference sequences being zero, and the load command does not change significantly, the feature extraction unit can directly output a transient feature scalar close to zero. If the pressure change rate exceeds the preset change rate threshold but the load command does not change, it indicates that the disturbance may come from sensor jitter or non-target process events. In this case, the feature extraction unit can use the frequency band selection capability of the bandpass filter and the attenuation mechanism of the forgetting factor to reduce the probability of false triggering. If the second time series data is missing within a certain window, the previous valid value or the zero-order hold value can be used to participate in the calculation of that window. If the continuous missing data exceeds the preset duration, the edge node can mark the window as low confidence and report it to the central side.

[0048] For example, during the load ramp-up phase of the same waste heat boiler, the edge node obtains a pressure sample every 5 milliseconds; the first to fourth points are 10.00, 9.98, 9.94, and 9.93 respectively, and the differential calculation unit obtains a change rate pattern of first rapid decrease and then slow decrease;

[0049] Meanwhile, the gas turbine load command increases from 82% to 95% within this window; the feature extraction unit inputs these two types of information into a digital bandpass filter with a forgetting factor and outputs a transient feature scalar of 0.83; this value will decay to 0.61 and 0.35 in the next two windows, and will not remain at a high value for a long time unless a new pressure surge occurs again.

[0050] The purpose of this step is to enable the edge side to not only quickly sense sudden changes in pressure, but also to retain the necessary historical inertia in short-term dynamic processes, thereby stabilizing transient feature extraction and improving the scheduling accuracy of the central side.

[0051] Furthermore, the central processing module includes: a mapping establishment unit, used to construct a mapping function between transient feature scalars and the observation noise covariance matrix; and a matrix modulation unit, used to dynamically modulate the observation noise covariance matrix according to the mapping function, wherein, when the transient feature scalar is greater than or equal to a preset scalar threshold, the noise variance value corresponding to the third time series data in the observation noise covariance matrix is ​​increased to be greater than a preset benchmark variance value, and integral calculation is performed based on the mass conservation prediction model built into the preset state observer; when the transient feature scalar is less than the preset scalar threshold, the noise variance value is reduced to the preset benchmark variance value.

[0052] Furthermore, the preset state observer is a Kalman filter; the central processing module also includes: a gain calculation unit, used to calculate the Kalman gain matrix based on the modulated observation noise covariance matrix; and a state update unit, used to update the state of the preset mass conservation prediction model using the Kalman gain matrix and the third time series data, so as to output the state estimation result.

[0053] This embodiment provides a central-side covariance dynamic modulation and state update mechanism. Specifically, in the aforementioned architecture, although the edge nodes have provided transient characteristic scalars, if the central side still uses the same set of observation noise parameters, the liquid level measurement will still be over-relied during the violent disturbance phase, making it difficult to suppress the interference of false water levels on control decisions. Therefore, this embodiment further introduces a mapping establishment unit, a matrix modulation unit, a gain calculation unit, and a state update unit to realize a coherent processing link of eigenvalue change—covariance change—filter gain change—state estimation change.

[0054] Specifically, the mapping unit constructs a mapping function between the transient feature scalar and the observation noise covariance matrix. For ease of engineering deployment, this mapping function can be piecewise rather than a complex continuous high-order function. For example, when the transient feature scalar falls within [0, 0.3), the noise variance of the liquid level measurement channel is set to 1; when it falls within [0.3, 0.6), it is set to 2; and when it is greater than or equal to 0.6, it is set to 5.

[0055] If the observation vector includes three quantities: liquid level, feedwater flow rate, and steam flow rate, then the observation noise covariance matrix can be simplified as a 3×3 diagonal matrix. In a specific example derivation, the reference matrix can be written as diagonal elements [1, 0.5, 0.5]. When the received eigenvalue is 0.75, the matrix modulation unit adjusts it to [5, 0.5, 0.5]. This means that only the confidence level of the liquid level measurement channel is significantly reduced, while the flow rate channel maintains its original confidence level.

[0056] After completing the noise variance modulation, the matrix modulation unit drives the preset state observer to enter the corresponding working mode. When the transient characteristic scalar is greater than or equal to the threshold, the central side determines that it is currently in a stage prone to false water levels. Therefore, it increases the noise variance value corresponding to the third time series data and makes the state observer rely more on the built-in mass conservation prediction model for integral calculation. The preset mass conservation prediction model has a discrete state equation form:

[0057]

[0058] in, This is the current discrete time step number. This is the sequence number of the previous discrete time step; To predict the actual water volume at the current moment, This is an estimate of the actual water volume at the previous moment. For water supply flow rate, For steam flow rate, The integral step size is used here; the mass conservation model can be used to obtain the stock change trend based on the water supply minus the steam supply.

[0059] To illustrate this more clearly, let's assume the estimated actual water volume at the previous moment was 100 units, the water supply for this cycle is converted to +6 units, and the steam volume is converted to -8 units. Then, the actual water volume calculated by integration should be 98 units. Although the level sensor may show a higher reading at this time due to the expansion of steam bubbles, for example, it may seem to correspond to 102 units after conversion, the Kalman filter will reduce the tracking gain of this apparent overshoot because the noise variance of the level channel has been increased.

[0060] When the transient characteristic scalar is less than the threshold, it indicates that the disturbance has weakened or disappeared. The matrix modulation unit reduces the noise variance of the liquid level channel, allowing the state observer to reuse the liquid level sensor to correct the internal model error. For example, if the characteristic value drops from 0.75 to 0.22 within a certain period, the noise variance of the liquid level channel can be reduced from 5 back to 1.2.

[0061] At this point, the gain calculation unit calculates a new Kalman gain matrix based on the modulated covariance matrix, enabling the level measurement to once again play a stronger corrective role in the state estimation; specifically, the Kalman gain matrix... The prior error covariance matrix predicted by the state model Observation mapping matrix and its transpose matrix and the modulated observation noise covariance matrix Together we obtain, and satisfy:

[0062]

[0063] Among them, superscript To represent the transpose of a matrix, the superscript... Represents the matrix inversion operation;

[0064] This processing mechanism shows that during strong transient disturbances... When the variance value representing the uncertainty of the liquid level channel measurement is independently amplified, it can be seen from the monotonic change effect of matrix inversion that the gain coefficient component of the corresponding dimension will shrink to zero; the state update unit uses this Kalman gain matrix and the liquid level data collected in this cycle to update the model state, thereby outputting the weighted estimation result that does not blindly follow the conflict variation characteristics.

[0065] To illustrate the actual effect of gain changes, microscopic numerical simulations can be used. Assume that at a certain moment, the model predicts the actual water level to be 98, while the level sensor measures 101, with a residual of 3. If the noise variance of the level channel is at the baseline value of 1, the Kalman gain can be approximated as 0.7, and the updated state becomes 98 + 0.7 × 3 = 100.1. If the noise variance of the level channel is temporarily amplified to 5, the Kalman gain can be approximated as decreasing to 0.2, and the updated state becomes 98 + 0.2 × 3 = 98.6. A comparison shows that during strong transient phases, the system significantly reduces the influence of the apparent level value on the estimation results, thus avoiding misleading subsequent control commands.

[0066] It should be noted that if the liquid level signal is smoothed directly on the central side with a low-pass filter, although it can suppress the upward surge to a certain extent, it will bring a non-negligible phase hysteresis. Especially during the load decline phase, the rate of decline of the actual low water level exceeds the preset rate of change threshold, and the excessive group delay can easily lead to water replenishment lag. This embodiment adjusts the covariance instead of forcibly smoothing the measured value, so that the liquid level signal still maintains high responsiveness in the steady phase, and can suppress the impact of disturbances by reducing its weight in the transient phase, thus taking into account both anti-disturbance and timeliness.

[0067] In terms of fault tolerance, if the transient characteristic scalar fluctuates repeatedly around the threshold, such as swinging between 0.58 and 0.61 for several consecutive master control cycles, a hysteresis interval can be set to avoid frequent jumps in the covariance matrix; for example, the threshold for entering high noise mode is set to 0.60, and the threshold for exiting high noise mode is set to 0.45.

[0068] If the Kalman filter exhibits non-positive definite covariance, missing input data terms, or residuals exceeding the preset residual limit within a certain period, the central processing module can activate a protection update strategy. This could include freezing the previous effective gain, limiting the single-period state correction magnitude, or performing only the prediction step without performing the measurement update step. If the water or steam input to the mass conservation model is temporarily unavailable, the system can degrade to using the average flow rate of the most recent few periods to estimate short-term stock changes and wait for the real data to recover.

[0069] For example, during the process of the gas turbine rapidly increasing from 70% load to 100%, the transient characteristic scalar reported by the edge node jumps from 0.18 to 0.79; accordingly, the mapping establishment unit increases the level channel noise variance from 1 to 5, and the level gain corresponding to the Kalman filter decreases from about 0.68 to about 0.21.

[0070] Although the level transmitter showed that the water level rose from 0 mm to +35 mm, the status update unit still judged that the actual inventory was decreasing based on the difference between the water supply and steam supply. Therefore, it maintained the output and slightly increased the control tendency of the water supply. After 150 milliseconds, the transient characteristic dropped back to 0.24, the noise variance of the level channel recovered to 1.2, and the filter enhanced the calibration of the level sensor again, finally making the estimated water level smoothly converge to near the actual operating condition.

[0071] The purpose of this mechanism is to convert the identification results of transient disturbances into weight scheduling actions within the observer, thereby achieving digital suppression of false water levels, adaptive adjustment of the balance between model prediction and measured correction, and outputting more reliable state estimation results.

[0072] Furthermore, the edge computing module operates at a first sampling frequency, and the central processing module operates at a second sampling frequency, with the first sampling frequency being greater than the second sampling frequency. The communication bus module uses a dual-rate asynchronous covariance scheduling mechanism to send transient feature scalars to the central processing module. The dual-rate asynchronous covariance scheduling mechanism specifically includes: buffering transient feature scalars according to the first sampling frequency, and extracting the buffered transient feature scalars according to the second sampling frequency, so as to send the transient feature scalars to the central processing module.

[0073] This embodiment provides a dual-rate asynchronous covariance scheduling mechanism. Specifically, in the aforementioned scheme, the edge computing module typically samples at the 5-millisecond level, while the central processing module typically performs state estimation at the 50-millisecond level. If both are forced to use the same sampling frequency, two types of defects will occur: first, if the central processing module is forced to receive all 5-millisecond data, it will substantially increase the bus load and computational load; second, if the edge computing module is forced to reduce its sampling to 50-millisecond, it will miss critical details of rapid pressure changes. Therefore, this embodiment uses high-frequency caching and low-frequency extraction to enable the edge computing module's recognition capability and the central processing module's execution cycle to work together at different rates.

[0074] Specifically, the edge computing module generates transient feature scalars according to the first sampling frequency and writes them sequentially into the buffer. The buffer can be a circular queue of fixed length. For example, if the first sampling frequency is 5 milliseconds and the second sampling frequency is 50 milliseconds, then one central processing cycle corresponds to 10 edge sampling points. If the edge computing module continuously generates feature values ​​[0.12, 0.18, 0.44, 0.71, 0.79, 0.73, 0.62, 0.40, 0.25, 0.20] within this cycle, the communication bus module does not need to upload all values ​​one by one. Instead, it can form a transmission value for covariance scheduling according to the preset extraction rules when 50 milliseconds arrive.

[0075] There are several ways to implement the extraction rules; one way is to take the maximum value in the cache to ensure that the central processing module does not miss short-term strong disturbances; for example, if the maximum value among the above 10 points is 0.79, then 0.79 will be sent in this cycle; another way is to take the last value to reflect the latest state when entering the boundary of the central processing cycle; the last value above is 0.20; there is also a compromise method to take a weighted combination, for example, if the maximum value accounts for 0.6 and the last value accounts for 0.4, then the sent value is about 0.55;

[0076] In practical engineering, to better adapt to the short and drastic characteristics of false water levels, this embodiment preferably adopts a maximum value priority extraction method with an accompanying final value. Correspondingly, the communication bus module configures the data message structure of the scheduling instruction to include: a main feature value, an auxiliary status value, a generation timestamp, and a corresponding sampling window number; that is, the scheduling instruction carries at least a maximum value as the main feature value and a final value as the auxiliary status value; the central processing module can thus sense whether a strong transient has occurred within the period and determine whether it is currently in the decay phase.

[0077] The reason for introducing asynchronous covariance scheduling is that if the edge computing module and the central processing module exchange data only through synchronous blocking, when a short-term congestion occurs on the communication bus, the central processing module will wait for the edge data to arrive, which will lengthen the 50-millisecond control cycle. Conversely, if the central processing module completely ignores the edge tick, it may read null or outdated values ​​in the window where the feature is most critical.

[0078] This embodiment uses caching and asynchronous retrieval to enable the edge computing module to run continuously at a high frequency without stopping due to the busyness or idleness of the central processing module; the central processing module, when a fixed beat arrives, reads the most representative transient feature scalar from the cache in the most recent cycle, thereby maintaining the stability of the control loop rhythm.

[0079] As a fault-tolerant mechanism, if the buffer is empty when a certain central processing cycle arrives, it means that the edge computing module may be temporarily offline or communication is interrupted. At this time, the communication bus module can send an empty flag and an abnormal status code, and the central processing module uses the decayed version of the eigenvalue of the previous cycle as a replacement to avoid the covariance matrix being suddenly reset.

[0080] If the buffer overflows, for example, if the central processing module fails to read the data in time for several consecutive cycles, the system can retain the latest window data and discard the oldest window data, while recording the discard count for subsequent fault diagnosis. If multiple peak values ​​appear within a 50-millisecond cycle, such as [0.20, 0.78, 0.22, 0.81, 0.19…], the maximum value extraction method can ensure that at least one high disturbance is transmitted to the central processing module. If it is necessary to further distinguish between single-peak disturbances and multi-peak disturbances, the number of peak values ​​or the peak position index can be added to the scheduling instruction.

[0081] For example, within 50 milliseconds before the load ramp-up of the same waste heat boiler, the edge computing module generates transient features every 5 milliseconds, namely 0.10, 0.14, 0.30, 0.65, 0.82, 0.76, 0.58, 0.37, 0.24, and 0.18; the buffer completely stores these 10 values; at the 50th millisecond, the communication bus module extracts the main feature value 0.82 according to the central beat, and simultaneously sends it to the central processing module along with the last value 0.18 and the corresponding timestamp; after reading it, the central processing module can determine that the 50-millisecond interval is a stage where a strong transient has occurred and is currently declining, maintain a high level channel noise variance in this cycle, and then gradually revert to the previous cycle based on the new buffer results;

[0082] The purpose of this mechanism is to achieve clock cycle decoupling and lightweight scheduling between the edge computing module and the central processing module without sacrificing the high-frequency sensing capability at the edge, thereby reducing communication and computing pressure and improving the alignment stability between transient features and the master control cycle.

[0083] Furthermore, the edge computing module also includes a data denoising unit, which is used to denoise the first time-series data based on the fast Fourier transform algorithm or wavelet denoising algorithm before performing first-order difference processing on the first time-series data, so as to remove power frequency interference data and high-frequency white noise data, and retain the time-series data of the preset frequency band to reflect the drop trend of the first time-series data.

[0084] This embodiment provides a noise reduction mechanism before edge-side differential calculation. Specifically, in the aforementioned edge extraction process, if the original pressure sequence is directly differentially analyzed, although the pressure drop trend can be rapidly amplified, power frequency interference and quantization noise will also be amplified simultaneously. In the waste heat boiler field, pressure transmitters, analog acquisition cards, and local cables are often accompanied by 50 Hz interference and high-frequency jitter caused by mechanical vibration. Direct differential calculation can easily cause transient characteristic values ​​to be mistakenly raised even under stable operating conditions. Therefore, this embodiment adds a data noise reduction unit before differential calculation.

[0085] Specifically, the data denoising unit can employ either the Fast Fourier Transform (FFT) algorithm or the wavelet denoising algorithm. If the FFT algorithm is used, the edge nodes first perform frequency domain decomposition on the pressure sampling sequence within a short time window, and then suppress known irrelevant frequency bands.

[0086] For example, within a window consisting of 20 sampling points, the system identifies a significant component near 50 Hz, while also exhibiting slight random jitter in frequency bands above 2 Hz. Since the effective frequency band reflecting the pressure drop trend under the target operating condition is mainly between 0.1 Hz and 2 Hz, the data denoising unit can retain the 0.1 Hz to 2 Hz frequency band components, weaken the 50 Hz and higher frequency band components, and then perform an inverse transform to output a smoothed time series. If a wavelet denoising algorithm is used, the high-frequency detail coefficients can be thresholded, suppressing white noise spikes while preserving low-frequency trends and mid-frequency transitions.

[0087] To illustrate more clearly, suppose a certain original pressure window sequence is normalized to [1.00, 0.99, 1.01, 0.97, 0.96]; where 1.01 may be short-term high-frequency pulse interference caused by power frequency superposition, and 0.97 and 0.96 reflect the true drop trend; after denoising, the sequence can be changed to [1.00, 0.99, 0.98, 0.97, 0.96].

[0088] Performing a first-order difference at this point yields [-0.01, -0.01, -0.01, -0.01], which more continuously reflects the downward pressure trend. If the difference is performed directly without denoising, it may yield [-0.01, +0.02, -0.04, -0.01], with positive differences appearing that contradict the actual physical process, thus affecting the stability of subsequent feature extraction.

[0089] The reason for introducing this unit is that although the previous scheme has already alleviated the feature jitter problem by using bandpass filters and forgetting factors, if the original input itself contains strong power frequency and high-frequency white noise, the differential step will amplify the noise at an earlier stage, and subsequent filtering can only remedy the situation afterward. The cost is that the filtering intensity needs to be increased, which may further reduce the sensitivity of the true transient. This embodiment moves the denoising to before the differential step, eliminates the interference components in the input signal through preprocessing, and then extracts the rate of change, which is more conducive to maintaining the data form with weak noise and strong trend.

[0090] As a fault-tolerant processing mechanism, if the noise level at the scene is low, the data denoising unit can use the first preset threshold setting to avoid excessive smoothing that weakens the real abrupt changes; if the fast Fourier transform window length is less than the preset lower limit, resulting in insufficient frequency resolution, it can automatically switch to wavelet denoising mode.

[0091] If the wavelet threshold is greater than the preset upper limit and the actual drop edge is flattened, it can be rolled back to the previous stable parameter combination; if a sudden isolated spike appears in the pressure sequence within a certain window and the amplitude exceeds the reasonable range of the instrument, the data denoising unit can first limit the amplitude and then enter the frequency domain or wavelet processing to prevent the anomaly from destroying the denoising result of the entire window.

[0092] For example, during the load ramp-up phase after gas turbine ignition, 50 Hz electromagnetic interference is superimposed on the original steam drum pressure sampling sequence, causing local samples to exhibit non-realistic fluctuation characteristics of drop-rebound-drop again. The data denoising unit at the edge node first performs frequency domain filtering on this segment of the sequence, retaining the trend component from 0.1 Hz to 2 Hz. The differential calculation unit obtains continuous negative differences from the denoised sequence, and the feature extraction unit outputs a more stable transient feature scalar accordingly. In this way, the dispatch instructions received by the central side can better reflect the real pressure drop, rather than the pseudo-transient constructed by electrical noise.

[0093] The purpose of this step is to suppress frequency components unrelated to the target operating condition before differential amplification, thereby enhancing the noise resistance of transient feature extraction and reducing misscheduling caused by noise.

[0094] Furthermore, the feature extraction unit is also used to normalize the transient feature scalar to a preset numerical range after extracting the transient feature scalar, and send the normalized transient feature scalar to the communication bus module.

[0095] This embodiment provides a transient feature scalar normalization mechanism. Specifically, in the same control system, the absolute amplitude of the pressure change rate may differ between different units, pressure transmitters with different ranges, and even the same unit in different seasons and load ranges. If edge nodes directly upload feature values ​​that have not been processed with a unified scale, it is difficult for the central side to use a stable covariance mapping rule, which can easily lead to problems such as features being too small and unable to trigger under certain operating conditions, and features being too large and saturating for a long time under other operating conditions. Therefore, this embodiment adds a normalization step after feature extraction.

[0096] Specifically, after generating the original transient feature scalar, the feature extraction unit maps it to a preset numerical range, such as [0,1]. A simplified implementation is to use a linear normalization algorithm with a boundary-crossing prevention strategy: the preset lower bound is mapped to 0, the preset upper bound to 1, and the mapping is performed proportionally within the range. From the perspective of numerical processing rules, when the original transient feature scalar is denoted as... The preset lower bound is denoted as The upper bound is preset as And ensure during system initialization Under the condition, the normalized transient characteristic scalar It can be expressed by the following formula:

[0097]

[0098] in, To find the maximum value function, To find the minimum value function; the above formula, while achieving a linear mapping, also... and Boundary constraints avoid spillover risks; if the denominator approaches zero, a minimal positive complement constant can be automatically added during the underlying implementation.

[0099] To illustrate its effect, assume that the system, based on historical sample statistics, considers the original feature value of 0.00 as a stationary baseline and 1.50 as a strong transient upper bound. If the current original feature value is 0.30, the normalized result after substitution is 0.20; if it is 0.90, the normalized result is 0.60; if it is 1.80, since it exceeds the upper bound, the normalization is truncated to 1.00 under the boundary protection mechanism. After adopting this boundary normalization, the central processing module only needs to deal with feature values ​​with uniform dimensions and strict constraints within a specific range, thus making it easier to formulate stability criteria such as entering a strong transient mode when the value is above 0.6.

[0100] The reason for introducing this step is that the aforementioned scheme can extract transient features, but if different edge nodes use different scales for output, for example, one node outputs in the range of 0 to 2.5 and another node outputs in the range of 0 to 0.3, then the same set of thresholds and mapping functions on the central side will not be universal; normalization is equivalent to completing scale alignment on the edge side first, so that the communication bus transmits values ​​with consistent interpretation, rather than the original amplitude directly affected by the range and noise limit of the field instruments;

[0101] During the normalization process, a certain dynamic boundary update capability can be retained. For example, the system can maintain the distribution of effective features over a recent period of operation. When long-term stable operation reveals that the actual original features are almost no more than 1.0, the upper bound can be lowered to 1.2 by a preset fixed step size to improve the resolution for moderate disturbances. Conversely, if effective features above 1.5 frequently appear, the upper bound can be increased according to the first preset adjustment rate to avoid a large number of samples being squeezed around 1 and losing their discriminative ability. However, this update should not exceed the preset rate threshold to prevent the boundary from fluctuating beyond the preset amplitude due to short-term noise.

[0102] Regarding the handling mechanism for abnormal boundaries, if the upper and lower bounds required for normalization are equal or too close, it may lead to an excessively small denominator and amplify numerical fluctuations. In this case, the feature extraction unit can switch to using a fixed default boundary or directly output the normalization rule of the previous stable period. If the original feature value is negative, and the preset design only targets non-negative scalars, then the absolute value or lower bound truncation can be performed first. If the normalized value is 1 for several consecutive periods, it indicates that the upper bound may be set too low or there may be continuous strong disturbances in the field. A saturation flag should be attached to the central side to facilitate the adoption of a more conservative covariance modulation strategy.

[0103] For example, when the same waste heat boiler is running at high load in summer, the pressure change rate is large, and the original transient characteristics are often distributed between 0.2 and 1.4; while in winter at low load steady state, the original characteristics are mostly distributed between 0 and 0.5. After the edge nodes uniformly map the original characteristics in both cases to the [0,1] interval, the central side can use consistent threshold logic regardless of the current absolute pressure level: for example, 0.00 to 0.30 is regarded as the stable area, 0.30 to 0.60 is regarded as the transition area, and above 0.60 is regarded as the strong transient area; in this way, the communication bus and the main control central processor process scheduling instructions of the same scale.

[0104] The purpose of this step is to unify the dimensions and constrain the amplitude of the transient features extracted from the edges, thereby standardizing the central side mapping rules and improving reusability under different working conditions and equipment configurations.

[0105] Furthermore, the preset state observer is a reduced-dimensional state observer.

[0106] This embodiment provides a data object limitation and dimensionality reduction state observation mechanism for waste heat boiler water level control scenarios. Specifically, in the aforementioned general description, the first time series data, the second time series data, the third time series data, and the state observer can all be selected from multiple sources. In order to make the scheme more suitable for the actual application of waste heat boilers in heavy-duty gas turbine generator sets, this embodiment specifies that: the first time series data adopts steam drum pressure measurement data, the second time series data adopts gas turbine load command data, the third time series data adopts steam drum liquid level measurement data, and the state observer adopts a dimensionality reduction state observer.

[0107] Specifically, pressure measurement data was chosen as the core input for edge feature extraction because, in scenarios of sudden load changes, changes in drum pressure usually precede the stable recovery of apparent liquid level, providing a feedforward indicator for false water levels. Load command data was chosen as the second input because the same pressure fluctuation has different technological implications depending on whether there is a load command change or not. Introducing load commands helps the edge computing module distinguish between legitimate transients caused by unit scheduling and occasional fluctuations caused by random disturbances. Liquid level measurement data was chosen as the third input because liquid level is still a direct observation that ultimately needs to be corrected and utilized; its reliability only needs to be temporarily lowered at a specific stage, rather than being completely discarded.

[0108] For the state observer, this embodiment adopts a dimensionality reduction approach because if all mechanistic variables such as the steam drum, evaporator, economizer, and superheater are included in a unified high-dimensional state space, the central processing module will find it difficult to stably complete the calculation within a 50-millisecond control cycle; the dimensionality-reduced state observer can focus on a small number of core states that are directly related to water level control.

[0109] Specifically, the dimensionality-reduced state observer internally maintains a three-dimensional state vector. ,in, This represents the equivalent water stock status. This indicates a shift in the liquid level observation status. The superscript represents the cumulative net difference between the feedwater and steam flows. This represents the transpose of a vector or matrix; thus, the central processing module only needs to perform prediction and update on the above three-dimensional state vector in each cycle, significantly reducing the computational pressure to within the preset computing power threshold range.

[0110] Taking a specific numerical extrapolation as an example, assume that the three-dimensional state predicted by the dimensionality reduction observer in a certain period is [100, +3, -2], where 100 represents the estimated inventory, +3 represents the liquid level observation offset trend, and -2 represents the downward trend caused by the net flow difference; the value measured by the liquid level sensor is displayed as 103 after conversion, but the edge computing module simultaneously reports a normalized transient feature of 0.81;

[0111] Based on this, the central processing module reduces its confidence in the measured liquid level value and mainly corrects the inventory status and net flow status during status updates, while conservatively updating the liquid level deviation status; in this way, even if the apparent liquid level temporarily surges, the estimated value of the actual inventory status can still converge to the actual value.

[0112] The use of a dimensionality-reduced state observer can also be combined with the aforementioned dual-rate scheduling mechanism; the edge computing module is responsible for high-frequency identification of pressure transients, while the central processing module only performs low-frequency estimation of a small number of key states. This avoids the computational bottleneck of centralized high-dimensional models on industrial control computers and reduces the difficulty of model parameter calibration. Especially in the transformation projects of old distributed control systems or programmable logic controllers, it is not necessary to reconstruct the entire mechanism model. Stable control of the target state can be achieved by using only a small number of key states and covariance modulation logic.

[0113] As a fault-tolerant mechanism, if the pressure measurement data fails, the edge computing module cannot reliably generate false water level intensity features. In this case, the system can revert to a conservative control mode based only on liquid level and flow rate. If the load command data is delayed or lost, the edge computing module can first maintain the most recent effective load command for a short period of time, and reduce the feature reliability after continuous loss for more than a set time.

[0114] If the liquid level measurement data is severely distorted, the central processing module can temporarily freeze the liquid level offset state in the dimensionality reduction observer and rely only on the inventory and net flow difference for transition estimation. If the deviation between the dimensionality reduction model and the field conditions increases, such as changes in the characteristics of the water supply system after maintenance, it can be quickly restored by recalibrating a small number of state transition parameters without rebuilding the high-dimensional model.

[0115] For example, when implemented on the same triple-pressure reheat waste heat boiler, the edge computing module consistently collects the steam drum pressure and gas turbine load commands, while the central processing module consistently receives the steam drum liquid level measurement value; when the gas turbine load jumps from 75% to 95%, the edge computing module extracts high-intensity transient features from the pressure measurement data and sends them to the central processing module;

[0116] The reduced-dimensional state observer running in the central processing module only maintains three core states: equivalent inventory, liquid level offset, and net flow difference. It can complete prediction, gain calculation, and state update within a 50-millisecond cycle, and finally output the water supply valve opening adjustment amount. The whole process retains the ability to identify false water levels and meets the real-time data processing requirements of industrial controllers.

[0117] The purpose of this mechanism is to limit the data input objects and state estimation objects to the key quantities that best reflect the target operating conditions, thereby reducing model complexity, ensuring real-time performance, and enabling direct engineering adaptation in the waste heat boiler water level control scenario.

[0118] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A smart control system for water level in a waste heat boiler of a heavy-duty gas generator set, characterized in that, The system includes: an edge computing module, used to receive first time-series data as pressure measurement data and second time-series data as load command data, perform differential calculation on the first time-series data to obtain the data change rate, and combine the data change rate with the second time-series data through a preset digital bandpass filter to extract transient feature scalars; A communication bus module, connected to the edge computing module, is used to receive the transient feature scalar and asynchronously transmit scheduling instructions containing the transient feature scalar. The central processing module, connected to the communication bus module, is used to receive the scheduling instructions and third time-series data as liquid level measurement data, run a preset state observer, and dynamically modulate the observation noise covariance matrix based on the transient characteristic scalar in the scheduling instructions to calculate the state estimation result, generate a target control instruction based on the state estimation result, and output the target control instruction to the water supply regulating actuator for execution.

2. The intelligent control system for waste heat boiler water level of heavy-duty gas generator set according to claim 1, characterized in that, The edge computing module includes: a differential computing unit for performing first-order differential processing on the first time-series data to obtain the data change rate; and a feature extraction unit for inputting the data change rate and the second time-series data into the preset digital bandpass filter with a preset forgetting factor to extract the transient feature scalar.

3. The intelligent control system for waste heat boiler water level of heavy-duty gas generator set according to claim 1, characterized in that, The central processing module includes: a mapping establishment unit, used to construct a mapping function between the transient feature scalar and the observation noise covariance matrix; and a matrix modulation unit, used to dynamically modulate the observation noise covariance matrix according to the mapping function, wherein, when the transient feature scalar is greater than or equal to a preset scalar threshold, the noise variance value corresponding to the third time-series data in the observation noise covariance matrix is ​​increased to be greater than a preset benchmark variance value, and integral calculation is performed based on the mass conservation prediction model built into the preset state observer; when the transient feature scalar is less than the preset scalar threshold, the noise variance value is reduced to the preset benchmark variance value.

4. The intelligent control system for waste heat boiler water level of heavy-duty gas generator set according to claim 3, characterized in that, The preset state observer is a Kalman filter; The central processing module further includes: a gain calculation unit, used to calculate the Kalman gain matrix based on the modulated observation noise covariance matrix; The state update unit is used to update the state of the preset mass conservation prediction model using the Kalman gain matrix and the third time series data, so as to output the state estimation result.

5. The intelligent control system for waste heat boiler water level of heavy-duty gas generator set according to claim 1, characterized in that, The edge computing module operates at a first sampling frequency, and the central processing module operates at a second sampling frequency, wherein the first sampling frequency is greater than the second sampling frequency. The communication bus module uses a dual-rate asynchronous covariance scheduling mechanism to send the transient feature scalar to the central processing module. The dual-rate asynchronous covariance scheduling mechanism specifically includes: caching the transient feature scalar according to the first sampling frequency, and retrieving the cached transient feature scalar according to the second sampling frequency, so as to send the transient feature scalar to the central processing module.

6. The intelligent control system for waste heat boiler water level of heavy-duty gas generator set according to claim 2, characterized in that, The edge computing module further includes a data denoising unit, which is used to denoise the first time series data based on a fast Fourier transform algorithm or a wavelet denoising algorithm before performing first-order difference processing on the first time series data, so as to remove power frequency interference data and high-frequency white noise data, and retain time series data of a preset frequency band to reflect the drop trend of the first time series data.

7. The intelligent control system for waste heat boiler water level of heavy-duty gas generator set according to claim 2, characterized in that, The feature extraction unit is further configured to, after extracting the transient feature scalar, normalize the transient feature scalar to a preset numerical range, and send the normalized transient feature scalar to the communication bus module.

8. The intelligent control system for waste heat boiler water level of heavy-duty gas generator set according to claim 1, characterized in that, The preset state observer is a dimensionality-reduced state observer.

Images