Centrifugal fan kiln head wind remote precise regulation system

Abstract

This invention discloses a remote precision control system for centrifugal fan kiln head air. This invention relates to the field of industrial control technology and includes a sensing layer, an edge computing layer, a control decision layer, and an execution layer. The sensing layer includes a multi-band industrial near-infrared camera installed at the fire observation hole of the rotary kiln head. The multi-band industrial near-infrared camera is equipped with a near-infrared (NIR) band sensor and a short-wave infrared (SWIR) band sensor, and is externally fitted with a high-temperature resistant cooling water jacket. This invention, through NIR and SWIR dual-band collaborative sensing and adaptive spatiotemporal joint filtering, can stably and quantitatively extract three core combustion state characteristic parameters—blackhead length, flame stiffness index, and flame divergence half-angle—in high-temperature and high-dust environments. The sensing availability rate is significantly improved compared to traditional differential pressure transmitter solutions, fundamentally solving the problem of traditional sensors easily failing under harsh operating conditions, and providing reliable quantitative input information for subsequent control decisions.

Description

Technical Field

[0001] This invention relates to the field of industrial control technology, specifically a remote and precise control system for centrifugal fan kiln head air. Background Technology

[0002] The rotary kiln is the core thermal equipment for clinker calcination in cement production. The highest operating temperature inside the kiln can reach 1450℃. The kiln head combustion system consists of a centrifugal fan, a multi-channel pulverized coal burner, and related pipelines and valves. The primary air output from the centrifugal fan is injected into the kiln through two channels: an internal air channel and an external air channel. The internal air generates a rotating airflow through the swirl blades, forming a recirculation zone at the burner outlet, which helps stabilize the flame and promote the ignition of pulverized coal. The external air is injected in a direct stream, enhancing the flame penetration and allowing the flame to reach deeper into the kiln. The ratio of the internal to external air flow, i.e., the swirl ratio, has a decisive influence on the flame morphology and the stability of the thermal regime inside the kiln, and is one of the most important adjustment methods in kiln process control.

[0003] Flame morphology is a direct and comprehensive indicator of the combustion state and clinker calcination quality within the kiln. An excessively long black flame head indicates delayed ignition, which will reduce clinker production and increase fly sand. Insufficient flame rigidity leads to flame deflection, causing localized high-temperature damage to the kiln lining and shortening the lifespan of refractory bricks. An excessively large flame divergence angle may cause abnormal expansion of the recirculation zone, resulting in localized oxygen deficiency, increased nitrogen oxide emissions, and failure to meet environmental regulations. Therefore, real-time and precise control of the primary air at the kiln head is a core technical requirement for ensuring the safe, efficient, and low-emission operation of the rotary kiln.

[0004] Existing technologies have the following three main shortcomings in achieving precise control of primary air at the kiln head: Firstly, the reliability of the sensing layer is insufficient. The ambient temperature at the kiln head exceeds 1200℃, and the dust concentration remains high for a long time. Moreover, the dust contains sticky components such as calcium oxide and silicon oxide. The air volume measurement instrument based on the differential pressure transmitter is prone to scaling and blockage due to the pressure measuring tube being subjected to alternating effects of dust scouring and high-temperature baking for a long time, which leads to measurement distortion or even complete failure. Ordinary visible light industrial cameras are affected by the strong scattering effect of dust on the visible light band. In the high dust environment at the kiln head, the image is blurred and cannot stably identify the flame boundary. Furthermore, it is impossible to quantitatively extract combustion state characteristic parameters such as black flame length, flame stiffness index, and flame divergence half angle. This causes the controller to lose accurate process feedback, the control precision to drop significantly, energy consumption to increase, and product quality fluctuations to intensify.

[0005] Secondly, the control of internal / external air lacks a physical mapping model and relies heavily on human experience. The influence mechanism of internal and external air on flame morphology is complex, with a strong nonlinear coupling relationship between the two. It also interacts with multiple operating variables such as pulverized coal particle size, pulverized coal flow rate, and negative pressure inside the kiln, making it difficult to describe with a simple linear model. Existing plants completely lack a physical model that can quantitatively map observable flame characteristics to fan control commands, making it impossible to achieve automatic intelligent decoupling control of internal / external air. Operators have to rely on subjective experience to adjust the air using a trial-and-error method, resulting in serious lag in the adjustment process and frequent over-adjustment and under-adjustment, leading to low combustion efficiency, difficulty in ensuring production continuity, and the quality of adjustment is highly dependent on the personal ability of experienced operators.

[0006] Third, random network delays cause control oscillations, seriously threatening the security of remote control. Remote control commands must be transmitted from the central control center to the field programmable logic controller via the public network, resulting in random time-varying delays on the order of 100ms to 500ms, as well as delay jitter. When the control commands calculated and issued by the controller based on the state information at a certain moment reach the field actuators, the actual operating conditions have already deviated due to the natural response of the system during the delay period. The control system is essentially driving the current actuator based on the outdated state. For centrifugal fan systems that already have large inertia and pure time delay characteristics, random delays are equivalent to introducing additional uncertainties and pure time delays into the control loop, which can easily cause continuous oscillations or even instability, posing risks of equipment damage and major safety accidents. This is a major technical obstacle to the promotion of remote centralized control. Summary of the Invention

[0007] This invention aims to solve the following three technical problems: First, to achieve stable, real-time, and quantitative non-contact sensing of combustion state characteristic parameters in a high-temperature and high-dust kiln head environment; Second, to establish a direct quantitative physical mapping from observable flame characteristics to fan control commands, realizing real-time online intelligent decoupled optimization control of internal / external air; Third, to ensure the asymptotic stability of remote closed-loop control of large-inertia centrifugal fans under the random latency conditions of OT / IT converged networks.

[0008] This invention provides a remote precision control system for centrifugal fan kiln head air, comprising a sensing layer, an edge computing layer, a control decision layer, and an execution layer. Each layer interacts with data via industrial Ethernet to form a complete remote precision control closed-loop system.

[0009] The sensing layer includes a multi-band industrial near-infrared camera installed at the inspection hole of the rotary kiln head. The camera is equipped with a near-infrared (NIR) band sensor and a short-wave infrared (SWIR) band sensor, and is fitted with a high-temperature resistant cooling water jacket to withstand the high-temperature working environment of the kiln head above 1200°C.

[0010] The edge computing layer includes edge computing nodes deployed in the kiln head on-site cabinet. The edge computing nodes sequentially perform adaptive spatiotemporal joint filtering to remove dust occlusion preprocessing, multi-spectral band adaptive weight fusion to enhance the signal-to-noise ratio, improved U-Net multi-scale semantic segmentation to accurately extract the flame region, and quantitative calculation of three combustion state feature parameters: black flame head length, flame stiffness index, and flame divergence half angle. The original video stream is converted into a structured feature vector F and uploaded to the control decision layer with a timestamp.

[0011] The control decision layer runs three core modules on the central control center server: the PINN decoupling optimization model, the network delay state observer, and the adaptive generalized prediction compensation controller. These three modules work together to generate control command frames with precise timestamps, which are then sent to the execution layer via the network.

[0012] The PINN decoupled optimization model takes the feature vector F and the operating condition vector W as inputs, and uses the combustion dynamics and fluid dynamics physical equations as penalty terms to constrain the network output, outputting the internal wind setpoint and the external wind setpoint. The network delay state observer uses a finite-state Markov chain to model the network delay from the central control center to the field and predict the future delay distribution. The adaptive generalized predictive compensation controller dynamically adjusts the prediction step size and control step size based on the predicted future time delay distribution. After identifying the centrifugal fan process model online, it solves the optimal control increment sequence and sends the control command frame with timestamp to the execution layer.

[0013] The execution layer consists of a field-programmable logic controller, a frequency converter, an internal butterfly valve, and an external butterfly valve.

[0014] The field programmable logic controller receives control command frames, verifies the timestamp, and determines whether to activate the remote status observer compensation mechanism based on the time delay prediction error. It then drives the frequency converter to adjust the total air volume, the internal air butterfly valve to adjust the internal air volume, and the external air butterfly valve to adjust the external air volume. The execution results are reported to the central control center to form a closed loop.

[0015] Compared with the prior art, the beneficial effects of the present invention by adopting the above technical solution are as follows: This invention utilizes dual-band NIR and SWIR sensing combined with adaptive spatiotemporal filtering to stably and quantitatively extract three core combustion state characteristic parameters—blackhead length, flame stiffness index, and flame divergence half-angle—in high-temperature and high-dust environments. The sensing availability is significantly improved compared to traditional differential pressure transmitter solutions, fundamentally solving the problem of traditional sensors being prone to failure under harsh conditions and providing reliable quantitative input information for subsequent control decisions.

[0016] This invention embeds three types of physical priors—mass conservation constraints, swirl ratio-ignition distance relationship constraints, and value range constraints—into a neural network loss function using PINN. This enables a direct quantitative physical mapping from observable flame conditions to fan control commands, completely replacing manual experience intervention. It can improve combustion efficiency by 3% to 8%, reduce nitrogen oxide emissions by 10% to 20%, and extend kiln lining life.

[0017] This invention guarantees the asymptotic stability of the remote control loop of a centrifugal fan under random time-varying network delay conditions in the range of 100ms to 500ms by employing a dual-end collaborative mechanism of Markov chain online delay modeling, adaptive generalized prediction compensation controller dynamic prediction step size adjustment, and field programmable logic controller remote state observer compensation. This supports safe operation of unattended remote centralized control. Attached Figure Description

[0018] Figure 1 This is a diagram showing the overall architecture of the centrifugal fan kiln head air remote precision control system of the present invention; Figure 2 This is a flowchart illustrating the entire process of implementing the system of the present invention; Figure 3 This is a flowchart of the multi-band infrared flame feature extraction algorithm of the present invention; Figure 4 This is a flowchart illustrating the offline pre-training and online inference process of the PINN model of this invention. Figure 5 This is a flowchart of the adaptive generalized predictive compensation control algorithm of the present invention. Detailed Implementation

[0019] The centrifugal fan kiln head air remote precision control system of this invention is physically deployed into four layers: a sensing layer, an edge computing layer, a control decision layer, and an execution layer. These layers interact via industrial Ethernet, forming a complete remote precision control closed-loop system. The overall system architecture is as follows: Figure 1-2 As shown.

[0020] The multi-spectral industrial near-infrared camera in the perception layer is installed at the kiln head inspection port, equipped with a high-temperature resistant cooling water jacket for thermal protection. The video stream is transmitted to the edge computing layer via gigabit industrial Ethernet fiber optic cable. The edge computing layer is deployed in the kiln head field cabinet and is supported by an industrial-grade GPU embedded platform with a computing power of no less than 8 TOPS. It sequentially completes adaptive spatiotemporal joint filtering to remove dust obstruction preprocessing, multi-spectral adaptive weight fusion to enhance the signal-to-noise ratio, improved U-Net multi-scale semantic segmentation to accurately extract the flame region, and quantitative calculation of three core feature parameters. The original video stream is converted into a structured feature vector with a timestamp and uploaded to the central control center. The control decision layer runs the PINN decoupling optimization model, network latency state observer, and adaptive generalized prediction compensation controller on the central control center server. The three work together to generate control command frames with precise timestamps, which are then sent to the execution layer via the network. The field programmable logic controller in the execution layer completes the timestamp verification and latency compensation of the state observer, driving the frequency converter and internal / external butterfly valves to implement precise airflow adjustment. The execution results are reported to the central control center in real time, forming a complete physical closed loop.

[0021] Module 1: Intelligent Flame Morphology Sensing Based on Multi-Band Infrared Fusion In terms of hardware configuration, a dedicated multi-band industrial-grade near-infrared camera is installed at the kiln head inspection port. This camera is equipped with two image sensors: NIR band (0.9μm to 1.7μm) and SWIR band (1.7μm to 2.5μm), supporting dual-band synchronous exposure acquisition with a time synchronization error of less than 1ms. The NIR band has rich texture information for the core area of ​​the high-temperature flame, while the SWIR band has a better penetration ability for water vapor and fine dust particles than the NIR band. The dual-band collaborative acquisition can obtain complementary and effective image information in high-temperature and high-dust environments. The camera is equipped with a dedicated high-temperature resistant stainless steel cooling water jacket, and the cooling water circulation temperature is maintained below 60℃, ensuring that the camera can work stably for a long time in the high-temperature environment of over 1200℃ at the kiln head. The camera transmits the original video stream (25fps, resolution 1280×1024 pixels) to the edge computing node installed in the on-site cabinet at the kiln head via industrial gigabit Ethernet fiber optic (transmission latency less than 1ms).

[0022] The complete processing chain of multi-spectral band adaptive spatiotemporal filtering and feature extraction algorithm is as follows: Figure 3 As shown, the specific implementation is as follows: The first step involves the synchronous acquisition of multi-spectral raw images. Edge computing nodes control the NIR and SWIR sensors of the multi-spectral industrial camera to acquire raw image frames simultaneously via hardware triggers, which are denoted as NIR band image frames. and SWIR band image frames ,in Image pixel coordinates ( =1280、 =1024 (unit: pixels) The current acquisition time (unit: s) is used to store the acquired image frames in the frame buffer queue in a first-in, first-out manner. The queue length is... =5 frames, used to support subsequent temporal filtering processing.

[0023] The second step is temporal adaptive inter-frame difference filtering. For image frames in both the NIR and SWIR bands, the following operations are performed: Calculate the pixel-wise absolute difference between the current frame and the previous frame, where the subscript... Indicates the band type. =1 / 25s is the time interval between adjacent frames: ; Calculate the global mean of the difference map to estimate the dust motion intensity in the current frame: ; Dynamically update the motion dust detection threshold using an exponentially weighted moving average strategy: ; in =0.3 is the forgetting factor, which enables the threshold to adaptively track the dynamic changes in dust concentration, and to detect dust exceeding the threshold in the difference graph. The pixel locations are marked as areas occluded by moving dust, and temporal mean replacement is performed on the original image pixels of the occluded areas: ; Replacement is performed only on pixels marked as obscured by moving dust; the remaining pixels retain their original values ​​to eliminate high-frequency flicker noise caused by dust movement, while preserving the true brightness information of the flame area. The output is a temporally filtered image frame pair. and .

[0024] The third step is spatial anisotropic diffusion filtering, applied to the time-domain filtered image. Applying the Perona-Malik anisotropic diffusion filter, the iterative update equation is: ; in The initial input image, =0.25 is the time step (satisfying the numerical stability condition). The image gradient vector, For gradient magnitude, For the divergence operator, the diffusion coefficient function is defined as: ; in The gradient-sensitive parameter is adaptively set based on the contrast between the flame and dust in the current frame, with a typical value of 30. This function's characteristics result in a large gradient magnitude and a diffusion coefficient approaching 0 at the flame boundary (boundary preservation), while a small gradient magnitude and a diffusion coefficient approaching 1 in the uniform dust region (strong smoothing and denoising). The total number of iterations is fixed at 10, outputting smooth image frame pairs with preserved boundaries. and .

[0025] The fourth step is multi-spectral image adaptive weight fusion, where the local information entropy of the NIR and SWIR filtered images is calculated in 8×8 pixel blocks: ; in In pixels Gray levels within the 8×8 block centered on Normalized frequency, =10 -8 To avoid small positive numbers overflowing from logarithmic calculations, the adaptive fusion weights for each band are normalized: ; Perform linear fusion by weights to output an enhanced fused image: ; This strategy tends to use NIR images in the high-temperature flame core region (where NIR information entropy is higher and texture information is richer), and tends to use SWIR images in the dust-rich region (where SWIR has stronger penetration and higher information entropy), thus achieving adaptive optimal information fusion.

[0026] The fifth step is to improve U-Net's multi-scale semantic segmentation by fusing the images. After normalization to the [0,1] interval, the improved U-Net network model is input. This network is pre-trained offline based on manually annotated flame area samples from historical DCS data of the factory and synthetic samples generated by CFD simulation. After on-site deployment, it is fine-tuned by a small number of manually annotated frames. The encoder of the improved U-Net contains a 5-level downsampling module (each level contains two 3×3 convolutional layers + batch normalization + ReLU activation, followed by 2×2 max pooling), and the decoder contains a 4-level upsampling module (bilinear upsampling + skip connection concatenation + convolutional thinning). The output layer uses the Sigmoid activation function to output two pixel-level semantic masks: a binary mask of the flame area. Binary mask for the black fire head region The classification thresholds are all 0.5.

[0027] Step 6: Morphological post-processing and contour refinement, for the mask. and Perform the following steps separately: Remove isolated noisy connected components with an area less than 0.5% of the total number of pixels; A morphological closing operation, first dilation and then erosion, is performed on a 5×5 rectangular structuring element to fill the holes inside the mask caused by local occlusion. The convex hull algorithm is applied to the largest connected region to fit the main contour, resulting in a refined set of flame main contour coordinates. Set of boundary coordinates of the Blackfire Head region .

[0028] Step 7: Quantitative calculation of the three core feature parameters.

[0029] Blackfire head length Calculation: Using the pixel position of the pulverized coal nozzle outlet in the image as the axial origin (determined through offline calibration), the mask of the black fire head area is measured along the kiln axis. Number of axially extended pixels covered Multiply by the pre-calibrated pixel scaling factor (Unit: mm / pixel, offline calibration by placing a calibration target of known size within the field of view of the fire-viewing hole): ; The unit is mm, which reflects the timeliness of pulverized coal ignition.

[0030] Flame stiffness index Calculation: Extracting the main contour of the flame The sequence of geometric center coordinates at each cross-section is fitted to a quadratic polynomial curve, and the average curvature of the fitted curve is used as the reference value. The reciprocal of (unit: 1 / pixel) multiplied by the normalization factor (Determined by the number of pixels in the kiln diameter): ; It is a dimensionless quantity; the larger the value, the straighter the flame axis and the higher the stiffness.

[0031] Flame spreads out at half an angle Calculation: The axial distance from the pulverized coal nozzle outlet (Measure the radial half-width of the flame profile at the cross-sectional position at the preset reference cross-sectional distance, typically 5 times the diameter of the pulverized coal nozzle.) (Unit: mm): ; The unit is degrees (°), which reflects the intensity of the internal wind vortex.

[0032] The three feature parameters are assembled into a feature vector. The data is then encapsulated into a structured data frame and uploaded to the PINN model inference service of the central control center via industrial Ethernet to complete one perception processing cycle. The data is then synchronized with the UTC timestamp of the IEEE1588 protocol (precision: microsecond level, synchronization error with the central control center server clock is less than 1μs).

[0033] Module 2: PINN-driven dynamic decoupling optimization control of internal / external swirl ratio In terms of the overall design of the PINN model, it uses observable flame morphology feature vectors. and current working condition vector As input, the internal air setting value and external wind setting value (Unit: m) 3 / h) is used as output to construct a direct quantitative physical mapping from observable flame state to fan control commands. The operating condition vector is defined as ,in The temperature at the kiln tail is... For pulverized coal flow rate, This represents the current total air volume. The current internal / external wind measurements are respectively given. The expanded network input vector has 8 dimensions. The training and inference process of the PINN model is as follows: Figure 4 As shown.

[0034] PINN backbone network adopts a fully connected deep residual network architecture: 8 nodes in the input layer; It contains 4 cascaded residual blocks, each consisting of 2 fully connected hidden layers (256 neurons per layer, Swish activation function) and 1 identity shortcut connection. When the input and output dimensions are different, the shortcut connection aligns the dimensions through a linear transformation. The output layer has 2 nodes, and the activation function is Sigmoid (with the range normalization constraint that the output value is in the range [0,1]). The total number of parameters in the entire network is about 270,000, and it can achieve millisecond-level forward inference on an edge GPU platform.

[0035] PINN's loss function consists of two parts: data fitting loss and physical constraint penalty loss. ; in These are the physical constraint weighting coefficients. The mean square error between the network output and the historical best internal / external wind setpoint labels. This is to account for the loss due to physical constraints.

[0036] Comprehensive physical constraint loss It consists of a weighted average of three sub-items: ; in =0.5、 =2.0 is the weight coefficient of each sub-constraint.

[0037] Mass conservation constraint loss The sum of the internal and external air volumes output by the network must be strictly equal to the total air volume setpoint to ensure mass conservation. ; in This represents the batch sample size. For the first The total air volume constraint value corresponding to each sample.

[0038] Swirl ratio-ignition distance relationship constraint loss Embedding prior knowledge of combustion physics into the network training process: ; in For the first The predicted swirl ratio for each sample The semi-empirical relationship between swirl ratio and ignition distance fitted from CFD data ( (The correlation coefficient of the working conditions is obtained by fitting the CFD simulation results). For the first The measured length of the black flame head of each sample.

[0039] Range constraint loss The network output exceeding the safe upper / lower limits of internal / external airflow is penalized using ReLU penalty to prevent the output from exceeding the engineering safety boundary. For each sample, the internal airflow is calculated to be below the lower limit. and exceeding the limit The penalty items, and the external air volume being lower than the lower limit. and exceeding the limit The penalty items are calculated using the average of the batches.

[0040] The PINN model employs a two-stage training strategy: The first stage is the offline pre-training stage, which extracts complete operation records from the factory's historical DCS database for nearly one year. Positive samples are selected based on the following rules: clinker free calcium oxide content is less than 1.5%, kiln tail temperature is within the range of 900℃ to 1100℃, and no abnormal alarms are issued by the operator. This forms the historical data positive sample set. Simultaneously, 78 CFD steady-state simulations were run under the Cartesian product combination of the swirl ratio value space (from 0.5 to 3.0, with a step size of 0.1, for a total of 26 levels) and the total air volume at 3 levels. The black flame head length, flame stiffness, and divergence angle features of each simulation output were extracted to form a CFD simulation sample set. The two datasets were mixed in a 4:1 ratio, linearly normalized according to the global maximum and minimum values ​​of their respective variables, and then divided into training and validation sets in an 80:20 ratio.

[0041] The training process employs a course-based learning strategy: physical constraint weighting coefficients. During training, the initial value is linearly increased from 0.1 to 1.0 (increasing by 0.1 every 100 training rounds). This allows the network to prioritize fitting the data in the early stages of training and gradually strengthen the physical constraints in the later stages, training a deep residual network with physical consistency. The Adam optimizer is used (initial learning rate of 10⁻³, decaying to 10⁻⁵ within 500 rounds using a cosine annealing strategy), with a batch size of 128. The early stopping condition is that the total loss of the validation set does not decrease for 20 consecutive rounds. After training is completed, the optimal model weights on the validation set are saved.

[0042] The second stage is the online few-sample transfer fine-tuning stage. After on-site deployment, a transfer fine-tuning is triggered every 50 operator-confirmed optimal adjustment records: all weight parameters of the first 3 residual blocks of the network are frozen (preserving the general feature representations learned in pre-training), and only the weight parameters of the 4th residual block and the output layer are unfrozen, using a smaller learning rate of 10. -4 and L2 regularization (coefficient 10) -4 By fine-tuning the data 20 times on a small sample, a precise adaptation to the specific operating characteristics of individual kilns can be achieved.

[0043] Module 3: AGPC Adaptive Generalized Predictive Compensation Control Regarding the control problem description and overall architecture, the adaptive generalized predictive compensation controller at the sampling time... Calculate the control increment (Corresponding to the correction amount for the frequency of the internal or external air inverter, unit: Hz). This command is transmitted via industrial Ethernet, and the actual one-way delay is... (Unit: ms, random time-varying) The adaptive generalized predictive compensation controller solves the random time delay problem through a two-end collaborative mechanism of dynamic step-size predictive compensation at the central control center and state observer correction at the field programmable logic controller. The complete algorithm flow is as follows: Figure 5 As shown.

[0044] Accurate real-time measurement of network round-trip latency: Adaptive generalized predictive compensation controller in the central control center for each control sampling cycle (Typical value: 500ms) Sends a data packet with a precise IEEE 1588 timestamp to the field programmable logic controller. Heartbeat detection package; Upon receiving the probe packet, the field-programmable logic controller (FPGA) immediately sends back an acknowledgment packet, along with a timestamp of the reception time. ; The control center records the time when it receives the confirmation packet. Calculate round-trip time delay: ; Take the estimated one-way delay value (Unit: ms, IEEE1588 protocol guarantees that the clock synchronization error between the two ends is less than 1μs, and the measurement accuracy meets the requirements), it is stored in a time delay sliding window buffer of length 50, and the current time delay sliding mean and variance are calculated for subsequent Markov chain state updates.

[0045] Online modeling and prediction of Markov chain delay states: The delay value space is divided into 5 state intervals to form a finite state space. The corresponding intervals are [0,100)ms, [100,200)ms, [200,300)ms, [300,400)ms, and [400,+∞)ms, respectively, with the central mean of each interval being... ( =1,...,5) takes values ​​of 50, 150, 250, 350, and 450ms respectively. A 5×5 state transition count matrix is ​​maintained (initialized as an all-1 matrix to avoid zero probability issues). Each control cycle increments the current state transition count by 1, and the normalized state transition probability matrix is ​​updated in real-time, where elements... Indicates from state Transition to state in the next moment The probability, and , This is the corresponding state transition count.

[0046] In the current state Given the initial state, predict the future using matrix exponentiation. The probability distribution of the delayed state of the step: ; in For the first One standard unit basis vector, Given the current state transition probability matrix, calculate the expected delay for each step: ; in For the future Step in state The probability, based on the base prediction step size =10 steps as the time window, calculate the expected peak delay within the window. .

[0047] The adaptive generalized predictive compensation controller dynamically adjusts the prediction step size and control step size based on the expected peak delay. The prediction step size of this control cycle is dynamically adjusted. and control step size : ; in =10 is the base step size. The equivalent delay step size (rounded down). =3 is the safety margin step size, and the control step size is determined by the following formula: ; in =5 controls the upper limit of the step size. To round up, the prediction step size is dynamically expanded so that the optimal prediction window of the adaptive generalized prediction compensation controller always fully covers the equivalent pure time delay introduced by network latency, thus fundamentally avoiding control mismatch.

[0048] Centrifugal fan ARX process model with forgetting factor RLS online identification: Independent ARX identification models are maintained for the internal and external air systems respectively. Taking internal air as an example, the ARX model is defined as follows: ; in The measured air volume (m³) at the current sampling time 3 / h), This is the historical output vector from the past two steps. The historical control input vectors from the past two steps (internal air inverter frequency command, unit: Hz). The parameter vector of the ARX model to be identified. It is residual white noise.

[0049] Centrifugal fan systems are low-order dynamic systems dominated by inertia. A second-order ARX model can effectively capture its dominant dynamic characteristics, and the appropriate order is beneficial to the online convergence of the recursive least squares method and the real-time computation within the control cycle.

[0050] Let the parameter vector to be identified Regression vector Execution with forgetting factor The RLS recursive update with a value of 0.98 is defined by the following update formulas: ; ; ; in The Kalman gain vector. The covariance matrix (initialized to) , (4th order identity matrix), forgetting factor =0.98<1 enables the identification algorithm to assign greater weight to recent data, effectively tracking the time-varying dynamic characteristics of centrifugal fans caused by changes in operating conditions.

[0051] Construction of the cost function and solution of the control increment sequence for the adaptive generalized predictive compensation controller: Calculation based on online identified ARX model parameters. Step-by-step response matrix and free response vector This indicates the future performance of the system when the current control quantity remains unchanged. (Based on the natural evolution trajectory of the steps), construct the cost function of the adaptive generalized prediction compensation controller: ; in In order to be in Given information at any given time, The predicted value output by the system at any given time; Internal airflow setpoint output by the PINN model Or external wind setting value The constructed reference trajectory remains at the target value within the prediction window; For the first Step prediction error weights, where =1.0、 =0.95, reflecting increased uncertainty in long-term forecasts; =0.1 is the first Step-by-step control of incremental weights prevents overly aggressive control actions from damaging the actuator; For the first Step control increment (variable to be optimized).

[0052] The cost function with respect to the control increment sequence Taking the partial derivative and setting it to zero, we obtain the analytical optimal solution: ; in , This is a weighted diagonal matrix. Using the reference trajectory vector, and following the rolling time-domain control principle, the first element of the control increment sequence is taken. This refers to the actual control commands issued during this cycle.

[0053] Control command frame encapsulation and timestamped delivery: This will increment the control volume for the current cycle. Expected delay value Expected reference state vector (Predicted by the internal state equations of the adaptive generalized predictive compensation controller, considering...) (The target state that the actuator should achieve) and the precise IEEE 1588 transmission timestamp. The control command is encapsulated into a four-element control command frame and sent to the field programmable logic controller via industrial Ethernet (TCP / IP protocol, using QoS priority queue to ensure priority transmission of control command frames).

[0054] Field-Programmable Logic Controller (FPGA) Delay Verification and Remote State Observer Compensation: After receiving a control command frame, the FPGA immediately records the reception time. (Accurately recorded by the IEEE1588 module on the programmable logic controller side), calculate the actual one-way arrival delay. With the predicted delay carried in the instruction frame The difference is used to obtain the time delay prediction error. .

[0055] when When the time exceeds 50ms, the remote status observer compensation mechanism is activated: the observer reads the current measured values ​​of the indoor / outdoor airflow. ,by As input, the Runge-Kutta fourth-order method is used to integrate the state equations of the process model forward. Milliseconds, yielding the state drift during the additional latency period. (That is, the deviation between the actual state and the controller's expected state), and based on this, calculate the compensation and then execute the control command: ; in The observer compensation gain matrix is ​​pre-designed using an offline pole placement method, and the poles are configured to ensure that the compensation response has no overshoot. When ≤50ms, directly use As an execution instruction, no observer compensation is required.

[0056] Actuator Drive and Result Feedback Reporting: The field-programmable logic controller (FPGA) converts the final execution command into inverter frequency adjustment commands (controlling total air volume), indoor butterfly valve opening adjustment commands, and outdoor butterfly valve opening adjustment commands, respectively, driving the corresponding actuators to complete the actions. After execution, the FPGA reports the measured indoor / outdoor air volume values. and the actual delay of this period Package and report to the central control center.

[0057] After receiving the data, the adaptive generalized predictive compensation controller at the central control center substitutes the measured air volume value into the RLS algorithm to update the model parameters. The measured time delay is used to update the Markov chain state transition counting matrix, refresh the time delay probability prediction model, and enter the calculation loop of the next control cycle, forming a continuously operating adaptive closed-loop control system.

Claims

1. A remote precision control system for centrifugal fan kiln head air, characterized in that, It includes a perception layer, an edge computing layer, a control and decision-making layer, and an execution layer; The sensing layer includes a multi-band industrial near-infrared camera installed at the fire observation hole at the kiln head of the rotary kiln. The multi-band industrial near-infrared camera is equipped with a near-infrared (NIR) band sensor and a short-wave infrared (SWIR) band sensor, and is externally fitted with a high-temperature resistant cooling water jacket. The edge computing layer includes edge computing nodes deployed in the kiln head field cabinet. The edge computing nodes receive NIR and SWIR images acquired by the multi-band industrial near-infrared camera, and sequentially perform adaptive spatiotemporal joint filtering, multi-band adaptive weight fusion, and multi-scale semantic segmentation to quantitatively extract three combustion state feature parameters: black flame length, flame stiffness index, and flame divergence half angle, forming a feature vector F, which is then uploaded to the control decision layer with a timestamp. The control decision layer includes a physical information neural network (PINN) decoupling optimization model running on the central control center server, a network delay state observer, and an adaptive generalized prediction compensation controller. The PINN decoupling optimization model takes the feature vector F and the operating condition vector W as inputs and outputs the internal wind setting value and the external wind setting value. The network delay state observer uses a finite-state Markov chain to model the network delay from the control center to the field and predict the future delay distribution. The adaptive generalized predictive compensation controller dynamically adjusts the prediction step size and control step size according to the predicted future time delay distribution, identifies the centrifugal fan process model online, solves the optimal control increment sequence, and sends the control instruction frame with timestamp to the execution layer via the network. The execution layer includes a field-programmable logic controller, a frequency converter, an internal butterfly valve, and an external butterfly valve; The field programmable logic controller receives the control command frame, verifies the timestamp and performs delay compensation, then drives the frequency converter to adjust the total air volume, drives the internal air butterfly valve to adjust the internal air volume, drives the external air butterfly valve to adjust the external air volume, and reports the execution result to the central control center to form a closed loop.

2. The centrifugal fan kiln head air remote precision control system according to claim 1, characterized in that, The NIR band sensor has a wavelength range of 0.9μm to 1.7μm, and the SWIR band sensor has a wavelength range of 1.7μm to 2.5μm. The two sensors support synchronous exposure acquisition with a time synchronization error of less than 1ms. The high-temperature resistant cooling water jacket uses circulating cooling water to maintain the operating temperature of the multi-band industrial near-infrared camera below 60°C. The edge computing node is equipped with an industrial-grade GPU with a computing power of no less than 8 TOPS; The multi-band industrial near-infrared camera transmits a video stream with a resolution of 1280×1024 pixels to the edge computing node at a frame rate of no less than 25fps via an industrial gigabit Ethernet fiber.

3. The centrifugal fan kiln head air remote precision control system according to claim 1, characterized in that, The adaptive spatiotemporal joint filtering includes two processing steps: temporal adaptive inter-frame difference filtering and spatial anisotropic diffusion filtering. The temporal adaptive inter-frame differential filtering estimates the dust motion intensity of the current frame using the global mean of the pixel-by-pixel absolute difference map of adjacent frames. It dynamically updates the motion dust detection threshold with a forgetting factor of 0.3 using an exponentially weighted moving average strategy. Pixels in the difference map that exceed the threshold are marked as motion dust occlusion areas. Pixels in the occlusion areas are replaced temporally using the mean of a buffer queue of 5 frames. Pixels in the non-occlusion areas retain their original values. The filtered image frame with motion dust noise eliminated is output. The spatial anisotropic diffusion filter applies Perona-Malik anisotropic diffusion iteration to the time-domain filtered image with a time step of 0.25 and a gradient sensitivity parameter of 30, for a total of 10 iterations. The diffusion coefficient is a negative exponential function of the gradient magnitude, which makes the diffusion coefficient at the flame boundary approach 0 to preserve the boundary gradient information, and makes the diffusion coefficient in the uniform dust region approach 1 for strong smoothing and noise reduction, outputting a smooth image frame with preserved boundaries.

4. The centrifugal fan kiln head air remote precision control system according to claim 1, characterized in that, The processing steps for the multi-band adaptive weight fusion are as follows: Calculate the local information entropy of the NIR and SWIR filtered images in 8×8 pixel blocks respectively; After normalizing the local information entropy of each band, the adaptive fusion weight of the two bands at each pixel position is obtained. The NIR image of the core area of ​​the high-temperature flame has a higher weight, while the SWIR image of the dust-rich area has a higher weight. The dual-band image is linearly weighted and fused according to the adaptive fusion weights to output an enhanced fused image.

5. The centrifugal fan kiln head air remote precision control system according to claim 1, characterized in that, The quantitative extraction steps for the three combustion state characteristic parameters are as follows: The enhanced fused image is segmented at the pixel level by improving the U-Net multi-scale semantic segmentation network, and the flame region mask and the black flame head region mask are output. Then, the image is further processed by removing isolated connected components with an area less than 0.5% of the total number of pixels, performing morphological closing operation of 5×5 structuring element, and refining the convex hull contour. The length of the black flame head is obtained by multiplying the number of pixels extending axially from the black flame head region mask along the kiln axis by an offline calibrated pixel scaling factor. The flame stiffness index is obtained by extracting the geometric center coordinate sequence of each section of the main contour of the flame, fitting a quadratic polynomial curve, and then multiplying the inverse of the average curvature by a normalized scaling factor calibrated offline by the number of pixels in the kiln inner diameter. The flame divergence half-angle is obtained by measuring the radial half-width of the flame profile at an offline calibration reference section located at an axial distance of 5 times the diameter of the pulverized coal nozzle outlet, and calculating the arctangent of the ratio of the radial half-width to the axial distance.

6. The centrifugal fan kiln head air remote precision control system according to claim 1, characterized in that, The loss function of the PINN decoupling optimization model is a weighted sum of data fitting loss and physical constraint penalty loss. The physical constraint penalty loss is the sum of mass conservation constraint loss, swirl ratio-ignition distance constraint loss with a weight coefficient of 0.5, and range constraint loss with a weight coefficient of 2.

0. The mass conservation constraint loss requires that the sum of the internal air volume and the external air volume output by the network equals the total air volume constraint value. The swirl ratio-ignition distance constraint loss embeds a semi-empirical relationship between the swirl ratio and the blackhead length fitted by CFD data into the network training process, wherein the swirl ratio is calculated by the quotient of the internal air volume and the external air volume. The range constraint loss is applied in the form of ReLU penalty to penalize the portion of the internal / external air volume that exceeds the safe upper and lower limits; The backbone network of the PINN decoupled optimization model is a fully connected deep residual network containing four cascaded residual blocks, each containing two hidden layers with 256 neurons each.

7. The centrifugal fan kiln head air remote precision control system according to claim 6, characterized in that, The PINN decoupling optimization model employs a two-stage training strategy: The first stage is the offline pre-training stage, in which the training set is constructed by mixing historical DCS operation data and CFD numerical simulation data in a 4:1 ratio. During the training process, the physical constraint weight coefficient is linearly increased from 0.1 to 1.0 in a step of 0.1 every 100 rounds, with the verification condition that the total set loss does not decrease for 20 consecutive rounds as an early stopping condition. The second stage is the online small sample transfer fine-tuning stage. Every 50 on-site optimal adjustment records are accumulated, a transfer fine-tuning is triggered. The weights of the first 3 residual blocks of the network are frozen, and only the weights of the 4th residual block and the output layer are unfrozen. Fine-tuning is performed on small sample data for 20 rounds to achieve adaptation to the specific kiln individual operating conditions.

8. The centrifugal fan kiln head air remote precision control system according to claim 1, characterized in that, The finite-state Markov chain divides the network latency into 5 state intervals, corresponding to [0,100)ms, [100,200)ms, [200,300)ms, [300,400)ms, and [400,+∞)ms respectively, with the mean values ​​of the center of each interval being 50, 150, 250, 350, and 450ms respectively; By maintaining a 5×5 state transition counting matrix and normalizing and updating the state transition probability matrix in each control cycle, the state probability distribution and expected delay of each step are predicted by matrix exponentiation. The adaptive generalized prediction compensation controller dynamically adjusts the current prediction step size based on the expected delay peak within a 10-step time window of the basic prediction step size, with a safety margin step size of 3 steps and a control step size upper limit of 5 steps, so that the prediction window always covers the equivalent pure delay introduced by network latency. The recursive least squares method with a forgetting factor of 0.98 was used to identify the ARX process model of the centrifugal fan online and track the time-varying dynamic characteristics of the centrifugal fan.

9. The centrifugal fan kiln head air remote precision control system according to claim 1, characterized in that, After receiving the control command frame, the field-programmable logic controller calculates the difference between the actual arrival delay and the predicted delay as the delay prediction error. When the absolute value of the delay prediction error exceeds 50ms, the remote state observer compensation mechanism is activated. Read the current measured value of the internal / external air volume on site, take the control increment as input, use the Runge-Kutta fourth-order method to integrate the state equation of the process model forward to predict the time corresponding to the error, and obtain the state drift. Then, after correcting the control increment with the compensation gain matrix designed by offline pole configuration, execute the program. Configure the poles to make the compensation response free of overshoot. When the absolute value of the time delay prediction error does not exceed 50ms, the original control increment is executed directly.

10. The centrifugal fan kiln head air remote precision control system according to claim 1, characterized in that, The control command frame contains four elements: The current optimal control increment, the expected value of the predicted delay, the expected reference state vector, and the precise IEEE 1588 transmission timestamp; The control command frame is transmitted to the field programmable logic controller via an industrial Ethernet network employing QoS priority queuing. After the field programmable logic controller completes its execution, it reports the measured values ​​of internal / external air volume and the actual time delay of the current cycle to the central control center. The central control center substitutes the measured values ​​of internal / external air volume into the recursive least squares method to update the parameters of the centrifugal fan process model, and uses the actual time delay for online refreshing of the Markov chain state transition counting matrix.

Images