A real-time data interaction processing method in a spray drying process

Abstract

The present application relates to the technical field of data processing, in particular to a real-time data interaction processing method in a spray drying process. The spray drying tower feed flow, atomization pressure, hot air temperature and exhaust temperature sensor data and basic sampling period are obtained; a sliding window based on the change of exhaust temperature is established according to the heat and mass transfer lag time constant of the material; the low-frequency sampling data is reconstructed by using the polynomial fitting method of Lagrange interpolation in the window; the high-frequency sampling data and the up-sampling data are time-aligned at the window termination time, and a multi-source feature vector is generated; the interactive message is packaged and sent to the atomization controller through the edge gateway, and the interactive closed loop is completed by receiving and executing feedback data. The present application eliminates the data timestamp drift caused by the difference in sensor sampling frequency, improves the data consistency at the time of issuing control instructions, and reduces the control system overshoot frequency caused by time sequence misalignment.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and specifically to a real-time data interaction processing method in the spray drying process. Background Technology

[0002] The spray drying process requires monitoring of feed flow rate, atomization pressure, hot air temperature, and exhaust temperature. Due to limitations in sensor physical characteristics, pressure and flow sensors have high sampling frequencies, while temperature sensors have low sampling frequencies. Existing main control systems typically use a fixed-period polling mechanism or a unified timestamp capture method to directly pull the latest sampling values ​​from each sensor from the data bus before issuing control commands to the atomization controller. After receiving these values, the controller adjusts the atomization valve opening and motor speed according to preset linkage logic. This data interaction method forcibly binds data from different sampling periods to be sent within the same interaction period, without addressing the differences in physical response times between different sensors.

[0003] In the actual operation of a spray drying tower, the material needs to undergo heat and mass transfer processes from the feed end to the exhaust detection point. Changes in exhaust temperature lag behind changes in pressure and flow rate. Using a fixed-period data extraction method leads to a physical time misalignment between high-frequency pressure data and low-frequency temperature data. The controller, based on this time-misaligned data, fails to match the actual physical state inside the drying tower, causing overshoot and lag in the control system. Summary of the Invention

[0004] The purpose of this invention is to provide a real-time data interaction processing method in the spray drying process, which can effectively solve the problems in the background art mentioned above.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A real-time data interaction processing method for a spray drying process includes: acquiring feed flow sensor data, atomization pressure sensor data, hot air temperature sensor data, and exhaust temperature sensor data of a spray drying tower, and determining the basic sampling period of the feed flow sensor data, the atomization pressure sensor data, the hot air temperature sensor data, and the exhaust temperature sensor data.

[0007] Based on the material heat and mass transfer lag time constant in the spray drying process, a sliding window is established with the change in the exhaust temperature sensor data as the reference.

[0008] Within the sliding window, a polynomial fitting method based on Lagrange interpolation is used to upsample and reconstruct the low-frequency sampling data in the basic sampling period to obtain upsampled data;

[0009] The high-frequency sampled data in the basic sampling period and the upsampled data are time-aligned with each other at the end of the sliding window to generate a multi-source feature vector.

[0010] The multi-source feature vectors are encapsulated into an interactive message in a preset format. The interactive message is sent to the atomization controller via the edge gateway, and the execution feedback data returned by the atomization controller is received to complete a data interaction loop.

[0011] Preferably, the step of establishing a sliding window based on the changes in the exhaust temperature sensor data, according to the material heat and mass transfer lag time constant in the spray drying process, includes: extracting multiple consecutive historical data points of the exhaust temperature sensor data before the current moment.

[0012] Calculate the slope of temperature change corresponding to the consecutive historical data points;

[0013] The temperature change slope is input into a preset lag time constant mapping model, and the target lag time constant corresponding to the current moment is output.

[0014] The sliding window is defined between the current time as the end point of the time axis and the time when the target lag time constant is subtracted from the current time as the start point of the time axis.

[0015] When the next data point of the exhaust temperature sensor is detected, the time axis start point and time axis end point of the sliding window are shifted according to the target lag time constant.

[0016] Preferably, the step of upsampling and reconstructing the low-frequency sampling data in the basic sampling period using a polynomial fitting method based on Lagrange interpolation within the sliding window to obtain upsampled data includes: identifying the low-frequency sampling data and high-frequency sampling data in the basic sampling period, and marking the timestamp of the low-frequency sampling data within the sliding window as the original interpolation node;

[0017] Obtain the high-frequency sampling frequency corresponding to the high-frequency sampling data, and use the high-frequency sampling frequency as the reconstruction frequency to uniformly insert virtual interpolation nodes between two adjacent original interpolation nodes;

[0018] Based on the low-frequency sampled data values ​​corresponding to all the original interpolation nodes within the sliding window, a Lagrange interpolation polynomial is constructed.

[0019] Substitute the timestamp coordinates of the virtual interpolation node into the Lagrange interpolation polynomial to calculate the low-frequency sampled data value corresponding to the virtual interpolation node, and use the data set containing the original interpolation node and the virtual interpolation node as the upsampled data.

[0020] Preferably, the step of aligning the high-frequency sampling data in the basic sampling period with the upsampled data at the end of the sliding window to generate a multi-source feature vector includes: extracting the timestamp corresponding to the end of the sliding window as a unified alignment timestamp;

[0021] Extract real-time high-frequency data points from the high-frequency sampled data whose timestamps are equal to the unified alignment timestamp, and extract real-time upsampled data points from the upsampled data whose timestamps are equal to the unified alignment timestamp;

[0022] The real-time high-frequency data points and the real-time upsampled data points are processed to unify their data formats. The processed real-time high-frequency data points and the real-time upsampled data points are then concatenated according to a preset arrangement order of the sensor physical addresses to generate a fixed-length multi-source feature vector.

[0023] The unified alignment timestamp and the target lag time constant are added to the beginning of the multi-source feature vector, and a cyclic redundancy check code is added to the end of the multi-source feature vector.

[0024] Preferably, the step of encapsulating the multi-source feature vector into an interactive message of a preset format and sending the interactive message to the atomization controller via the edge gateway includes: parsing the communication protocol stack of the edge gateway and determining the destination port number and physical layer frame length limit corresponding to the atomization controller;

[0025] The multi-source feature vector is segmented according to the physical layer frame length limit to obtain multi-frame feature segmentation data.

[0026] A message header field containing the destination port number is added to the header of each frame of the multi-frame feature fragment data, and a fragment sequence number field is added to the tail of each frame to generate the interaction message;

[0027] The interaction messages are sent to the industrial Ethernet bus through the sending queue of the edge gateway, in the order of the fragment sequence number field.

[0028] The system monitors the industrial Ethernet bus. When it receives a response frame that matches the fragment sequence number field, it extracts the status register value from the response frame as the execution feedback data.

[0029] Preferably, receiving the execution feedback data returned by the atomization controller to complete a data interaction loop includes: inputting the execution feedback data into a preset status word parsing mapping table, and outputting the valve opening update value and motor speed update value of the atomization controller;

[0030] The updated valve opening value and the updated motor speed value are superimposed on the corresponding feed flow rate dimension and atomization pressure dimension in the multi-source feature vector to generate the updated feature vector.

[0031] The updated feature vector is stored in a circular buffer queue, and the write pointer of the circular buffer queue is moved forward by one step.

[0032] When the difference between the write pointer and the read pointer of the circular buffer queue is detected to reach a preset queue depth threshold, the encapsulation operation of the next interaction message is triggered, and the read pointer is moved to a position that coincides with the write pointer.

[0033] Preferably, the step of inputting the temperature change slope into a preset lag time constant mapping model and outputting the target lag time constant corresponding to the current moment includes: the lag time constant mapping model is a set of piecewise linear interpolation functions, and the set of piecewise linear interpolation functions includes multiple sets of slope interval thresholds and multiple sets of constant interval values;

[0034] The slope of the temperature change is compared one by one with the multiple sets of slope interval thresholds to determine the target slope interval in which the slope of the temperature change is located.

[0035] Extract the slope values ​​of the left and right endpoints corresponding to the target slope interval, and extract the constant values ​​of the left and right endpoints corresponding to the target slope interval;

[0036] A local linear equation is constructed based on the slope values ​​of the left and right endpoints and the constant values ​​of the left and right endpoints. The slope of the temperature change is substituted into the local linear equation to calculate the target lag time constant.

[0037] Preferably, the step of constructing a Lagrange interpolation polynomial based on the low-frequency sampled data values ​​corresponding to all the original interpolation nodes within the sliding window includes: counting the total number of the original interpolation nodes within the sliding window;

[0038] When the total number is greater than the preset node number threshold, the preset node number threshold target interpolation nodes that are closest to the end time of the sliding window are extracted from the original interpolation nodes in chronological order.

[0039] For each target interpolation node, calculate the absolute value of the difference between the timestamp of other target interpolation nodes (excluding the current target interpolation node) and the timestamp of the current target interpolation node;

[0040] The absolute values ​​of the differences are multiplied in ascending order to obtain the product of the denominator and the product of the numerator.

[0041] The low-frequency sampled data values ​​corresponding to each target interpolation node, the product of the denominator and the product of the numerator are weighted and summed to generate the coefficients of the Lagrange interpolation polynomial.

[0042] Preferably, the data format unification processing of the real-time high-frequency data points and the real-time upsampled data points includes: extracting the first data bit width of the real-time high-frequency data points and the second data bit width of the real-time upsampled data points, and determining the maximum bit width between the first data bit width and the second data bit width as the target data bit width;

[0043] When it is detected that the data bit width of the real-time high-frequency data point is less than the target data bit width, a first preset value is filled before the most significant bit of the real-time high-frequency data point; when it is detected that the data bit width of the real-time upsampled data point is less than the target data bit width, a second preset value is filled before the most significant bit of the real-time upsampled data point.

[0044] Both the filled real-time high-frequency data points and the filled real-time upsampled data points are converted into data streams in a preset floating-point format.

[0045] Preferably, the step of segmenting the multi-source feature vector according to the physical layer frame length limit to obtain multi-frame feature segmentation data includes: calculating the difference between the physical layer frame length limit and the length of the message header field and the length of the segmentation sequence number field to obtain the maximum effective payload length of a single frame;

[0046] Divide the total data length of the multi-source feature vector by the maximum effective payload length of a single frame, and round up to obtain the total number of target fragment frames.

[0047] The multi-source feature vector is divided into equal-length segments according to the total number of target segments. When the data length of the last segment is less than the maximum effective payload length of a single frame, a third preset value is filled at the end of the last segment so that the data length of the last segment after filling is equal to the maximum effective payload length of a single frame. All segments after segmentation are used as the multi-frame feature segment data.

[0048] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0049] 1. This invention establishes a sliding window based on the material heat and mass transfer lag time constant and the exhaust temperature change. Within the window, a polynomial fitting method using Lagrange interpolation is employed to upsample and reconstruct low-frequency sampled data, ensuring that the low-frequency data has the same sampling density as the high-frequency data. The high-frequency sampled data and the upsampled data are time-aligned at the end of the sliding window and encapsulated into an interactive message for transmission. This processing method eliminates data timestamp drift caused by differences in sensor sampling frequencies and physical response lags, ensuring that the multi-source feature vectors transmitted to the atomization controller are physically consistent, thus reducing the overshoot frequency of the control system caused by data timing misalignment.

[0050] 2. By extracting historical data points of exhaust temperature to calculate the slope of temperature change, and using a mapping model constructed from piecewise linear interpolation functions to output the target lag time constant, adaptive adjustment of the sliding window period according to operating conditions is achieved. When constructing the Lagrange interpolation polynomial, the computational load on edge computing nodes is reduced by limiting the number of nodes participating in the interpolation operation. Aligned data is padded according to the target data bit width and converted into a floating-point format data stream. Combined with equal-length segmentation and end padding based on the maximum effective payload length of a single frame, frame structure consistency of feature vectors during transmission on the industrial Ethernet bus is ensured. By mapping execution feedback data to valve opening and motor speed update values ​​and superimposing them onto the feature vector and storing them in a circular buffer queue, the next encapsulation operation is triggered based on the queue depth threshold, maintaining the temporal continuity of the data interaction loop. Attached Figure Description

[0051] Figure 1 This is a flowchart of the overall method for real-time data interaction processing in spray drying according to the present invention;

[0052] Figure 2 The flowchart of the adaptive sliding window creation method of the present invention is shown below;

[0053] Figure 3 This is a flowchart of the Lagrange interpolation upsampling reconstruction method of the present invention;

[0054] Figure 4 This is a flowchart of the multi-source data temporal alignment and feature vector generation process of the present invention;

[0055] Figure 5 This is a flowchart illustrating the interactive message encapsulation and bus transmission process of the present invention.

[0056] Figure 6 This is a flowchart of the execution feedback data parsing and interactive closed-loop control process of the present invention. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] Please refer to Figure 1 This embodiment provides a real-time data interaction processing method in the spray drying process. This embodiment is applied to a real-time data interaction processing system for a spray drying tower. The system includes a feed flow sensor, an atomization pressure sensor, a hot air temperature sensor, and an exhaust air temperature sensor that are matched with the spray drying tower cavity, an edge computing node connected to the sensor signal output terminal, an edge gateway that is communicatively connected to the edge computing node, and an atomization controller that is connected to the edge gateway via an industrial Ethernet bus.

[0059] Specifically, the edge computing node establishes signal connections with each sensor through analog acquisition channels to acquire real-time data from the feed flow sensor, atomizing pressure sensor, hot air temperature sensor, and exhaust temperature sensor. The feed flow sensor and atomizing pressure sensor use piezoelectric sensing elements, and their output analog signals exhibit high-frequency response characteristics as the measured physical quantity changes. The hot air temperature sensor and exhaust temperature sensor use platinum resistance sensing elements, and their output analog signals exhibit low-frequency response characteristics as the measured temperature changes. The edge computing node performs analog-to-digital conversion on the analog signals output by each sensor to obtain the corresponding digital sensor data. Simultaneously, it extracts the inherent sampling period of each sensor to determine the basic sampling period corresponding to all sensor data.

[0060] Define the inherent sampling period of the feed flow sensor as follows: The inherent sampling period of the atomizing pressure sensor is The inherent sampling period of the hot air temperature sensor is The inherent sampling period of the exhaust temperature sensor is Basic sampling period The smallest positive value that satisfies the following condition: ;

[0061] in, The least common multiple (LCM) operation function converts the inherent sampling periods of each sensor to a unified time unit during the calculation process, ensuring that all input parameters have consistent dimensions. In this embodiment, the inherent sampling periods of the feed flow sensor and the atomizing pressure sensor are both 10ms, and the inherent sampling periods of the hot air temperature sensor and the exhaust air temperature sensor are both 100ms. The LCM operation yields a basic sampling period of 100ms. After determining the basic sampling period, the edge computing nodes use this period as a reference to perform time-series marking on all sensor data, providing a unified time reference for subsequent data processing.

[0062] Furthermore, the edge computing nodes establish a sliding window based on the material heat and mass transfer lag time constant in the spray drying process, using changes in exhaust temperature sensor data as a reference. The material heat and mass transfer lag time constant refers to the time interval between the material entering the chamber from the inlet of the spray drying tower, undergoing atomization and dispersion, exchanging heat and mass with hot air, and ultimately generating a detectable temperature change at the exhaust detection point. This constant corresponds to the physical lag characteristic between changes in feed parameters and the exhaust temperature response during the spray drying process. The material heat and mass transfer lag time constant in the spray drying process is defined as... The current data processing time is The time interval of the sliding window is defined as: ;

[0063] in, For the current moment The corresponding sliding window has its timeline ending at the current moment. The starting point of the timeline is the current moment minus the target lag time constant. In this embodiment, the lag time constant is predetermined through no-load and load calibration experiments of the spray drying tower, and is fixed at 1.5s, corresponding to a sliding window time span of 1.5s. The edge computing node monitors the update status of the exhaust temperature sensor data in real time. When the next data point of the exhaust temperature sensor data is detected, the starting point and ending point of the sliding window time axis are synchronously shifted according to the fixed lag time constant, so that the sliding window always covers the complete lag time interval before the current moment, ensuring that the sensor data within the window corresponds to the complete heat and mass transfer process of the same batch of materials.

[0064] After the sliding window is defined, the edge computing nodes within the sliding window use a polynomial fitting method based on Lagrange interpolation to upsample and reconstruct the low-frequency sampling data in the basic sampling period, obtaining upsampled data. The edge computing nodes first identify the low-frequency and high-frequency sampling data in the basic sampling period. Hot air temperature sensor data and exhaust air temperature sensor data with an inherent sampling period equal to the basic sampling period are marked as low-frequency sampling data, while feed flow sensor data and atomizing pressure sensor data with an inherent sampling period less than the basic sampling period are marked as high-frequency sampling data. The edge computing nodes extract all data points of the low-frequency sampling data within the sliding window and mark the timestamps corresponding to each data point as the original interpolation nodes. The set of timestamps for the original interpolation nodes is denoted as […]. The corresponding set of low-frequency sampled data values ​​is denoted as ,in This is the maximum index value of the original interpolation node within the sliding window.

[0065] Edge computing nodes acquire the high-frequency sampling frequency corresponding to the high-frequency sampling data, and use this high-frequency sampling frequency as the reconstruction frequency to uniformly insert virtual interpolation nodes between two adjacent original interpolation nodes. High-frequency sampling frequency The sampling period is the reciprocal of the inherent sampling period of the high-frequency sampling data. In this embodiment, the high-frequency sampling period is 10ms, corresponding to a high-frequency sampling frequency of 100Hz. The low-frequency sampling period is also 100ms, corresponding to a low-frequency sampling frequency of 10Hz. The reconstruction frequency is consistent with the high-frequency sampling frequency of 100Hz. The time interval between two adjacent original interpolation nodes is 100ms. Within this time interval, nine virtual interpolation nodes are evenly inserted. The time interval between two adjacent interpolation nodes (including original interpolation nodes and virtual interpolation nodes) is 10ms, which is completely consistent with the sampling interval of the high-frequency sampling data. The set of timestamps for the virtual interpolation nodes is denoted as... ,in For the first The and the first Between the nth original interpolation nodes The timestamp of each virtual interpolation node.

[0066] Based on the low-frequency sampled data values ​​corresponding to all the original interpolation nodes within the sliding window, the edge computing nodes construct a Lagrange interpolation polynomial. The expression for the nth degree Lagrange interpolation polynomial is: ;

[0067] in, To construct the complete Lagrange interpolation polynomial, The timestamp coordinates of the interpolation nodes. For the first The low-frequency sampled data values ​​corresponding to each original interpolation node. For the first The interpolation basis functions corresponding to the original interpolation nodes. The expressions for the interpolation basis functions are: ;

[0068] in, For the first The timestamps of the original interpolation nodes To exclude the first The timestamps of other original interpolation nodes besides the original interpolation nodes. This is the multiplication operator. The interpolation basis function satisfies the following property: when the input is a timestamp... hour, When entering a timestamp and hour, This ensures that the calculated values ​​of the Lagrange interpolation polynomial at all original interpolation nodes are completely consistent with the values ​​of the original low-frequency sampled data.

[0069] After the Lagrange interpolation polynomial is constructed, the edge computing nodes sequentially substitute the timestamp coordinates of all virtual interpolation nodes into the Lagrange interpolation polynomial to calculate the low-frequency sampled data value corresponding to each virtual interpolation node. The data set containing the timestamps and corresponding values ​​of all original interpolation nodes and virtual interpolation nodes is used as upsampled data. The sampling frequency of the upsampled data is completely consistent with the sampling frequency of the high-frequency sampled data, giving the originally low-frequency sampled temperature data the same time sampling density as the high-frequency sampled pressure and flow data, providing a unified time dimension benchmark for subsequent time-series alignment processing.

[0070] refer to Figure 4 After upsampling and reconstruction, the edge computing nodes time-align the high-frequency sampled data from the basic sampling period with the upsampled data at the end of the sliding window, generating a multi-source feature vector. The edge computing nodes then extract the timestamp corresponding to the end of the sliding window, which is the current time. The timestamp is used as a unified alignment. Edge computing nodes traverse all high-frequency sampled data within the sliding window, extracting real-time high-frequency data points whose timestamps are completely equal to the unified alignment timestamp. These real-time high-frequency data points include feed flow rate data and atomization pressure data. Simultaneously, they traverse upsampled data, extracting real-time upsampled data points whose timestamps are completely equal to the unified alignment timestamp. These real-time upsampled data points include hot air temperature data and exhaust air temperature data. Through this processing, all sensor data participating in feature vector construction corresponds to the same physical moment, eliminating the timing misalignment problem caused by differences in sampling frequency.

[0071] The edge computing node performs data format unification processing on the extracted real-time high-frequency data points and real-time upsampling data points. The processed real-time high-frequency data points and real-time upsampling data points are then concatenated according to a preset arrangement order of sensor physical addresses to generate a fixed-length multi-source feature vector. In this embodiment, the preset arrangement order of sensor physical addresses is: feed flow sensor, atomizing pressure sensor, hot air temperature sensor, and exhaust air temperature sensor. The main data segment of the concatenated feature vector contains four consecutive sensor data values. The edge computing node adds a unified alignment timestamp and a target hysteresis time constant to the beginning of the multi-source feature vector and a cyclic redundancy check code to the end, completing the construction of the multi-source feature vector. The complete structural expression of the multi-source feature vector is: ;

[0072] in, To construct the complete multi-source feature vector, For the aligned feed flow rate data, For the aligned atomization pressure data, For the aligned hot air temperature data, For the aligned exhaust temperature data, This is a Cyclic Redundancy Check (CRC) code. The calculation scope of the CRC code includes all bytes of data in the multi-source feature vector except for the check code itself. It uses the CRC32 check algorithm and generates a polynomial of 0x04C11DB7 to ensure the integrity and correctness of data during transmission.

[0073] After the multi-source feature vectors are constructed, the edge computing node encapsulates them into an interactive message in a preset format. This message is then sent to the atomization controller via the edge gateway. The edge node also receives execution feedback data from the atomization controller to complete a data interaction loop. The edge computing node transmits the multi-source feature vectors to the edge gateway, parses the gateway's communication protocol stack, and determines the destination port number and physical layer frame length limit corresponding to the atomization controller. In this embodiment, the edge gateway's communication protocol stack uses the PROFINETIO industrial Ethernet protocol, the destination port number for the atomization controller is 0x8892, and the physical layer frame length is limited to the Ethernet standard maximum transmission unit of 1500 bytes.

[0074] Edge computing nodes fragment multi-source feature vectors according to physical layer frame length limitations, resulting in multi-frame feature fragment data. A header field containing the destination port number is added to the beginning of each frame, and a fragment sequence number field is added to the end, generating an interactive message conforming to the protocol specification. The header field also includes fixed protocol fields such as source port number, protocol type, and message length. The fragment sequence number field contains the current fragment's sequence number and the total number of fragments, used for fragment reassembly at the receiving end. The edge gateway stores all generated interactive messages in a sending queue and sends them sequentially to the industrial Ethernet bus according to the fragment sequence number field order, ensuring that the atomization controller can receive and reassemble the multi-source feature vectors in the correct order.

[0075] While sending interaction messages, the edge gateway continuously monitors the industrial Ethernet bus. When it receives a response frame carrying a fragment sequence number field matching the fragment number field, it confirms that all interaction messages have been correctly received by the atomization controller and extracts the status register value from the response frame as execution feedback data. After receiving the complete multi-source feature vector, the atomization controller calculates the control parameters according to the preset control logic, executes the corresponding valve opening and motor speed adjustment operations, and writes the execution result and current operating status into the status register, generates a response frame, and sends it to the industrial Ethernet bus. The edge computing node parses the received execution feedback data and stores the parsed execution status parameters in its local storage unit, completing a complete data interaction loop.

[0076] Table 1. Basic sampling parameters of each sensor in this embodiment.

[0077] Table 1 clarifies the basic sampling parameters of each sensor involved in data interaction in this embodiment, providing a benchmark for determining the basic sampling period, identifying high and low frequency sampling data, and subsequent data format unification processing, ensuring that the processing of all sensor data has a unified parameter benchmark.

[0078] This embodiment fully realizes real-time interactive processing of multi-source sensor data during spray drying. A sliding window based on the material's heat and mass transfer lag time constant is used to determine the data processing time range, ensuring that the sensor data being processed corresponds to the complete heat and mass transfer process of the same batch of material within the drying tower. A polynomial fitting method based on Lagrange interpolation is used to upsample and reconstruct low-frequency data, giving it the same sampling density as high-frequency data. Time-series alignment at the end of the sliding window eliminates timestamp drift caused by differences in sampling frequencies among the multi-source data. Standardized message encapsulation and feedback data parsing complete the end-to-end data interaction loop, ensuring that the multi-source data sent to the atomization controller matches the actual physical state inside the drying tower.

[0079] In a preferred embodiment, this embodiment refines the process of establishing the sliding window based on the foregoing embodiments, so as to realize the adaptive adjustment of the sliding window period according to the spray drying conditions, and optimizes the encapsulation processing of interactive messages and the closed-loop processing of execution feedback data.

[0080] Specifically, refer to Figure 2 The edge computing node extracts multiple consecutive historical data points from the exhaust temperature sensor prior to the current moment. The number of historical data points is fixed at 5 using a preset parameter, corresponding to the exhaust temperature data from 5 consecutive sampling cycles prior to the current moment. The current moment is defined as... The set of timestamps corresponding to the extracted historical data points is ,in The corresponding set of exhaust temperature values ​​is The edge computing nodes calculate the corresponding temperature change slope based on extracted historical data points. A least squares method is used to perform linear fitting on the historical data points to obtain the calculated value of the temperature change slope. The formula for calculating the fitted slope is: ;

[0081] in, The slope of the temperature change of the exhaust air. In this embodiment, the total number of historical data points is... , For the first The timestamps corresponding to each historical data point For the first The exhaust temperature values ​​corresponding to each historical data point. The temperature change slope obtained by least squares fitting can eliminate random noise interference from individual data points, reflect the overall trend of exhaust temperature change under the current operating conditions, and provide stable input parameters for subsequent lag time constant calculation.

[0082] After the temperature change slope is calculated, the edge computing node inputs the temperature change slope into a pre-set lag time constant mapping model and outputs the target lag time constant corresponding to the current moment. The lag time constant mapping model is a set of piecewise linear interpolation functions, pre-constructed through full-condition calibration experiments of a spray drying tower. It includes multiple sets of slope interval thresholds and multiple sets of constant interval values, with each slope interval corresponding to a set of linear interpolation parameters. The edge computing node compares the calculated temperature change slope with each of the multiple slope interval thresholds to determine the target slope interval in which the temperature change slope falls. It extracts the slope values ​​at the left and right endpoints of the target slope interval, as well as the constant values ​​at the left and right endpoints of the target slope interval. Based on the extracted parameters, it constructs a local linear equation, substitutes the temperature change slope into the local linear equation, and calculates the target lag time constant. The formula for calculating the local linear interpolation is: ;

[0083] in, Current slope The corresponding target lag time constant, This represents the slope value at the left endpoint of the target slope interval. This represents the slope value at the right endpoint of the target slope interval. This represents the lag time constant value corresponding to the slope at the left endpoint. This represents the lag time constant value corresponding to the slope at the right endpoint. When the slope of the temperature change exceeds the preset maximum slope range, the edge computing node uses the constant value corresponding to the interval endpoint closest to the current slope as the target lag time constant to avoid calculation exceeding the limit.

[0084] After the target lag time constant is calculated, the edge computing node defines a sliding window between the current time as the end point of the time axis and the time point minus the target lag time constant as the start point of the time axis. The time span of the sliding window is completely consistent with the target lag time constant under the current operating condition and can adaptively adjust according to the trend of exhaust air temperature changes. When the exhaust air temperature changes drastically, the absolute value of the temperature change slope increases, the corresponding target lag time constant decreases, and the time span of the sliding window shortens, ensuring that the data within the window can quickly respond to changes in operating conditions. When the exhaust air temperature changes gradually, the absolute value of the temperature change slope decreases, the corresponding target lag time constant increases, and the time span of the sliding window extends, ensuring that the data within the window can cover the complete heat and mass transfer process. The edge computing node monitors the update status of the exhaust air temperature sensor data in real time. When it detects that the next data point of the exhaust air temperature sensor data has arrived, it recalculates the temperature change slope and the target lag time constant, and shifts the start and end points of the time axis of the sliding window according to the new target lag time constant, completing the adaptive update of the sliding window.

[0085] Table 2. Piecewise linear interpolation parameters for the time lag constant mapping model.

[0086] Table 2 clarifies the mapping relationship between the temperature change slope and the lag time constant in this embodiment, providing a calibration benchmark for the adaptive adjustment of the sliding window period, ensuring that the time span of the sliding window can match the actual heat and mass transfer lag process in the drying tower in real time, and covering the material change process under different operating conditions.

[0087] Further, refer to Figure 5After constructing the multi-source feature vectors, the edge computing node optimizes the encapsulation process of the interactive messages. The edge computing node parses the communication protocol stack of the edge gateway, determines the destination port number corresponding to the atomization controller and the physical layer frame length limit, and then calculates the difference between the physical layer frame length limit and the length of the message header field and the fragment sequence number field to obtain the maximum effective payload length of a single frame. The calculation formula is as follows: ;

[0088] in, This represents the maximum effective payload length per frame. Due to physical layer frame length limitations, For fixed length of message header fields, The fragment sequence number field has a fixed length. In this embodiment, the physical layer frame length is limited to 1500 bytes, the header field has a fixed length of 14 bytes, and the fragment sequence number field has a fixed length of 2 bytes, resulting in a maximum payload length of 1484 bytes per frame.

[0089] After determining the maximum effective payload length per frame, the edge computing nodes calculate the total data length of the multi-source feature vectors, divide the total data length by the maximum effective payload length per frame, and round up to obtain the total number of frames in the target slice. The calculation formula is as follows: ;

[0090] in, The target is the total number of frames in the fragment. The total data length of the multi-source feature vectors. This is the round-up operator. Edge computing nodes perform equal-length segmentation of the multi-source feature vectors according to the total number of target fragment frames. The data length of each segment is consistent with the maximum effective payload length of a single frame. When the data length of the last segment is less than the maximum effective payload length of a single frame, a third preset value (0x00) is added to the end of the last segment, making the data length of the padded last segment equal to the maximum effective payload length of a single frame. All segments after segmentation are used as multi-frame feature fragment data. Through equal-length segmentation and end-padding, the length of all feature fragment data is completely consistent, conforming to the transmission specifications of the Industrial Ethernet protocol and avoiding transmission scheduling delays or data packet loss caused by inconsistent frame lengths.

[0091] Edge computing nodes add a header field containing the destination port number to the beginning of each frame of multi-frame feature fragment data and a fragment sequence number field to the end of each frame. After generating the interaction message, the interaction message is sent to the industrial Ethernet bus through the sending queue of the edge gateway in the order of the fragment sequence number field. At the same time, the edge computing node continuously listens to the industrial Ethernet bus. When it receives a response frame that carries a fragment sequence number field matching the fragment sequence number field, it extracts the status register value in the response frame as the execution feedback data.

[0092] refer to Figure 6 The edge computing node inputs the execution feedback data into a pre-defined state word parsing mapping table and outputs the updated valve opening and motor speed values ​​of the atomizing controller. The state word parsing mapping table predefines the mapping relationship between state register values ​​and engineering physical quantities. In this embodiment, the state register is a 32-bit unsigned integer; the high 16 bits correspond to the valve opening value, and the low 16 bits correspond to the motor speed value. The mapping table linearly maps the range of the 16-bit unsigned integer data (0x0000-0xFFFF) to the engineering value range of valve opening (0%-100%) and motor speed (0-3000 r / min). After parsing the execution feedback data according to the mapping table, the edge computing node superimposes the updated valve opening and motor speed values ​​onto the corresponding feed flow rate and atomization pressure dimensions in the multi-source feature vector, generating an updated feature vector. The updated feature vector contains the execution result of this control operation, providing feedback for the next control parameter calculation.

[0093] Edge computing nodes store the updated feature vectors into a circular buffer queue, and move the write pointer of the circular buffer queue forward by one step, the step size being consistent with the data length of a single feature vector. The circular buffer queue uses a fixed-length contiguous storage space, and both the write and read pointers are moved cyclically using modulo operations. The formula for calculating the pointer movement is as follows: ; ;

[0094] in, This is the write pointer to the circular buffer queue. This is the read pointer for the circular buffer queue. This represents the maximum depth of the circular buffer queue. This is the modulo operator. In this embodiment, the maximum depth of the circular buffer queue is 16, which corresponds to storing a maximum of 16 sets of updated feature vectors.

[0095] Edge computing nodes monitor the positional relationship between the write and read pointers in real time. When the difference between the write pointer and the read pointer of the circular buffer queue reaches a preset queue depth threshold, the next interaction message encapsulation operation is triggered, and the read pointer is moved to a position coinciding with the write pointer. The queue depth threshold is determined as follows: ;

[0096] in, For absolute value operations, In this embodiment, a preset queue depth threshold is used. When this condition is met, it indicates that the circular buffer queue has stored a sufficient amount of feedback data, which can provide complete feedback basis for the next control parameter calculation. At this time, the next sensor data acquisition, processing and message encapsulation operation is triggered, and the read pointer is moved to the position that coincides with the write pointer, and the processed historical data is cleared to avoid data overflow in the queue.

[0097] This embodiment achieves adaptive adjustment of the sliding window period according to the spray drying conditions by calculating the temperature change slope based on the least squares method and combining it with the lag time constant mapping model of piecewise linear interpolation. This allows the time range of data processing to match the heat and mass transfer process in the drying tower in real time. Standardized equal-length fragmentation and end padding ensure the frame structure consistency of interactive messages during transmission on the industrial Ethernet bus, reducing the probability of packet loss. Standardized parsing of execution feedback data is achieved through a status word parsing mapping table, and the threshold triggering mechanism of the circular buffer queue maintains the temporal continuity of the data interaction closed loop, ensuring the stable operation of the feedback link in the control process.

[0098] In another preferred embodiment, this embodiment optimizes the upsampling and reconstruction process of low-frequency sampled data based on the foregoing embodiments, reducing the computational load of edge computing nodes, and refines the data format unification process in the time alignment process to ensure the compatibility and consistency of multi-source data.

[0099] Specifically, refer to Figure 3 After the sliding window is defined, the edge computing nodes identify low-frequency and high-frequency sampling data within the basic sampling period, marking the timestamps of the low-frequency sampling data within the sliding window as the original interpolation nodes. The edge computing nodes count the total number of original interpolation nodes within the sliding window and compare this total number with a preset node count threshold of 5, corresponding to the construction of a 5th-order Lagrange interpolation polynomial. When the total number of original interpolation nodes exceeds the preset node count threshold, the edge computing nodes, in chronological order, extract the preset node count threshold of target interpolation nodes closest to the end time of the sliding window. The formula for extracting the timestamp set and corresponding value set of the target interpolation nodes is as follows: ; ;

[0100] in, The set of timestamps for the target interpolation nodes. The set of low-frequency sampled data values ​​corresponding to the target interpolation node. This represents the maximum index value of the original interpolation nodes within the sliding window. In this embodiment, a preset threshold for the number of nodes is used. When the total number of original interpolation nodes is less than or equal to a preset node number threshold, the edge computing node uses all original interpolation nodes as target interpolation nodes to participate in the subsequent interpolation polynomial construction. By limiting the number of nodes participating in the interpolation operation, the Runge phenomenon caused by high-order interpolation is avoided, while significantly reducing the computational load of the edge computing node, improving the real-time performance of the upsampling reconstruction process, and adapting to the limited computing resources of the edge computing node.

[0101] Edge computing nodes acquire the high-frequency sampling frequency corresponding to the high-frequency sampling data. Using the high-frequency sampling frequency as the reconstruction frequency, virtual interpolation nodes are uniformly inserted between two adjacent original interpolation nodes. In this embodiment, the high-frequency sampling frequency is 100Hz, corresponding to a sampling period of 10ms; the low-frequency sampling frequency is 10Hz, corresponding to a sampling period of 100ms; and the time interval between two adjacent original interpolation nodes is 100ms. Within this time interval, nine virtual interpolation nodes are uniformly inserted, ensuring that the time interval between two adjacent interpolation nodes is 10ms, perfectly matching the sampling interval of the high-frequency sampling data. The timestamps of the virtual interpolation nodes are generated in a linearly increasing manner to ensure that the time series of the interpolation nodes completely overlaps with the time series of the high-frequency sampling data, providing a unified time reference for subsequent time-series alignment processing.

[0102] For each target interpolation node, the edge computing node calculates the absolute value of the difference between the timestamps of all other target interpolation nodes and the timestamp of the current target interpolation node. These absolute differences are then multiplied in ascending order to obtain the product of the denominator and numerator. For the ... The product of the denominators of the interpolation basis functions for each target interpolation node is: ;

[0103] in, For the first The result of the chain of denominators for each target interpolation node. For the first The timestamps of the target interpolation nodes To exclude the first The timestamps of all target interpolation nodes other than the target interpolation node. The product of the numerators of the interpolation basis functions is: ;

[0104] in, For the first The numerator chain product results corresponding to each target interpolation node The input is the timestamp coordinates of the interpolation nodes.

[0105] The edge computing nodes perform a weighted summation of the low-frequency sampled data values, the product of the denominator and the product of the numerator for each target interpolation node, generating the coefficients of the Lagrange interpolation polynomial. The completed Lagrange interpolation polynomial is as follows: ;

[0106] in, To construct the complete Lagrange interpolation polynomial, For the first The low-frequency sampled data values ​​corresponding to each target interpolation node are calculated. The edge computing node sequentially substitutes the timestamp coordinates of all virtual interpolation nodes into this polynomial to calculate the low-frequency sampled data value corresponding to each virtual interpolation node. The data set containing both the original interpolation nodes and the virtual interpolation nodes is then used as the upsampled data. Through this optimization, the computational complexity of the upsampling reconstruction process is limited to a fixed range and does not increase with the length of the sliding window, ensuring the real-time performance and stability of the upsampling process.

[0107] Table 3 Lagrange Interpolation Upsampling Parameter Configuration Table

[0108] Table 3 clarifies the core configuration parameters of the Lagrange interpolation upsampling process in this embodiment, providing a standardized execution basis for upsampling reconstruction of low-frequency sampled data. While ensuring interpolation accuracy, it also achieves effective control of the computing load of edge computing nodes, ensuring that the upsampling process can meet the timing requirements of real-time processing.

[0109] After upsampling reconstruction, the edge computing nodes time-align the high-frequency sampled data and the upsampled data at the end of the sliding window to generate multi-source feature vectors. The edge computing nodes extract the timestamp corresponding to the end of the sliding window as a unified alignment timestamp, and extract real-time high-frequency data points from the high-frequency sampled data whose timestamps are equal to the unified alignment timestamp, as well as real-time upsampled data points from the upsampled data whose timestamps are equal to the unified alignment timestamp.

[0110] Edge computing nodes perform data format unification processing on real-time high-frequency data points and real-time upsampled data points. First, the first data bit width of the real-time high-frequency data points and the second data bit width of the real-time upsampled data points are extracted. The larger of the first and second data bit widths is determined as the target data bit width. The formula for calculating the target data bit width is: ;

[0111] in, For the target data bit width, The first data bit width for real-time high-frequency data points. The second data bit width is the real-time upsampled data point. In this embodiment, the first data bit width of the real-time high-frequency data point is 12 bits, the second data bit width of the real-time upsampled data point is 16 bits, and the calculated target data bit width is 16 bits.

[0112] After the target data bit width is determined, the edge computing node detects the actual data bit width of the real-time high-frequency data points and the real-time upsampled data points. When the data bit width of the real-time high-frequency data point is detected to be less than the target data bit width, a first preset value (0) is filled before the most significant bit of the real-time high-frequency data point. When the data bit width of the real-time upsampled data point is detected to be less than the target data bit width, a second preset value (0) is filled before the most significant bit of the real-time upsampled data point. Through high-bit padding, the data bit width of all data points is unified to the target data bit width, and the padding operation does not change the numerical value of the original data, ensuring data accuracy.

[0113] After the bit width is uniformly processed, the edge computing node converts both the padded real-time high-frequency data points and the padded real-time upsampled data points into data streams in a preset floating-point format. The preset floating-point format adopts the IEEE 754 single-precision floating-point format, with each data point occupying 4 bytes of fixed storage space. Through floating-point format conversion, all sensor data are stored and transmitted in a unified numerical format, avoiding parsing errors caused by data type differences, and providing standardized numerical input for subsequent control logic calculations.

[0114] The edge computing node concatenates the real-time high-frequency data points (after standardization) with the real-time upsampled data points according to a preset order of sensor physical addresses to generate a fixed-length multi-source feature vector main data segment. The expression for the main data segment is as follows: ;

[0115] in, This is the main data segment of the multi-source feature vector. For alignment time The corresponding feed flow rate data, For alignment time The corresponding atomization pressure data, For alignment time

[0116] The corresponding hot air temperature data, For alignment time The corresponding exhaust temperature data. The main data segment has a fixed length of 16 bytes and contains 4 single-precision floating-point numbers.

[0117] Edge computing nodes add a unified alignment timestamp and a target lag time constant to the header of the main data segment of the multi-source feature vector. The unified alignment timestamp uses a 64-bit unsigned integer format, occupying 8 bytes of storage space, while the target lag time constant uses a single-precision floating-point format, occupying 4 bytes of storage space. A cyclic redundancy check (CRC) code is added to the tail of the main data segment. The CRC code uses a 32-bit unsigned integer format, occupying 4 bytes of storage space. The complete multi-source feature vector has a total length of 32 bytes and a fixed frame structure, providing a unified format benchmark for subsequent message encapsulation and transmission.

[0118] After the multi-source feature vectors are constructed, the edge computing node encapsulates them into an interactive message in a preset format, sends it to the atomization controller via the edge gateway, and receives the execution feedback data returned by the atomization controller, completing the data interaction loop. The formula for calculating the cyclic redundancy check code is: ;

[0119] in, This refers to all bytes of data in the multi-source feature vector except for the CRC field. For generating polynomials Cyclic Redundancy Check (CRC) calculations are used. By adding CRC codes, any single-bit errors and multi-bit burst errors that occur during data transmission can be detected, ensuring the integrity of data transmission.

[0120] This embodiment avoids the Runge phenomenon caused by high-order interpolation by limiting the number of nodes for Lagrange interpolation, reduces the computational load on edge computing nodes, and ensures the real-time performance of the upsampling reconstruction process. Through standardized data bit width matching, high-bit padding, and floating-point format conversion, it achieves format unification of multi-source data, ensuring the consistency and compatibility of feature vectors after time alignment. Through the design of a fixed-length multi-source feature vector structure, it simplifies the subsequent message encapsulation and transmission process, further improving the stability and reliability of data interaction.

Claims

1. A real-time data interaction processing method for a spray drying process, characterized in that, include: Acquire data from the feed flow sensor, atomization pressure sensor, hot air temperature sensor, and exhaust temperature sensor of the spray drying tower, and determine the basic sampling period for the feed flow sensor data, the atomization pressure sensor data, the hot air temperature sensor data, and the exhaust temperature sensor data; Based on the material heat and mass transfer lag time constant in the spray drying process, a sliding window is established with the change in the exhaust temperature sensor data as the reference. The material heat and mass transfer lag time constant is the time interval from when the material enters the chamber from the feed inlet of the spray drying tower, is atomized and dispersed, exchanges heat and mass with hot air, and finally generates a detectable temperature change at the exhaust detection point. This constant corresponds to the physical lag characteristics between the changes in feed parameters and the exhaust temperature response during the spray drying process. Within the sliding window, a polynomial fitting method based on Lagrange interpolation is used to upsample and reconstruct the low-frequency sampling data in the basic sampling period to obtain upsampled data; The high-frequency sampled data in the basic sampling period and the upsampled data are time-aligned with each other at the end of the sliding window to generate a multi-source feature vector. The multi-source feature vectors are encapsulated into an interactive message in a preset format. The interactive message is sent to the atomization controller via the edge gateway, and the execution feedback data returned by the atomization controller is received to complete a data interaction loop. The step of establishing a sliding window based on the changes in the exhaust temperature sensor data, according to the material heat and mass transfer lag time constant in the spray drying process, includes: extracting multiple consecutive historical data points of the exhaust temperature sensor data before the current moment. Calculate the slope of temperature change corresponding to the consecutive historical data points; The temperature change slope is input into a preset lag time constant mapping model, and the target lag time constant corresponding to the current moment is output. The sliding window is defined between the current time as the end point of the time axis and the time when the target lag time constant is subtracted from the current time as the start point of the time axis. When the next data point of the exhaust temperature sensor is detected, the time axis start point and time axis end point of the sliding window are shifted according to the target lag time constant. The lag time constant mapping model is a set of piecewise linear interpolation functions. This set is pre-constructed through full-condition calibration experiments of the spray drying tower and includes multiple sets of slope interval thresholds and multiple sets of constant interval values. Each set of slope intervals corresponds to a set of linear interpolation parameters. The slope of the temperature change is compared one by one with the multiple sets of slope interval thresholds to determine the target slope interval in which the slope of the temperature change is located. Extract the slope values ​​of the left and right endpoints corresponding to the target slope interval, and extract the constant values ​​of the left and right endpoints corresponding to the target slope interval; A local linear equation is constructed based on the slope values ​​of the left and right endpoints and the constant values ​​of the left and right endpoints. The slope of the temperature change is substituted into the local linear equation to calculate the target lag time constant.

2. The real-time data interaction processing method in the spray drying process according to claim 1, characterized in that, Within the sliding window, a polynomial fitting method based on Lagrange interpolation is used to upsample and reconstruct the low-frequency sampling data in the basic sampling period to obtain upsampled data, including: identifying the low-frequency sampling data and high-frequency sampling data in the basic sampling period, and marking the timestamp of the low-frequency sampling data within the sliding window as the original interpolation node; Obtain the high-frequency sampling frequency corresponding to the high-frequency sampling data, and use the high-frequency sampling frequency as the reconstruction frequency to uniformly insert virtual interpolation nodes between two adjacent original interpolation nodes; Based on the low-frequency sampled data values ​​corresponding to all the original interpolation nodes within the sliding window, a Lagrange interpolation polynomial is constructed. Substitute the timestamp coordinates of the virtual interpolation node into the Lagrange interpolation polynomial to calculate the low-frequency sampled data value corresponding to the virtual interpolation node, and use the data set containing the original interpolation node and the virtual interpolation node as the upsampled data.

3. The real-time data interaction processing method in the spray drying process according to claim 2, characterized in that, The step of aligning the high-frequency sampling data in the basic sampling period with the upsampled data at the end of the sliding window to generate a multi-source feature vector includes: extracting the timestamp corresponding to the end of the sliding window as a unified alignment timestamp. Extract real-time high-frequency data points from the high-frequency sampled data whose timestamps are equal to the unified alignment timestamp, and extract real-time upsampled data points from the upsampled data whose timestamps are equal to the unified alignment timestamp; The real-time high-frequency data points and the real-time upsampled data points are processed to unify their data formats. The processed real-time high-frequency data points and the real-time upsampled data points are then concatenated according to a preset arrangement order of the sensor physical addresses to generate a fixed-length multi-source feature vector. The unified alignment timestamp and the target lag time constant are added to the beginning of the multi-source feature vector, and a cyclic redundancy check code is added to the end of the multi-source feature vector.

4. The real-time data interaction processing method in the spray drying process according to claim 3, characterized in that, The step of encapsulating the multi-source feature vectors into an interactive message of a preset format and sending the interactive message to the atomization controller via the edge gateway includes: parsing the communication protocol stack of the edge gateway and determining the destination port number and physical layer frame length limit corresponding to the atomization controller; The multi-source feature vector is segmented according to the physical layer frame length limit to obtain multi-frame feature segmentation data. A message header field containing the destination port number is added to the header of each frame of the multi-frame feature fragment data, and a fragment sequence number field is added to the tail of each frame to generate the interaction message; The interaction messages are sent to the industrial Ethernet bus through the sending queue of the edge gateway, in the order of the fragment sequence number field. The system monitors the industrial Ethernet bus. When it receives a response frame that matches the fragment sequence number field, it extracts the status register value from the response frame as the execution feedback data.

5. The real-time data interaction processing method in the spray drying process according to claim 4, characterized in that, The step of receiving the execution feedback data returned by the atomization controller to complete a data interaction loop includes: inputting the execution feedback data into a preset status word parsing mapping table, and outputting the valve opening update value and motor speed update value of the atomization controller; The updated valve opening value and the updated motor speed value are superimposed on the corresponding feed flow rate dimension and atomization pressure dimension in the multi-source feature vector to generate the updated feature vector. The updated feature vector is stored in a circular buffer queue, and the write pointer of the circular buffer queue is moved forward by one step. When the difference between the write pointer and the read pointer of the circular buffer queue is detected to reach a preset queue depth threshold, the encapsulation operation of the next interaction message is triggered, and the read pointer is moved to a position that coincides with the write pointer.

6. The real-time data interaction processing method in the spray drying process according to claim 5, characterized in that, The step of constructing a Lagrange interpolation polynomial based on the low-frequency sampled data values ​​corresponding to all the original interpolation nodes within the sliding window includes: counting the total number of the original interpolation nodes within the sliding window; When the total number is greater than the preset node number threshold, the preset node number threshold target interpolation nodes that are closest to the end time of the sliding window are extracted from the original interpolation nodes in chronological order. For each target interpolation node, calculate the absolute value of the difference between the timestamp of other target interpolation nodes (excluding the current target interpolation node) and the timestamp of the current target interpolation node; The absolute values ​​of the differences are multiplied in ascending order to obtain the product of the denominator and the product of the numerator. The low-frequency sampled data values ​​corresponding to each target interpolation node, the product of the denominator and the product of the numerator are weighted and summed to generate the coefficients of the Lagrange interpolation polynomial.

7. The real-time data interaction processing method in a spray drying process according to claim 6, characterized in that, The process of unifying the data format of the real-time high-frequency data points and the real-time upsampling data points includes: extracting the first data bit width of the real-time high-frequency data points and the second data bit width of the real-time upsampling data points, and determining the maximum bit width between the first data bit width and the second data bit width as the target data bit width. When it is detected that the data bit width of the real-time high-frequency data point is less than the target data bit width, a first preset value is filled before the most significant bit of the real-time high-frequency data point; when it is detected that the data bit width of the real-time upsampled data point is less than the target data bit width, a second preset value is filled before the most significant bit of the real-time upsampled data point. Both the filled real-time high-frequency data points and the filled real-time upsampled data points are converted into data streams in a preset floating-point format.

8. The real-time data interaction processing method in the spray drying process according to claim 7, characterized in that, The step of segmenting the multi-source feature vector according to the physical layer frame length limit to obtain multi-frame feature segmentation data includes: calculating the difference between the physical layer frame length limit and the length of the message header field and the length of the segmentation sequence number field to obtain the maximum effective payload length of a single frame; Divide the total data length of the multi-source feature vector by the maximum effective payload length of a single frame, and round up to obtain the total number of target fragment frames. The multi-source feature vector is divided into equal-length segments according to the total number of target segments. When the data length of the last segment is less than the maximum effective payload length of a single frame, a third preset value is filled at the end of the last segment so that the data length of the last segment after filling is equal to the maximum effective payload length of a single frame. All segments after segmentation are used as the multi-frame feature segment data.

Images