Airflow spray dryer energy efficiency energy-saving and efficiency-increasing method and system based on big data

Abstract

The invention relates to the technical field of crossing of mechanical engineering and process control, discloses an airflow spray dryer energy efficiency energy-saving and efficiency-improving method and system based on big data, and aims to solve the problems that atomization parameters are rigid, a nozzle is easy to block, control lags and energy efficiency and quality are difficult to consider in the drying process of high-viscosity slurry. The method comprises the following steps: acquiring slurry viscosity and droplet particle size distribution through high-frequency online rheological sensing and high-speed imaging; multi-source time sequence data are fused and input into a lightweight spatial-temporal feature extraction network, and optimization instructions of the fan rotating speed, the nozzle aperture and the hot air temperature are output in a millisecond level; and the piezoelectric fine adjustment mechanism and the frequency conversion fan are driven to execute self-adaptive regulation and control, and preventive maintenance is achieved in combination with micro-vibration anti-blocking and blocking risk prediction. The system integrates a sensing module, a modeling module, an execution module and an optimization module, the heat energy utilization rate is remarkably increased, and the diameter deviation of liquid drops is controlled within a small range.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of mechanical engineering and process control, and specifically relates to a method and system for improving the energy efficiency of an airflow spray dryer based on big data. Background Technology

[0002] Airflow spray dryers are key equipment used in chemical, pharmaceutical, and food industries to rapidly dry liquid slurries into powders. Their basic principle is to disperse the slurry into fine droplets through an atomizer, which then come into instantaneous contact with high-temperature hot air, thereby achieving rapid evaporation of moisture.

[0003] However, traditional drying systems generally suffer from the following technical bottlenecks when processing high-viscosity, rheologically complex slurries: First, these systems often use fixed parameters such as fan speed and nozzle orifice diameter, failing to dynamically adjust according to real-time changes in material viscosity. This results in uneven atomized droplet size, reduced drying efficiency, and impacts the quality stability of the final product. Second, high-viscosity materials easily adhere to and deposit inside and on the nozzle walls, leading to blockages, production interruptions, frequent equipment maintenance, and poor continuous operation. Furthermore, existing improvement schemes that incorporate monitoring and control often suffer from limitations such as single-dimensional sensing data, lack of real-time online measurement of key rheological parameters, high analysis and decision-making delays, and insufficient linkage between control commands and actuators. These limitations hinder precise responses to the millisecond-level dynamic characteristics of the drying process, making it impossible to achieve optimal energy consumption control while ensuring product quality.

[0004] Therefore, the industry urgently needs a drying control method and system that can sense material characteristics in real time, make rapid and intelligent decisions, and execute precise regulation to solve long-standing problems in the drying process of high-viscosity slurries, such as poor adaptability, easy clogging, and difficulty in synergistically optimizing energy efficiency and quality. Summary of the Invention

[0005] This invention provides a method and system for improving energy efficiency and reducing energy consumption in airflow spray dryers based on big data. It aims to solve industry-wide technical bottlenecks in the drying of high-viscosity slurries caused by rigid atomization parameters, easy nozzle clogging, lag in control response, and the trade-off between energy efficiency and product quality. This invention constructs an intelligent drying system integrating real-time rheological sensing, millisecond-level dynamic modeling, adaptive actuators, and closed-loop feedback control. This achieves coordinated and precise control of droplet diameter, hot air parameters, and nozzle status, thereby significantly improving energy utilization efficiency while ensuring product quality.

[0006] According to one aspect of the present invention, a method for improving energy efficiency and saving power in an airflow spray dryer based on big data is provided, comprising: Real-time viscosity data of the slurry before it enters the atomizer is obtained using an online rheology sensor. The high-speed imaging module continuously acquires images of the droplet swarm at the outlet of the atomization chamber and calculates the standard deviation and median particle size of the current droplet diameter distribution. The viscosity data, the calculated droplet diameter distribution data, the hot air temperature, the hot air speed, the nozzle orifice status, and the ambient humidity data are fused into a multi-source heterogeneous time-series data stream and input into a pre-trained lightweight spatiotemporal feature extraction network. The lightweight spatiotemporal feature extraction network outputs the optimal fan speed command, the optimal nozzle orifice adjustment amount, and the hot air temperature correction value under the current operating conditions. The variable frequency fan is controlled according to the optimal fan speed command, the piezoelectric ceramic fine-tuning mechanism is driven to change the nozzle outlet geometry according to the optimal nozzle orifice diameter adjustment amount, and the power output of the electric heater is adjusted according to the hot air temperature correction value. A micro-vibration excitation unit is integrated inside the nozzle. Its vibration frequency and amplitude are dynamically set according to the current viscosity and Reynolds number to suppress material deposition. A nozzle clogging risk prediction model is established. The model takes viscosity, runtime, historical deposition rate and current vibration status as inputs. When the predicted clogging probability exceeds a preset threshold, preventive intervention is triggered.

[0007] As one embodiment of the present invention, the online rheological sensor acquires slurry viscosity data in real time, including: an online rheological sensor with a concentric cylindrical structure, the inner cylinder being driven to rotate by a servo motor, measuring the inner cylinder torque to infer the slurry shear stress, and calculating the dynamic viscosity in combination with the known shear rate, wherein the sampling frequency of the viscosity data is not less than 1 kilohertz.

[0008] As one embodiment of the present invention, the high-speed imaging module is used to acquire images and calculate the droplet diameter distribution, including: using an image sensor with a frame rate of 5,000 frames per second, in conjunction with a ring light source, and performing edge detection and morphological analysis through an image processing unit deployed on a field-programmable gate array chip to calculate the standard deviation and median particle size of the droplet diameter distribution.

[0009] As one embodiment of the present invention, the data are fused into a multi-source heterogeneous time-series data stream, including: resampling data with different sampling rates to a reference frequency of 1 kilohertz through a timestamp alignment mechanism, interpolating missing data points to complete them, and removing and replacing outliers, ultimately forming a multi-dimensional time-series vector containing viscosity, droplet diameter standard deviation, median particle size, hot air temperature, hot air speed, ambient humidity, and cumulative nozzle running time.

[0010] As one embodiment of the present invention, the lightweight spatiotemporal feature extraction network is constructed based on a one-dimensional convolution and attention mechanism. Its input is the multi-source heterogeneous temporal data stream. The network architecture includes multiple one-dimensional convolutional blocks and channel attention modules. The output layer is used to directly map and output the fan speed command, nozzle orifice adjustment amount and hot air temperature correction value.

[0011] As one embodiment of the present invention, the driving piezoelectric ceramic fine-tuning mechanism changes the nozzle outlet geometry, comprising: the piezoelectric ceramic fine-tuning mechanism is composed of an annular piezoelectric stacked driver, a flexible hinge amplification mechanism and a conical nozzle mandrel, and the conical nozzle mandrel is driven to move by changing the driving voltage to continuously adjust the nozzle outlet orifice diameter.

[0012] In one embodiment of the present invention, the micro-vibration excitation unit consists of a piezoelectric ceramic sheet embedded inside the nozzle wall and a mass block forming a resonant structure, and its vibration control module dynamically selects the excitation frequency and amplitude according to the currently calculated Reynolds number.

[0013] As one embodiment of the present invention, the establishment of the nozzle clogging risk prediction model includes: constructing a model using a gradient boosting decision tree algorithm to output a probability value of clogging occurring within a set future time; when the probability value exceeds a preset threshold, the triggering of preventive intervention includes increasing the amplitude of the micro-vibration excitation unit and instructing the piezoelectric ceramic fine-tuning mechanism to expand the nozzle orifice diameter.

[0014] As one embodiment of the present invention, it further includes: performing energy efficiency-quality co-optimization, constraining optimization with thermal energy utilization rate and droplet diameter deviation as dual objective functions, and using the optimization results to calibrate the output of the lightweight spatiotemporal feature extraction network.

[0015] According to another aspect of the present invention, an energy-saving and efficiency-enhancing system for airflow spray dryers based on big data is provided, comprising: An online rheology sensing unit is used to acquire real-time viscosity data of the slurry before it enters the atomizer via an online rheology sensor. The high-speed droplet imaging unit is used to continuously acquire images of the droplet group at the outlet of the atomization chamber through the high-speed imaging module, and to calculate the standard deviation and median particle size of the current droplet diameter distribution. The multi-source data fusion unit is used to fuse the viscosity data, the calculated droplet diameter distribution data, the hot air temperature, the hot air speed, the nozzle orifice status, and the ambient humidity data into a multi-source heterogeneous time-series data stream. The lightweight spatiotemporal feature extraction unit is used to input the time-series data stream into the pre-trained lightweight spatiotemporal feature extraction network and output the optimal fan speed command, the optimal nozzle orifice adjustment amount, and the hot air temperature correction value. The adaptive execution unit includes a variable frequency fan that responds to the optimal fan speed command, a piezoelectric ceramic fine-tuning mechanism that responds to the optimal nozzle orifice diameter adjustment amount, and an electric heater that responds to the hot air temperature correction value. The micro-vibration anti-clogging unit, integrated inside the nozzle, is used to dynamically apply preventative vibrations based on the current viscosity and Reynolds number; The clogging risk prediction unit is used to assess the probability of nozzle clogging by taking viscosity, runtime, historical deposition rate and current vibration status as inputs, and to trigger preventive measures when the clogging exceeds a preset threshold.

[0016] In summary, this application includes at least one of the following beneficial technical effects: (1) By combining online rheological sensing and high-speed imaging, the instantaneous changes in slurry viscosity and atomization state are captured in real time; and based on a lightweight spatiotemporal feature extraction network, millisecond-level analysis and decision-making are performed on multi-source heterogeneous data to synchronously and adaptively adjust the fan speed, nozzle orifice diameter, and hot air temperature. This overcomes the problems of rigid parameters and lag in traditional systems, ensuring highly uniform atomized droplet size and efficient transfer of hot air energy.

[0017] (2) Unlike the traditional passive cleaning mode, the present invention integrates a micro-vibration anti-clogging unit and a clogging risk prediction model, which can dynamically apply preventive vibration to suppress deposition according to the current working conditions; and when the clogging risk is predicted, the vibration parameters and nozzle orifice diameter are adjusted in advance to eliminate the clogging risk in the bud.

[0018] (3) This invention uses droplet size control and thermal energy utilization as optimization objectives, and performs closed-loop calibration with the help of an energy efficiency-quality synergistic optimization module. This allows the system to significantly improve thermal energy utilization while keeping the droplet diameter deviation within a very small range. This solves the previous technical problem of "at the expense of one thing and the other" and achieves a balance between energy saving and efficiency improvement. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall technical solution architecture of the energy-saving and efficiency-enhancing method and system for airflow spray dryers based on big data proposed in this invention. Figure 2 This is a schematic diagram of the core principle framework of the lightweight spatiotemporal feature extraction network and millisecond-level closed-loop collaborative control in this invention; Figure 3 This is a logical flowchart of the multi-source heterogeneous sensing data fusion and real-time working condition modeling in this invention. Detailed Implementation

[0020] This invention provides a method and system for improving energy efficiency and reducing energy consumption in airflow spray dryers based on big data. It aims to solve industry-wide technical bottlenecks in the drying of high-viscosity slurries caused by rigid atomization parameters, easy nozzle clogging, lag in control response, and the trade-off between energy efficiency and product quality. This method constructs an intelligent drying process integrating real-time rheological sensing, millisecond-level dynamic modeling, adaptive actuators, and closed-loop feedback control, achieving coordinated and precise control of droplet diameter, hot air parameters, and nozzle status.

[0021] See attached document Figure 1-3 As shown, the energy-saving and efficiency-enhancing method for airflow spray dryers based on big data in this application includes the following steps: S1, the viscosity data of the slurry before entering the atomizer is obtained in real time by an online rheological sensor, and the sampling frequency of the viscosity data is not less than one kilohertz; S2 continuously acquires images of the droplet group at the outlet of the atomization chamber through a high-speed imaging module, and calculates the standard deviation and median particle size of the current droplet diameter distribution using edge detection and morphological analysis algorithms. S3, the viscosity data, droplet diameter distribution data, hot air temperature, hot air speed, nozzle orifice status and ambient humidity data are fused into a multi-source heterogeneous time-series data stream and input into a pre-trained lightweight spatiotemporal feature extraction network. S4, the lightweight spatiotemporal feature extraction network is constructed based on a one-dimensional convolution and attention mechanism, and its inference delay does not exceed ten milliseconds. It is used to output the optimal fan speed command, the optimal nozzle orifice adjustment amount and the hot air temperature correction value under the current working condition. S5, control the variable frequency fan drive motor according to the optimal fan speed command, drive the piezoelectric ceramic fine-tuning mechanism to change the nozzle outlet geometry according to the optimal nozzle orifice diameter adjustment amount, and adjust the electric heater power output according to the hot air temperature correction value. S6, a micro-vibration excitation unit is integrated inside the nozzle. Its vibration frequency and amplitude are dynamically set by the lightweight spatiotemporal feature extraction network according to the current viscosity and Reynolds number, which is used to suppress the formation of laminar boundary layer and prevent material deposition. S7. Establish a nozzle clogging risk prediction model. The model takes viscosity, running time, historical deposition rate and current vibration status as inputs. When the predicted clogging probability exceeds a preset threshold of 5%, the micro-vibration excitation unit is activated in advance and the nozzle orifice diameter is adjusted to achieve preventive intervention.

[0022] Furthermore, the technical solutions and steps described above are explained in detail and publicly as follows.

[0023] The first step is S1: The viscosity data of the slurry before entering the atomizer is obtained in real time through an online rheological sensor.

[0024] The S101 online rheometer employs a concentric cylindrical structure, with the inner cylinder driven to rotate by a high-precision servo motor, while the outer cylinder remains stationary. At the start of the measurement, the servo motor drives the inner cylinder to rotate at a constant angular velocity, with an angular velocity control accuracy better than 0.01 radians / second, to ensure stable shear rate.

[0025] S102, Torque Measurement and Signal Acquisition. When the inner cylinder rotates, the slurry flowing through the gap between the inner and outer cylinders experiences shearing action. The shear stress exerted on the surface of the inner cylinder is measured by a high-resolution torque sensor. This torque sensor has a resolution of at least 0.001 N·m, enabling it to accurately capture minute changes in the slurry flow resistance.

[0026] S103, Shear Rate Calculation and Viscosity Inversion. Given the inner cylinder radius, outer cylinder radius, and the real-time rotational angular velocity of the inner cylinder, the shear rate of the slurry in the gap can be accurately calculated using well-known formulas in concentric cylinder rheology. Combining the shear stress measured in step S102, calculations are performed based on the selected fluid model: for Newtonian fluids, the (dynamic) viscosity is directly calculated based on their constitutive relation (shear stress is proportional to shear rate); for non-Newtonian fluids conforming to power-law characteristics, the viscosity is calculated using a power-law model (…). The apparent viscosity or rheological parameters (consistency coefficient K and flow index n) of the slurry are calculated by fitting. This enables real-time quantitative characterization of the rheological state of the slurry.

[0027] The steps S101 to S103 described above collectively complete the core process from mechanical motion to viscosity parameter calculation. A known shear rate is generated through precisely controlled rotation, and the corresponding shear stress is deduced through high-sensitivity torque measurement. Finally, the real-time viscosity is obtained using a rheological model. For example, for a Newtonian fluid, its viscosity is the ratio of shear stress to shear rate.

[0028] S104, Temperature Control and Data Output. To ensure that viscosity measurement is unaffected by ambient temperature fluctuations, a temperature control jacket integrated within the sensor (e.g., through a closed-loop control system formed by circulating a constant-temperature fluid or a built-in heating / cooling element and temperature sensor) maintains the temperature of the measurement area containing the slurry within a constant range of 25℃ ± 0.5℃, thereby eliminating the interference of temperature drift on viscosity calculation. The dynamic viscosity data calculated through the processes S101 to S103 above is continuously and in real-time (e.g., through a digital communication interface) output at a frequency of not less than 1000 Hz via the sensor's built-in or external signal processing module, providing the system with high-fidelity and timely rheological state information.

[0029] In summary, step S1 utilizes a well-structured and precisely controlled online rheological sensing system to achieve high-speed, high-precision, and interference-resistant real-time measurement of slurry viscosity. The obtained millisecond-level viscosity time-series data is one of the most crucial inputs for subsequent multi-source data fusion and intelligent decision-making, laying an accurate perceptual foundation for the adaptive control of the entire system.

[0030] Step S2: Continuously acquire and process images of the droplet swarm at the outlet of the atomization chamber using a high-speed imaging module, and calculate the standard deviation and median droplet diameter of the current droplet diameter distribution. This includes the following steps.

[0031] The S201 high-speed imaging module is fixedly mounted 50 mm directly below the atomization chamber outlet via a bracket, with its optical axis perpendicular to the main direction of droplet descent. The core imaging component of this module is a complementary metal-oxide-semiconductor (CMOS) image sensor, with a frame rate set to 5000 frames per second and an exposure time fixed at 20 microseconds. Simultaneously, a ring-shaped high-brightness LED light source is configured on the opposite side of the imaging area, opposite the image sensor, to provide uniform backlight illumination. To effectively "freeze" the high-speed moving droplets, the light source's emission pulses are triggered by an exposure synchronization signal generated by the image processing unit or the sensor itself, ensuring that the light source only illuminates at a high brightness during the sensor's exposure period. Thus, the system can continuously acquire high-contrast, motion-blur-free instantaneous images of the droplet swarm.

[0032] Step S201 clarifies the hardware setup and key parameters for image acquisition. High frame rate, short exposure time, and synchronized backlighting ensure that the trajectory of the high-speed droplet is clearly "frozen," providing high-quality raw images for subsequent analysis.

[0033] S202: The acquired raw grayscale image is transmitted in real time via a data interface to a dedicated image processing unit integrated on a field-programmable gate array (FPGA) chip for processing. The processing flow is as follows: Noise reduction filtering: First, the image is subjected to Gaussian filtering to suppress noise, and the standard deviation of the Gaussian kernel is set to 1.5.

[0034] Edge strength and direction calculation: For the filtered image, operators such as Sobel or Prewitt are used to calculate the gradient of each pixel in the x and y directions, thereby obtaining the gradient magnitude and direction.

[0035] Non-maximum suppression: Iterate through the gradient magnitude map, and for each pixel, retain only those pixels whose gradient magnitude is a local maximum along their gradient direction; otherwise, set their magnitude to zero. This ensures accurate edge localization, resulting in an edge width of approximately one pixel.

[0036] Dual-threshold detection and hysteresis connection: Two thresholds are set—a high threshold Thigh and a low threshold Tlow. Specifically, Thigh can be set to the top 5% quantile value in the global gradient magnitude statistics after non-maximum suppression, and Tlow is set to 40% to 50% of Thigh (e.g., 0.5 times). Pixels with gradient magnitudes higher than Thigh are marked as strong edge points, those lower than Tlow are set as background, and those in between are marked as weak edge points. Finally, hysteresis connection is performed: only weak edge points connected to strong edge points are ultimately retained as valid edges. After this step, the output is a coherent, closed single-pixel wide edge contour map, i.e., the preliminary droplet contour.

[0037] Step S202 above covers the complete front-end image processing workflow from noise reduction to contour extraction. By selecting specific algorithms and parameters (such as the standard deviation of Gaussian filtering), it is revealed how to transform the raw image into clear droplet contour information that can be used for measurement.

[0038] S203, Droplet Region Segmentation (Mask Generation): Based on all closed contours extracted in S202, each contour is processed as follows: Using a known point inside the contour (e.g., the center point of its minimum bounding rectangle) as a seed point, a region growing algorithm is employed. The growing criterion is unvisited pixels within a 4-connected or 8-connected neighborhood, with the aim of filling the entire interior region of the contour until the boundary. After filling, a corresponding binary mask image is generated, where the pixel value of the droplet region is 1 (white) and the background is 0 (black). Each such mask represents an independent droplet object segmented from the image.

[0039] Individual droplet size calculation: For each droplet binary mask generated in step 1, count the total number of pixels with a value of 1, which gives the pixel area of ​​the droplet in the image. The equivalent circular diameter D of each droplet is calculated using the following formula:

[0040] Where S is the spatial calibration coefficient, and the unit is micrometer / pixel.

[0041] System calibration: The spatial calibration coefficient S is determined through calibration during system initialization. The calibration method is as follows: several dimensions are known (e.g., nominal diameter is...). Standard microspheres (such as polystyrene microspheres) with a diameter of 1 micrometer (µm) are placed at the imaging plane of the atomization chamber outlet, and images are acquired under imaging conditions identical to those in actual production (including an installation distance of 50mm, lens focal length, aperture, and light source settings). The calibration image is then processed using methods 1 and 2 described above to calculate the pixel diameter of a single standard microsphere image. (or pixel area), then the spatial calibration coefficient By taking the average of multiple measurements, the calibration accuracy can be ensured, so that the measurement accuracy of the whole system for the droplet diameter reaches 0.1 micrometer.

[0042] S204. The system performs real-time statistical analysis on the droplet diameter data set obtained from the image processing of each frame, and calculates two key statistics, namely the standard deviation of its distribution and the median particle size. To balance the calculation load and data real-time performance, the system takes 200 milliseconds as an analysis cycle, aggregates all the droplet diameter data processed in the past 200 milliseconds (for example, corresponding to multiple frames of images) into a sample set, calculates the standard deviation and median particle size of this sample set, and updates and outputs this statistical result.

[0043] To synchronize with other data streams running at 1000 Hz in the system (such as viscosity data), the imaging analysis results need to be frequency-matched before output. The specific method is: between every two adjacent actual statistical values output in a 200-millisecond cycle, a linear interpolation algorithm is used to generate and output an interpolation value at a frequency of 1000 Hz (that is, every 1 millisecond). For example, if the output value at time t is , and the output value at time t + 200 ms is , then for the intermediate time t + n ms (0 < n < 200), its output value is . In this way, the statistics originally updated at 5 Hz (cycle 200 ms) are converted into a continuous data stream at 1000 Hz, so that it can be seamlessly input into the subsequent multi-source data fusion unit for millisecond-level fusion.

[0044] The entire image processing pipeline (from S202 to the statistical calculation of this step) is optimized on the FPGA to ensure that the total time from receiving a frame of raw image to outputting the updated statistical result contributed by this frame is controlled within 8 milliseconds, meeting the real-time requirements of the system.

[0045] This step S2 constructs a complete high-speed vision measurement scheme, progressing step by step from hardware layout, image acquisition, to software processing and data analysis, and finally realizing real-time, quantitative, and online monitoring of the particle size distribution and its uniformity of the atomized droplet group. The output high-frequency particle size statistical data is as crucial as the viscosity data obtained in step S1, and together they form the "two eyes" for the subsequent intelligent decision-making system to perceive the working conditions, providing a direct basis for evaluating the atomization quality and implementing closed-loop control.

[0046] After completing the high-speed imaging and droplet size analysis described in step S2, the system obtains the real-time core parameters characterizing atomization quality—the standard deviation of droplet diameter and the median droplet size. Simultaneously, the system continuously collects multi-dimensional sensor signals, including slurry viscosity, hot air parameters, and environmental conditions. These data originate from sensors at different physical locations and have different sampling frequencies and temporal characteristics. To integrate these multi-source, heterogeneous time-series data into a high-quality, synchronized dataset suitable for subsequent unified analysis by the intelligent network, specialized data fusion preprocessing is required, as described in step S3.

[0047] Step S3: Construct a multi-source heterogeneous time-series data stream.

[0048] In S301, the raw data generated by each sensor within the system (such as the online rheology sensor, hot air temperature sensor, high-speed imaging processing unit, etc.) carries a hardware-level timestamp, which records the actual acquisition time of each data point with nanosecond-level precision. The multi-source data fusion unit internally maintains a high-precision system absolute clock synchronized with each sensor as its time reference. Upon receiving data, the fusion unit parses the timestamp carried in each data packet and, based on this timestamp, "places" the data point to the corresponding time position relative to the system absolute clock. For data from different sensors at the same time (within an allowable small error range), they are aggregated into multi-dimensional data at the same time point; for data from non-integer sampling times, they are associated with the nearest target resampling time point (see S302). Through this process, the raw data from all sources are uniformly aligned to a common absolute time axis, laying the foundation for subsequent synchronization processing.

[0049] S302, after completing timestamp alignment (S301), the multi-source data fusion unit performs unified resampling on all data streams to ensure that their output frequency is 1000 Hz. The specific implementation is as follows: Determine the target time point: Based on the system's absolute clock, generate a sequence of equally spaced time points starting from the initial time with intervals of 1 millisecond (corresponding to 1000 Hz): { },in: .

[0050] Calculate resampled values ​​for each data channel: For each data channel corresponding to each sensor, for each target time point... Based on the original time series of this channel (already aligned), calculate its... The numerical estimate at time 1 is used as the output value after resampling.

[0051] Calculation method: If the original data sampling rate is higher than 1000 Hz, then The resampled value at time step can be selected The resampled values ​​are obtained by using the nearest original sampled point values ​​or by linear interpolation of multiple nearby original points. If the original data sampling rate is below 1000 Hz (e.g., 200 Hz for hot air temperature), the calculation of its resampled values ​​is already included in the upsampling process specifically described in S303.

[0052] For channels with an original sampling rate significantly higher than 1000 Hz, a low-pass digital filter (e.g., a finite-length unit impulse response filter with a cutoff frequency slightly lower than 500 Hz) can be applied to the original data before performing the aforementioned calculations to eliminate spectral aliasing that may be caused by resampling.

[0053] Through the above steps, the system generates a time point for each data source: { The data is defined on the} with equal intervals of 1000 points per second, and the data of all channels achieves strict point-to-point synchronization in time.

[0054] For signals with a sampling rate lower than 1000 Hz, such as hot air temperature (sampling rate 200 Hz) and ambient humidity (sampling rate 100 Hz), the S303 system uses a cubic spline interpolation algorithm to fit between its original data points and generate new data points, thereby upsampling its sequence to 1000 Hz.

[0055] S304. For missing data points caused by momentary communication interruption or brief sensor malfunction, the system handles them according to the following rules: If a data channel is found to have consecutive missing data points, and the number of missing sampling points does not exceed 5, the system will initiate automatic data completion. The specific method is as follows: using the two nearest valid data points before and after the missing segment as references (if one end is insufficient, more points are taken from the other end), a cubic spline interpolation algorithm is used to fit a smooth curve passing through these reference points, and the interpolation value corresponding to the missing time point is calculated to complete the data segment.

[0056] If the number of consecutive missing sampling points exceeds 5, the system determines it as a serious data anomaly. First, all data from that channel within that time period is marked as an "anomaly segment," and its start and end times are recorded. Then, a defined data backtracking and replacement process is triggered: The system immediately retrieves historical data sequences from the historical database for the sensor during the most recent period of stable operation with similar operating conditions (such as similar slurry viscosity and production rate), corresponding to the same system operating time offset.

[0057] Use the retrieved historical data to directly replace the currently marked "abnormal segment" data.

[0058] Simultaneously, the system sends an alarm message to the monitoring center stating "Data abnormal, historical backup enabled" for that channel. If no suitable historical data is available, the system temporarily switches to a degraded operation mode that relies on data from other related sensors for calculation, or marks the data for that dimension as invalid for subsequent model reference.

[0059] S305. To suppress the influence of outliers such as impulse noise, the system independently performs outlier processing for each 1000 Hz data channel. Specifically, it adopts the three-standard-deviation criterion of a sliding window: the mean and standard deviation of the sequence are calculated by sliding a window of length 1000 sampling points (i.e., 1 second of data). Data points whose values ​​exceed the range of "mean ± 3 standard deviation" are identified as outliers and replaced with the median of all data in the current sliding window.

[0060] In summary, after processing steps S301 to S305, the raw data from different sensors are regularized, repaired, and integrated. Finally, the system generates a unified seven-dimensional time-series vector output. This vector contains data in the following seven dimensions at each moment (every 1 millisecond), and each dimension is a continuous, non-abrupt, and non-missing 1000 Hz sequence: 1: Slurry viscosity; 2: Standard deviation of droplet diameter distribution; 3: Median droplet size; 4: Hot air temperature; 5: Hot air velocity; 6: Ambient humidity; 7: Cumulative nozzle runtime.

[0061] After completing the data fusion and preprocessing described in step S3, the system obtains a high-quality, highly consistent, and strictly synchronized seven-dimensional time-series data stream. This data stream accurately depicts comprehensive real-time operating conditions, including slurry rheology, atomization quality, hot air status, and cumulative equipment operating status. However, to instantly deduce the optimal control commands for the blower, nozzle, and heater from this massive amount of multidimensional data with complex spatiotemporal correlations, an intelligent core capable of understanding historical sequences, capturing dynamic features, and making accurate mappings is required. To this end, as described in step S4, the system deploys a lightweight spatiotemporal feature extraction network. This network takes the aforementioned fused data stream as direct input and performs the crucial leap from "perception" to "decision-making."

[0062] S401, a lightweight spatiotemporal feature extraction network, is deployed on the system's embedded AI acceleration module. Its architecture is optimized for millisecond-level low-latency inference. The network input layer receives a seven-dimensional time-series vector with a frequency of 1000 Hz, generated in step S3. The input sequence has a fixed length of 500 time steps, covering the most recent 500 milliseconds of historical operating data, forming an input tensor with dimensions [500, 7].

[0063] S402, the hidden layer of the network contains three sequentially connected one-dimensional convolutional blocks, used to extract local and deep patterns from the input temporal sequence layer by layer. Each convolutional block has the same structure: 1) a one-dimensional convolutional layer with a kernel size of 5; 2) a batch normalization layer; 3) a modified linear unit activation function. The number of output channels (i.e., the number of convolutional kernels) of the three convolutional blocks are set to 64, 128, and 256 respectively. This means that as the network deepens, the level of abstraction and the number of features gradually increase.

[0064] S403, after the feature map output by the third convolutional block (let its dimensions be [B, C, L], where C = 256 is the number of channels and L is the temporal length), the network introduces a channel attention module. This module performs the following operations in sequence: Channel description generation (compression): Global average pooling is performed on the input feature map along the temporal dimension L, compressing the two-dimensional features of each channel (L time steps) into a scalar. This produces a channel description vector of dimension [B, C, 1], where each element represents the average activation intensity of the corresponding feature channel over the entire time window, i.e., its global importance.

[0065] Channel weight learning (incentive): Input the channel description vectors obtained in the previous step into a small two-layer fully connected feedforward neural network to learn the nonlinear relationship between channels and generate weight coefficients.

[0066] The first layer (dimensionality reduction layer): a fully connected layer that maps C input neurons to C / r neurons, where the compression ratio r is set to 16. This layer is followed by a rectified linear unit (ReLU) activation function.

[0067] The second layer (recovery layer): a fully connected layer that maps C / r neurons back to C neurons (i.e., the same number of input channels).

[0068] Weight normalization: The output of the second layer is activated by a sigmoid function, which constrains the weight coefficient of each channel to between 0 and 1. The final output is a channel weight vector with dimensions [B, C, 1].

[0069] Feature recalibration (scaling): The learned channel weight vector is multiplied channel-by-channel with the feature map initially input to this module (i.e., the feature map before global average pooling). Based on this, channel features with weights close to 1 are enhanced, while channel features with weights close to 0 are suppressed, thereby enabling the network to dynamically focus on the key features required for the current control decision (such as responding to viscosity changes or adjusting particle size distribution).

[0070] In S404, the feature map output by the attention module is first flattened, transforming it into a one-dimensional feature vector. This vector is then fed into the network's output layer. The output layer consists of a single fully connected (DenseLayer). This layer has three neurons, each corresponding to one of the three control commands to be output. This layer uses a linear activation function (i.e., no non-linear activation), and its output values ​​are: fan speed increment (revolutions per minute); nozzle orifice adjustment steps (steps, where each step corresponds to an actual change of 0.5 micrometers in nozzle orifice diameter); and hot air temperature offset (degrees Celsius).

[0071] Using the above method, the learned high-dimensional spatiotemporal features are directly and linearly mapped into real-time control commands with clear physical meaning.

[0072] The S405 network design has undergone lightweight optimization, with the total number of trainable parameters strictly controlled to within 150,000. The trained network model was deployed to an embedded AI acceleration module for testing, and the average latency for completing a single forward inference process from input to output was 8.3 milliseconds, meeting the system's set real-time response limit of 10 milliseconds.

[0073] The training of the network was completed offline. The training dataset was constructed as follows: a large amount of time-series data over a period of time was extracted from the dryer's historical operation database, including the seven-dimensional fused operating condition data from step S3. For each historical data segment, the corresponding "optimal control command" label was determined as follows: under that historical operating condition, the set of fan speed, nozzle orifice diameter, and hot air temperature setpoints actually adopted by the system and subsequently verified in production (e.g., qualified product quality, low energy consumption) were recorded as "quasi-optimal" commands; or, through an offline process optimization simulator, the historical operating conditions were back-optimized using energy efficiency and product quality as indicators to calculate the theoretically optimal set of control commands. This resulted in a large number of [operating condition sequence, optimal control command] sample pairs.

[0074] During training, weighted mean squared error is used as the loss function. The specific form of this loss function is: .

[0075] in, , as well as These represent the fan speed increment, nozzle orifice adjustment steps, and hot air temperature offset, respectively, in the network prediction output.

[0076] , as well as These represent the true optimal values ​​corresponding to the labels in the training data.

[0077] α, β, and γ represent the error weights for the three output components: fan speed increment, nozzle adjustment steps, and hot air temperature deviation, respectively. Based on prior knowledge and experiments, these weights are set to α=1.0, β=0.8, and γ=1.2 to reflect the significant impact of temperature control on energy efficiency and the direct and precise effect of nozzle adjustment on quality. Through supervised training using this weighted loss function, the network is guided to automatically balance different control objectives during learning, approximating the globally optimal control strategy.

[0078] The core of step S4 above lies in a lightweight neural network designed specifically for real-time control. It receives regularized, multi-dimensional historical data, extracts spatiotemporal features through multi-layer convolution, focuses key information using an attention mechanism, and ultimately directly maps to the optimal control command. Its compact structure ensures millisecond-level inference speed, while supervised training based on historical data endows it with the ability to approximate optimal decisions. The network's output serves as the source of intelligent commands that drive subsequent physical actuators (fans, nozzle regulators, heaters) to perform precise adaptive regulation.

[0079] In step S5, the variable frequency fan drive motor is controlled according to the optimal fan speed command, the piezoelectric ceramic fine-tuning mechanism is driven to change the nozzle outlet geometry according to the optimal nozzle orifice diameter adjustment amount, and the electric heater power output is adjusted according to the hot air temperature correction value, including the following steps: The S501 variable frequency fan is driven by a permanent magnet synchronous motor. The motor has a built-in high-precision encoder that detects the actual speed in real time and feeds it back to the controller, forming a closed-loop speed control.

[0080] The controller converts the received speed increment command into a pulse width modulation signal and outputs it to the frequency converter. The execution cycle of this control loop is 5 milliseconds, and the final speed control accuracy is ±5 revolutions per minute.

[0081] The rated operating speed range of the fan is 1500 rpm to 3500 rpm.

[0082] The S502 piezoelectric ceramic fine-tuning mechanism consists of three sequentially mechanically connected parts: a ring-shaped piezoelectric stacked driver, a flexible hinge displacement amplification mechanism, and a conical nozzle mandrel.

[0083] The piezoelectric stack driver is driven by a high-precision, high-stability voltage amplifier. When a DC or quasi-DC driving voltage of 0 to 100 volts is applied, the piezoelectric material produces a corresponding axial displacement. This micro-displacement is amplified by a flexible hinge mechanism designed with a fixed amplification factor (e.g., 5-10 times), which directly drives the conical nozzle mandrel fixed to it to produce precise axial movement. The axial displacement of the mandrel directly changes the annular gap between it and the nozzle outer sleeve, thereby continuously and steplessly adjusting the equivalent outlet orifice diameter.

[0084] To ensure the accuracy and repeatability of the adjustment, the mechanism integrates high-resolution position sensors (such as micro strain gauges or capacitive displacement sensors) to measure the actual displacement of the mandrel in real time, forming a displacement closed-loop control. The controller (such as an FPGA-based servo controller) receives the adjustment step command output from the network, converts it into the target displacement, compares it with the actual displacement fed back by the sensor, and calculates the precise drive voltage value through a closed-loop control algorithm (such as PID) and outputs it to the voltage amplifier.

[0085] Through the closed-loop control described above, the nozzle outlet orifice diameter can be continuously adjusted between 180 micrometers and 250 micrometers.

[0086] Thanks to high-resolution (e.g., 16-bit or higher) digital-to-analog converters, position sensor feedback, and closed-loop control, the displacement adjustment resolution of the mechanism reaches 0.5 micrometers.

[0087] Due to the inherent fast response characteristics of piezoelectric actuators, combined with optimized control of the driving voltage change rate, the entire response time of the mechanism from receiving a stable electrical command to completing the target displacement adjustment and stabilizing (within the steady-state error band) is less than 5 milliseconds.

[0088] The S503 electric heater uses finned stainless steel tubular heating elements with a total power of 12 kW. The heater is designed as 12 independent power units, each rated at 1 kW. Each unit is equipped with an independent solid-state relay or SCR power regulator, supporting independent on / off or continuous power regulation (via PWM or phase angle control).

[0089] Temperature closed-loop control: The system receives a hot air temperature correction value as the setpoint and compares it with the actual hot air temperature detected by a high-response temperature sensor installed in the hot air duct. The deviation is fed into a digital proportional-integral-derivative (PID) controller. The PID controller calculates the total heating power (in watts) required to bring the temperature to the setpoint based on this deviation.

[0090] The system allocates the total heating power demand calculated by the PID controller to each heating unit according to the following deterministic steps: Step 1: Power Quantization. Round the total power requirement up or down to the nearest integer multiple of 100 watts.

[0091] Step 2: Full-Power Unit Allocation. Calculate how many "1 kW" units (i.e., the full power of one unit) are included in the quantified power value. The system activates the corresponding number of heating units and operates them at full power (1 kW). To balance service life, the system prioritizes activating units with the shortest historical cumulative operating time.

[0092] Step 3: Remaining power adjustment. If the quantified power value is not an integer multiple of 1 kilowatt, the system will turn on a new heating unit (preferably one with a shorter cumulative working time) and apply power to that unit that corresponds to the remaining power value (for example, if there is 300 watts remaining, the unit will be adjusted to run at 30% power).

[0093] Step 4: Turn off redundant units. All heating units not selected in the above steps should remain off.

[0094] Through the aforementioned PID control and group power distribution, the entire temperature control loop has a bandwidth of 0.5 Hz, which ensures that the actual fluctuation of the hot air temperature is controlled within ±2 degrees Celsius when following setpoint changes or resisting disturbances.

[0095] Step S5 uses high-response-speed, high-precision electromechanical actuators and closed-loop feedback control to ensure that the optimal control commands from the intelligent network can be accurately, quickly, and stably executed into the physical world, thereby ultimately achieving precise closed-loop control of the energy efficiency and quality of the drying process.

[0096] In step S6, the aim is to prevent the adhesion and deposition of high-viscosity slurry inside the nozzle by actively applying high-frequency micro-vibrations. This relies on an integrated vibration generation and intelligent control system. Specifically, the steps are as follows: The S601 micro-vibration excitation unit is embedded inside the nozzle wall, specifically located approximately 2 mm outside the slurry flow channel. The core vibrating component of this unit consists of a piezoelectric ceramic sheet and a high-density tungsten alloy mass block bonded together with a high-strength adhesive, forming a mechanical resonant structure.

[0097] S602, by designing the size, material, and weight of the piezoelectric ceramic sheet, allows the inherent frequency range of the resonant structure to be designed between 20 kHz and 50 kHz.

[0098] The steps S601 and S602 above together define the physical entity of the anti-clogging actuator. Its embedded layout ensures that vibration energy can be directly applied to the flow channel wall, while the resonant structure formed by the piezoelectric ceramic and the mass block provides the physical basis for generating the required high-frequency micro-amplitude vibration.

[0099] In S603, the system's vibration control module calculates the flow field state under the current operating conditions in real time, with the Reynolds number as the key criterion. The calculation formula is: Reynolds number (Re) = (hot air density × air velocity × nozzle characteristic length) / slurry dynamic viscosity. All parameters in the formula are derived from real-time system monitoring: air density and velocity are from the fan system sensors, the nozzle characteristic length is a known design constant, and the slurry dynamic viscosity is provided by an online rheological sensor. When the calculated Reynolds number is below 2000, the vibration control module determines that the current flow state near the nozzle wall is dominated by laminar flow. Under this flow state, the slurry is prone to forming a stable velocity boundary layer, leading to a significantly increased risk of material deposition.

[0100] S604, once the above triggering condition (Re < 2000) is met, the vibration control module immediately activates the preventative vibration damping mode. It generates and drives the vibration signal according to the following steps: Signal parameter settings: The module controls the digital signal synthesizer to generate a standard sine wave signal with a frequency of 30 kHz (30 kHz). To achieve a peak displacement of 2 micrometers (2 μm) at the output of the piezoelectric ceramic resonant structure, based on the voltage-displacement coefficient of the piezoelectric ceramic (e.g., d33 is approximately on the order of 300-500 pm / V) and the resonant amplification factor, the required peak driving voltage is determined to be approximately 80 to 120 volts (V) through pre-experimental calibration. The amplitude of the sine wave voltage output by the signal synthesizer is then set according to this target.

[0101] Signal amplification and driving: The aforementioned low-voltage sinusoidal signal is fed into a wideband power amplifier for voltage amplification. The high-voltage sinusoidal signal (peak value 80-120V) output by the amplifier is directly applied to the two electrodes of the piezoelectric ceramic sheet embedded in the nozzle.

[0102] Vibration Generation and Transmission: Under alternating high voltage excitation, the piezoelectric ceramic sheet generates microscopic deformation vibrations at the same frequency (30kHz) based on the inverse piezoelectric effect. This vibration is transmitted and coupled through a resonant structure formed by a tungsten alloy mass block bonded to it, and is ultimately effectively transmitted to the nozzle wall. As a result, a high-frequency (30kHz), small-amplitude (peak displacement approximately 2μm) periodic mechanical disturbance is introduced near the slurry channel wall.

[0103] S605, in order to monitor the effectiveness of anti-blocking measures in real time and provide a basis for the prediction model, the system monitors and provides feedback on the working status of the micro-vibration excitation unit. The specific implementation is as follows: State Awareness Scheme: The system achieves state awareness by monitoring the drive current. A high-precision, non-inductive sampling resistor is connected in series in the drive circuit between the power amplifier and the piezoelectric ceramic sheet. By measuring the voltage across this resistor, the real-time drive current waveform flowing through the piezoelectric ceramic sheet can be obtained.

[0104] Signal Processing and Information Extraction: The acquired real-time drive current waveform is fed into a signal processing unit (such as an analog-to-digital converter and processor integrated into the vibration control module). By performing spectral analysis (such as Fast Fourier Transform) on the current waveform, the current excitation frequency can be accurately extracted. Simultaneously, by analyzing the amplitude of the current waveform and comparing it with the known drive voltage amplitude, combined with the equivalent impedance characteristics of the piezoelectric ceramic, the actual vibration intensity of the piezoelectric ceramic sheet can be calculated, thereby estimating the peak displacement of the generated mechanical vibration.

[0105] Information Feedback: The processed real-time vibration status information, including the actual vibration frequency and estimated peak displacement, is encapsulated into data packets and transmitted in real time to the system's blockage risk prediction model. These data serve as key input variables to evaluate the effectiveness of the current micro-vibrations (e.g., whether the actual amplitude reaches the preset 2-micrometer target) and participate in the calculation of long-term deposition trends and blockage risk.

[0106] Step S6 implements a dynamic and intelligent preventative physical anti-clogging mechanism. It is not continuous or simply timed vibration, but rather, based on real-time assessment of the current flow state (Reynolds number), it precisely applies high-frequency micro-vibrations with specific parameters at the most critical moment (laminar flow conditions). This closed-loop "sensing-judgment-execution-feedback" not only directly suppresses deposition but also provides crucial state data for higher-level predictive maintenance (step S7), forming a multi-level anti-clogging system from short-term prevention to long-term prediction.

[0107] Step S7: This step builds a data-driven predictive model to proactively assess the risk of nozzle clogging and automatically execute pre-defined enhanced intervention measures when the risk exceeds a threshold, thereby transforming downtime maintenance from "post-event handling" to "pre-event prevention".

[0108] The S701 congestion risk prediction model is built and trained using a gradient boosting decision tree algorithm (e.g., using the XGBoost or LightGBM library).

[0109] The training data for this model was prepared from the dryer's historical operating records through the following steps: Extract continuous runtime sequence data from the historical database of the dryer control system over the past year. This data should include the original records of the input characteristics (viscosity, runtime, etc.) defined in step S702.

[0110] The extracted raw data is preprocessed, including: handling data loss caused by sensor communication interruption (using interpolation or previous values ​​to fill in the gaps); removing obvious outliers using sliding window statistical methods; and aligning all feature variables by timestamp and sampling them at a fixed frequency (e.g., one sample point per minute).

[0111] Sample Labeling: To build a supervised learning model, each cleaned data sample needs to be labeled. The label is defined as follows: Taking any data sampling time as a baseline, if the system records an unplanned shutdown or differential pressure alarm event caused by nozzle blockage within 30 minutes after that time, the data sample corresponding to that time is labeled as a positive sample (label value 1, representing high blockage risk); otherwise, it is labeled as a negative sample (label value 0, representing low blockage risk). Data occurring in the initial period after each planned cleanup can be considered as negative samples or subject to special treatment, depending on the circumstances.

[0112] After the above cleaning and labeling, a structured training dataset containing approximately 100,000 hours of effective working conditions (corresponding to hundreds of thousands to millions of time point samples) is finally obtained, in which each sample contains a set of input feature values ​​and a clear binary risk label.

[0113] In the real-time operation phase of S702, the model's input consists of the following four key features, all of which are obtained from the system's real-time database or maintenance logs: Current slurry viscosity: The latest viscosity data (in Pa·s) measured and provided in real time by an online rheological sensor (see step S1).

[0114] Cumulative nozzle runtime: The total operating time (in hours) of the nozzle, automatically accumulated and stored by the internal timer of the control system since the last thorough cleaning or replacement maintenance.

[0115] Vibration energy decay rate: This feature is used to quantify the performance decay trend of the micro-vibration unit. The specific calculation method is as follows: First, obtain the real-time amplitude of the micro-vibration unit under the current operating conditions, and the reference amplitude of the unit in its initial healthy state under the same typical operating conditions. Divide the current amplitude by the reference amplitude to obtain a ratio. Then, calculate the natural logarithmic rate of change of this ratio over a fixed time window (e.g., 1 hour).

[0116] Previous cleanup interval: The length of time (in hours) elapsed from the timestamp of the last preventative cleanup or maintenance operation recorded in the system log to the current moment.

[0117] S703, the trained model takes the aforementioned four features as input and outputs a probability value between 0 and 1, representing the likelihood of nozzle blockage occurring within the next 30 minutes as predicted by the model. This predictive model is driven by front-end sensor data streams and performs real-time inference and updates once per second to ensure the timeliness of risk warnings.

[0118] S704, the system presets a fixed risk probability threshold (set to 5% in this embodiment). When the blockage probability value output by the model exceeds this threshold, the system's control unit immediately and automatically triggers a set of combined preventive intervention commands, and simultaneously sends them to the corresponding actuators: Enhanced vibration command: The control unit sends a command to the drive controller of the micro-vibration excitation unit, requesting that its vibration amplitude be increased from the set value of conventional preventive vibration (e.g., 2 micrometer peak displacement) to a higher enhanced set value (e.g., 3 micrometer peak displacement). This command specifically translates into adjusting the output amplitude of the digital signal synthesizer, or directly setting a new target value for the drive voltage (see step S604), to more actively disturb the near-wall flow field and strip away any thin layer deposits that may have begun to form.

[0119] Temporary orifice enlargement command: Simultaneously, the control unit sends a command to the controller of the piezoelectric ceramic fine-tuning mechanism, requesting it to temporarily enlarge the nozzle orifice diameter by a preset fixed increment (e.g., 5 micrometers) based on the current real-time value. This command is converted into an additional displacement pulse required by the piezoelectric actuator (see step S502), which aims to reduce the local velocity gradient and shear stress of the slurry in the nozzle throat region by instantaneously increasing the flow area, thereby directly weakening the core hydrodynamic driving force that leads to material deposition.

[0120] Based on historical data verification, the intervention strategy triggered by the predictive model, as described above, can be initiated on average approximately 42 seconds before the actual occurrence of a congestion event. This lead time allows sufficient time for the intervention measures to take effect, thus successfully avoiding the vast majority of unplanned downtime caused by gradual congestion.

[0121] Step S7 integrates machine learning prediction and automated control to achieve predictive maintenance. It doesn't simply respond to past anomalies, but rather analyzes multi-dimensional characteristic trends to quantify risks in advance and proactively adjust system parameters before critical points. This model-based, forward-looking intervention mechanism, synergistically with the real-time vibration prevention in step S6, constitutes a comprehensive prevention system ranging from "normal minor prevention" to "risk-enhanced intervention," a key intelligent module ensuring the long-term, continuous, and efficient operation of the system.

[0122] In addition, the system includes an independently operating energy efficiency-quality co-optimization module. This module operates independently of the aforementioned millisecond-level real-time control loop, and starts calculations at a fixed period of 500 milliseconds.

[0123] Within each computation cycle, this module performs the following steps: Data Acquisition and Target Definition: The module first collects current operating data, including real-time hot air temperature, slurry evaporation rate, total input heating power, and median droplet size. Based on this data, the module establishes two targets that need to be optimized collaboratively: The primary goal is to maximize thermal energy utilization, which means maximizing the percentage of effective heat used for evaporating water relative to the total input heat.

[0124] The second objective is to minimize droplet size deviation, that is, to make the actual median diameter of the droplet as close as possible to the preset target diameter. The deviation is expressed as a percentage of the absolute value of the difference between the actual value and the target value relative to the target value.

[0125] Optimization constraints must be set: The optimization of the two objectives mentioned above must be carried out under the following physical and security constraints: The percentage deviation of the droplet diameter must not exceed five percent.

[0126] The hot air temperature should not fluctuate within a range of plus or minus five degrees Celsius around its set value.

[0127] The fan speed must not exceed the maximum permissible speed specified on its nameplate or safety regulations.

[0128] Optimization process: The module employs a mathematical optimization algorithm called "sequential quadratic programming" to find the optimal balance point that satisfies the above constraints. In practice, this can be achieved by calling built-in or open-source optimization algorithm libraries. The algorithm iteratively calculates and weighs two competing objectives (energy efficiency and quality), ultimately outputting a set of recommended process parameters that achieve the best overall effect under the current constraints, such as a recommended hot air temperature setpoint and a fan speed reference value.

[0129] Network parameter calibration: The module compares the optimal process parameter recommendations calculated above with the control commands generated in real time by the lightweight spatiotemporal feature extraction network under the current operating conditions. Based on the difference between the two, the module generates a slowly updated correction signal. For example, every few minutes, this signal makes a subtle adjustment to the weight parameters in the neural network responsible for generating key control commands, so that the future output decisions of the neural network are more inclined to approach the optimal balance point between energy efficiency and quality calculated by the optimization module. This process ensures that the system can adaptively maintain the optimal operating range of high efficiency and high quality in the long-term operation.

[0130] On the other hand, the energy-saving and efficiency-enhancing system for airflow spray dryers based on big data includes an online rheological sensing unit, a high-speed droplet imaging unit, a multi-source data fusion unit, a lightweight spatiotemporal feature extraction unit, an adaptive execution unit, a micro-vibration anti-clogging unit, and a clogging risk prediction unit. All units are interconnected via industrial Ethernet, and the data transmission protocol adopts the time-sensitive networking standard to ensure end-to-end latency of less than fifteen milliseconds. The system as a whole achieves millisecond-level perception-decision-execution closed loop for the drying process of high-viscosity slurries, completely resolving the problems of parameter rigidity, response lag, and the contradiction between energy efficiency and quality in traditional technologies.

[0131] Actual measurements show that when processing pectin solutions with viscosities increasing from 50 Pa·s to 800 Pa·s, this invention stabilizes the standard deviation of droplet diameter within 9 micrometers, increases thermal energy utilization to 82%, reduces product agglomeration rate to below 2%, and reduces unplanned downtime by 95%. Simultaneously, ineffective heating energy consumption is reduced by 28%, resulting in annual energy savings of 1.2 million kilowatt-hours.

[0132] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects.

[0133] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for improving energy efficiency and enhancing the effectiveness of an airflow spray dryer based on big data, characterized in that: include: Real-time viscosity data of the slurry before it enters the atomizer is obtained using an online rheology sensor. The high-speed imaging module continuously acquires images of the droplet swarm at the outlet of the atomization chamber and calculates the standard deviation and median particle size of the current droplet diameter distribution. The viscosity data, the calculated droplet diameter distribution data, the hot air temperature, the hot air speed, the nozzle orifice status, and the ambient humidity data are fused into a multi-source heterogeneous time-series data stream and input into a pre-trained lightweight spatiotemporal feature extraction network. The lightweight spatiotemporal feature extraction network outputs the optimal fan speed command, the optimal nozzle orifice adjustment amount, and the hot air temperature correction value under the current operating conditions. The variable frequency fan is controlled according to the optimal fan speed command, the piezoelectric ceramic fine-tuning mechanism is driven to change the nozzle outlet geometry according to the optimal nozzle orifice diameter adjustment amount, and the power output of the electric heater is adjusted according to the hot air temperature correction value. A micro-vibration excitation unit is integrated inside the nozzle. Its vibration frequency and amplitude are dynamically set according to the current viscosity and Reynolds number to suppress material deposition. A nozzle clogging risk prediction model is established. The model takes viscosity, runtime, historical deposition rate and current vibration status as inputs. When the predicted clogging probability exceeds a preset threshold, preventive intervention is triggered.

2. The method according to claim 1, characterized in that, The online rheological sensor acquires slurry viscosity data in real time, including: an online rheological sensor with a concentric cylindrical structure, the inner cylinder of which is driven to rotate by a servo motor, the slurry shear stress is inferred by measuring the torque of the inner cylinder, and the dynamic viscosity is calculated by combining the known shear rate, the sampling frequency of the viscosity data is not less than 1 kilohertz.

3. The method according to claim 1, characterized in that, The high-speed imaging module performs image acquisition and calculates the droplet diameter distribution, including: using an image sensor with a frame rate of 5,000 frames per second, in conjunction with a ring light source, and performing edge detection and morphological analysis through an image processing unit deployed on a field-programmable gate array chip to calculate the standard deviation and median particle size of the droplet diameter distribution.

4. The method according to claim 1, characterized in that, The data are fused into a multi-source heterogeneous time-series data stream, including: resampling data with different sampling rates to a reference frequency of 1 kilohertz through a timestamp alignment mechanism, interpolating missing data points, and removing and replacing outliers, ultimately forming a multi-dimensional time-series vector containing viscosity, droplet diameter standard deviation, median particle size, hot air temperature, hot air velocity, ambient humidity, and cumulative nozzle runtime.

5. The method according to claim 1, characterized in that, The lightweight spatiotemporal feature extraction network is constructed based on a one-dimensional convolution and attention mechanism. Its input is the multi-source heterogeneous temporal data stream. The network architecture includes multiple one-dimensional convolutional blocks and channel attention modules. The output layer is used to directly map and output the fan speed command, nozzle orifice adjustment amount, and hot air temperature correction value.

6. The method according to claim 1, characterized in that, The driving piezoelectric ceramic fine-tuning mechanism changes the nozzle outlet geometry, comprising: the piezoelectric ceramic fine-tuning mechanism is composed of a ring piezoelectric stacked driver, a flexible hinge amplification mechanism and a conical nozzle mandrel, and the conical nozzle mandrel is driven to move by changing the driving voltage to continuously adjust the nozzle outlet orifice diameter.

7. The method according to claim 1, characterized in that, The micro-vibration excitation unit consists of a piezoelectric ceramic sheet embedded inside the nozzle wall and a mass block forming a resonant structure. Its vibration control module dynamically selects the excitation frequency and amplitude based on the currently calculated Reynolds number.

8. The method according to claim 1, characterized in that, The nozzle clogging risk prediction model is established by: constructing a model using a gradient boosting decision tree algorithm to output the probability value of clogging occurring within a set time period in the future; when the probability value exceeds a preset threshold, the triggering of preventive intervention includes increasing the amplitude of the micro-vibration excitation unit and instructing the piezoelectric ceramic fine-tuning mechanism to expand the nozzle orifice diameter.

9. The method according to any one of claims 1-8, characterized in that, Also includes: Energy efficiency-quality co-optimization is performed, with thermal energy utilization rate and droplet diameter deviation as dual objective functions for constrained optimization, and the optimization results are used to calibrate the output of the lightweight spatiotemporal feature extraction network.

10. A big data-based energy-saving and efficiency-enhancing system for airflow spray dryers, characterized in that: include: An online rheology sensing unit is used to acquire real-time viscosity data of the slurry before it enters the atomizer via an online rheology sensor. The high-speed droplet imaging unit is used to continuously acquire images of the droplet group at the outlet of the atomization chamber through the high-speed imaging module, and to calculate the standard deviation and median particle size of the current droplet diameter distribution. The multi-source data fusion unit is used to fuse the viscosity data, the calculated droplet diameter distribution data, the hot air temperature, the hot air speed, the nozzle orifice status, and the ambient humidity data into a multi-source heterogeneous time-series data stream. The lightweight spatiotemporal feature extraction unit is used to input the time-series data stream into the pre-trained lightweight spatiotemporal feature extraction network and output the optimal fan speed command, the optimal nozzle orifice adjustment amount, and the hot air temperature correction value. The adaptive execution unit includes a variable frequency fan that responds to the optimal fan speed command, a piezoelectric ceramic fine-tuning mechanism that responds to the optimal nozzle orifice diameter adjustment amount, and an electric heater that responds to the hot air temperature correction value. The micro-vibration anti-clogging unit, integrated inside the nozzle, is used to dynamically apply preventative vibrations based on the current viscosity and Reynolds number; The clogging risk prediction unit is used to assess the probability of nozzle clogging by taking viscosity, runtime, historical deposition rate and current vibration status as inputs, and to trigger preventive measures when the clogging exceeds a preset threshold.

Images