A Real-Time Online Viscosity Monitoring Method and System for Millet Powder Processing Based on Edge Computing

Abstract

This invention relates to the field of online monitoring technology for food processing, specifically a method and system for real-time online viscosity monitoring of millet flour processing based on edge computing. The method includes: simultaneously acquiring screw torque timing signals, an infrared thermogram of the discharge port, and material moisture content at the millet flour extrusion and maturation outlet; using edge computing nodes to perform temperature field gradient analysis on the infrared thermogram to extract thermal entropy feature values ​​characterizing the uniformity of material gelatinization; and constructing a rheological correction factor based on the thermal entropy feature values ​​and the material moisture content. This invention, by simultaneously acquiring screw torque, an infrared thermogram of the discharge port, and material moisture content, and extracting thermal entropy feature values ​​from the infrared thermogram through temperature field gradient analysis, can quantitatively characterize the uniformity of material gelatinization during millet flour extrusion, overcoming the shortcomings of traditional methods that rely solely on a single torque or pressure signal and cannot perceive spatial differences in the gelatinization state.

Description

Technical Field

[0001] This invention relates to the field of online monitoring technology for food processing, specifically to a method and system for real-time online viscosity monitoring of millet flour processing based on edge computing. Background Technology

[0002] Extrusion and curing of millet flour is a crucial process in the deep processing of grains. Inside the extruder, the material undergoes high temperature, high pressure, and high shear, resulting in complex physicochemical changes such as starch gelatinization, degradation, and cross-linking, ultimately leading to extrusion molding. During this process, the real-time viscosity of the material is a core process parameter characterizing the degree of starch gelatinization, extrusion flow state, and final product quality. High viscosity leads to poor material flowability and an increased risk of die blockage, while low viscosity indicates excessive starch degradation and product collapse due to puffing. Therefore, accurate online monitoring of the material viscosity during extrusion is a key prerequisite for achieving stable process control and consistent product quality.

[0003] Traditional viscosity monitoring methods in millet flour extrusion processing mainly rely on three approaches: First, offline sampling and testing, which involves taking samples from the extrudate at regular intervals and then measuring them in the laboratory using a rapid viscosity analyzer or a Brabender viscometer. While this method offers high accuracy, the lag time is typically on the order of several minutes to tens of minutes, failing to reflect the instantaneous state of the material within the extruder. Second, empirical judgment based on a single parameter, such as screw torque or die pressure, where operators adjust equipment operating parameters based on experience using torque or pressure readings. This method is highly subjective, inconsistent, and the torque-viscosity correspondence is easily misjudged due to interference from multiple factors, including material moisture content and gelatinization uniformity. Third, soft viscosity measurement methods based on indirect sensing signals such as vibration or acoustic emission. These methods attempt to establish a statistical mapping relationship between vibration spectrum or acoustic characteristics and viscosity. However, the complex mechanical vibration background generated by the extruder's operation and the nonlinear changes in vibration transmission paths under no-load and loaded conditions make signal-noise separation difficult, and the calibration model's generalization ability for different material types and screw wear conditions is severely insufficient. Summary of the Invention

[0004] To achieve the above objectives, the present invention provides the following technical solution: a method for real-time online viscosity monitoring in millet flour processing based on edge computing, comprising: Simultaneously collect screw torque timing signals, discharge port infrared thermal images, and material moisture content at the millet flour extrusion and cooking outlet; The infrared thermal image is analyzed by edge computing nodes to extract thermal entropy feature values ​​that characterize the uniformity of material gelatinization. Based on the thermal entropy characteristic value and the moisture content of the material, a rheological correction factor is constructed. The rheological correction factor is used to perform non-Newtonian fluid shear compensation calculation on the screw torque time series signal to obtain an equivalent shear stress sequence. The instantaneous apparent viscosity of millet powder is calculated using the capillary rheological inversion equation based on the preset extrusion die geometry parameters and the equivalent shear stress sequence. The instantaneous apparent viscosity is compared with a preset standard viscosity threshold range. If the viscosity exceeds the range, an adjustment command is generated and fed back to the extruder control system to achieve closed-loop real-time monitoring of viscosity.

[0005] Preferably, at the millet flour extrusion and cooking outlet, the screw torque timing signal, the discharge port infrared thermal image, and the material moisture content are simultaneously collected, including: A strain gauge torque sensor installed on the screw drive shaft of the extruder continuously acquires screw torque timing signals at a first preset sampling frequency and records a first timestamp sequence. An infrared thermal imager mounted above the extrusion die outlet is used to continuously capture infrared thermal images of the material discharge flow at a second preset sampling frequency, and a second timestamp sequence is recorded. The moisture content of the material at the moment of extrusion is obtained online by a near-infrared moisture content measuring probe set at the discharge port, and the third timestamp sequence is recorded. In the edge computing node, the second and third timestamp sequences are interpolated and resampled using the first timestamp sequence as a reference, so that the screw torque timing signal, infrared thermal image and material moisture content are aligned on the time axis, and synchronous acquisition is achieved.

[0006] Preferably, edge computing nodes are used to perform temperature field gradient analysis on the infrared thermal image to extract thermal entropy feature values ​​characterizing the uniformity of material gelatinization, including: The edge computing node acquires the timestamp-aligned infrared thermal image, calculates the local temperature gradient amplitude pixel by pixel for a single frame of infrared thermal image, and obtains a gradient amplitude map of the same size as the original infrared thermal image. The gradient magnitude map is divided into a preset number of equal-area sub-regions. The pixel distribution of gradient magnitude in each sub-region is statistically analyzed, and a gradient magnitude histogram of each sub-region is constructed. The gradient magnitude histogram of each sub-region is normalized to obtain the gradient probability density distribution of the sub-region, and the local thermal entropy of the temperature field of the sub-region is calculated based on the gradient probability density distribution. Arrange the local thermal entropy values ​​of the temperature field of all sub-regions according to their spatial location to form the thermal entropy distribution matrix corresponding to the infrared thermal image frame. The coefficient of variation is calculated pixel by pixel for multiple thermal entropy distribution matrices corresponding to multiple consecutive frames of infrared thermal images within a time window to obtain the thermal entropy coefficient of variation matrix; wherein, each element in the coefficient of variation matrix represents the temporal stability of the temperature field order at the corresponding spatial location. The thermal entropy variation coefficient matrix is ​​weighted and pooled to obtain a scalar value as the thermal entropy characteristic value representing the uniformity of material gelatinization.

[0007] Preferably, a rheological correction factor is constructed based on the thermal entropy characteristic value and the material moisture content, including: Obtain the timestamp-aligned material moisture content sequence and the thermal entropy feature value sequence, and perform a sliding window average on the material moisture content sequence to obtain the material moisture content trend value; The thermal entropy feature value sequence is matched one-to-one with the material moisture content trend value according to the timestamp to form a time-varying two-dimensional state vector; From a preset gelatinization state reference library, retrieve the reference state point with the smallest Euclidean distance to the time-varying two-dimensional state vector; wherein, the gelatinization state reference library contains a reference record consisting of multiple reference state points and the standard rheological coefficients corresponding to each reference state point; Extract the standard rheological coefficient corresponding to the reference state point with the smallest distance as the reference rheological coefficient; The deviation rate between the thermal entropy characteristic value and its ideal homogeneous state thermal entropy value is calculated, and the deviation rate between the material moisture content and its nominal moisture content is calculated. The two deviation rates are weighted and fused to obtain the time-varying dynamic bias. The rheological correction factor is obtained by superimposing the reference rheological coefficient with the time-varying dynamic bias.

[0008] Preferably, the two bias rates are weighted and fused to obtain the time-varying dynamic bias, including: Based on the rate of change of the thermal entropy feature value sequence within the sliding time window, the first weighting coefficient of the thermal entropy deviation rate is dynamically determined. The second weighting coefficient of the moisture content deviation rate is dynamically determined based on the rate of change of the material moisture content sequence within the sliding time window; wherein the sum of the first weighting coefficient and the second weighting coefficient is 1. The time-varying dynamic bias is obtained by weighting and summing the deviation rates of the thermal entropy characteristic value and the material moisture content using the first weighting coefficient and the second weighting coefficient.

[0009] Preferably, the screw torque time-series signal is subjected to non-Newtonian fluid shear compensation calculation using the rheological correction factor to obtain an equivalent shear stress sequence, including: Obtain the timestamp-aligned screw torque timing signal and the rheological correction factor sequence; The screw torque timing signal is subjected to baseline drift removal and power frequency notch filtering to obtain a preprocessed torque signal; The preprocessed torque signal is substituted into the preset screw torque-shear stress conversion equation to obtain the nominal shear stress sequence; the screw torque-shear stress conversion equation is pre-established based on the extruder screw geometric parameters and the power-law fluid constitutive equation; Obtain the rheological correction factor value that matches the timestamp corresponding to each nominal shear stress value in the rheological correction factor sequence, and use the rheological correction factor value as the exponential term to perform power exponential compensation operation on the nominal shear stress sequence point by point to obtain the equivalent shear stress sequence.

[0010] Preferably, the screw torque-shear stress conversion equation is pre-established based on the extruder screw geometry parameters and the power-law fluid constitutive equation, including: Obtain the helix angle, groove depth, and groove width of the extruder screw; Determine the non-Newtonian exponents based on the constitutive equation of power-law fluids; Substituting the helix angle, groove depth, groove width, and non-Newtonian exponent into the analytical expression for screw extrusion shear stress, we obtain the screw torque-shear stress conversion equation with torque as input and nominal shear stress as output.

[0011] Preferably, the instantaneous apparent viscosity of millet flour is calculated using capillary rheological inversion equations based on preset extrusion die geometry parameters and the equivalent shear stress sequence, including: Obtain the die hole diameter and die hole length of the extrusion die head as the geometric parameters of the extrusion die head; Multiply each equivalent shear stress value in the equivalent shear stress sequence by the diameter of the die hole, and divide by four times the length of the die hole to obtain the wall shear stress sequence. The extrusion volume of the extruder per unit time is obtained, and the extrusion volume is divided by the cross-sectional area of ​​the die orifice to obtain the apparent shear rate. Divide each wall shear stress value in the wall shear stress sequence by the apparent shear rate to obtain the initial apparent viscosity sequence. Divide the difference in equivalent shear stress between adjacent sampling points in the equivalent shear stress sequence by the sampling time interval to obtain the shear stress change rate sequence. Based on the absolute value of the shear stress change rate sequence, time-varying inertia compensation is performed on the initial apparent viscosity sequence to obtain the instantaneous apparent viscosity sequence.

[0012] Preferably, the instantaneous apparent viscosity is compared with a preset standard viscosity threshold range. If it exceeds the range, a control command is generated and fed back to the extruder control system to achieve closed-loop real-time viscosity monitoring, including: The instantaneous apparent viscosity sequence is obtained, and the instantaneous apparent viscosity value of each sampling point in the sequence is compared with the preset standard viscosity threshold range point by point. In response to the instantaneous apparent viscosity value at the current sampling point exceeding the standard viscosity threshold range, a timer is triggered to start accumulating the abnormal duration; When the duration of the abnormality reaches a preset threshold for continuous abnormality duration, the average of all instantaneous apparent viscosity values ​​that exceed the interval within the duration of the abnormality is calculated as the deviation viscosity value. The difference between the deviation viscosity value and the center value of the standard viscosity threshold range is calculated to obtain the viscosity deviation amount; The direction of adjustment is determined based on the sign of the viscosity deviation; wherein, the direction of adjustment includes increasing or decreasing the viscosity. The adjustment step size level is determined based on the preset grade range in which the absolute value of the viscosity deviation falls; The control direction and the control step size are combined into a control command code, which is fed back to the feed rate control unit and heating power control unit of the extruder control system to adjust the extruder operating parameters in a closed loop, so that the subsequent instantaneous apparent viscosity returns to the standard viscosity threshold range.

[0013] The edge computing-based online viscosity real-time monitoring system for millet flour processing is applicable to the aforementioned edge computing-based online viscosity real-time monitoring method for millet flour processing, including: The multimodal acquisition module is used to simultaneously acquire screw torque timing signals, discharge port infrared thermal images, and material moisture content at the millet flour extrusion and cooking outlet. The thermal entropy feature extraction module is used to perform temperature field gradient analysis on the infrared thermal image using edge computing nodes, and extract thermal entropy feature values ​​that characterize the uniformity of material gelatinization. The rheological stress compensation module is used to construct a rheological correction factor based on the thermal entropy characteristic value and the material moisture content, and to perform non-Newtonian fluid shear compensation calculation on the screw torque time series signal using the rheological correction factor to obtain an equivalent shear stress sequence. The viscosity inversion calculation module is used to calculate the instantaneous apparent viscosity of millet powder based on the preset extrusion die geometry parameters and the equivalent shear stress sequence through the capillary rheological inversion equation. The closed-loop control command module is used to compare the instantaneous apparent viscosity with a preset standard viscosity threshold range. If the viscosity exceeds the range, a control command is generated and fed back to the extruder control system to achieve closed-loop real-time monitoring of viscosity.

[0014] Compared with the prior art, the beneficial effects of the present invention are: This invention simultaneously acquires screw torque, outlet infrared thermal images, and material moisture content, and extracts thermal entropy feature values ​​from the infrared thermal images through temperature field gradient analysis. This quantitatively characterizes the uniformity of material gelatinization during millet flour extrusion, overcoming the shortcomings of traditional methods that rely solely on a single torque or pressure signal and cannot perceive spatial differences in gelatinization states. Furthermore, by jointly constructing a rheological correction factor with the thermal entropy feature value and material moisture content, and performing non-Newtonian fluid shear compensation calculations on the screw torque signal, this invention effectively eliminates the cross-interference between gelatinization inhomogeneity and moisture content fluctuations on the torque-viscosity relationship, making the calculation of equivalent shear stress closer to the actual rheological state of the material. Moreover, the above compensation process is completed entirely in a closed loop within the edge computing node, without the need for no-load calibration or shutdown sampling of the extruder. This ensures real-time performance while avoiding model inaccuracies caused by equipment wear and changes in material type in traditional vibration methods. This invention directly calculates the instantaneous apparent viscosity of a material by substituting the equivalent shear stress sequence and the geometric parameters of the extrusion die into the capillary rheological inversion equation. The calculation path conforms to the basic principles of non-Newtonian fluid rheology and has stronger process adaptability compared to empirical fitting or statistical mapping methods. Furthermore, by judging the duration of continuous anomalies between the instantaneous apparent viscosity and the standard threshold range, a graded control command is generated and fed back to the extruder control system only after the anomaly has persisted for a set duration. This avoids erroneous adjustments caused by instantaneous random fluctuations and automatically matches the control step size according to the deviation amplitude, achieving precise closed-loop adjustment of the extruder feeding rate and heating power, significantly improving the batch consistency of millet flour extrusion and cooking products. Attached Figure Description

[0015] Figure 1 This is a schematic flowchart of the overall method in one embodiment of the present invention; Figure 2 This is a schematic diagram of the overall system architecture in one embodiment of the present invention.

[0016] In the diagram: 1. Multimodal acquisition module; 2. Thermal entropy feature extraction module; 3. Rheological stress compensation module; 4. Viscosity inversion calculation module; 5. Closed-loop control command module. Detailed Implementation

[0017] 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 embodiments of the present invention, and not all embodiments. 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.

[0018] Example 1, please refer to Figure 1This invention provides a technical solution: a method for real-time online viscosity monitoring in millet flour processing based on edge computing, comprising: S1. Simultaneously collect screw torque timing signal, discharge port infrared thermal image and material moisture content at the millet flour extrusion and cooking outlet; S2. Use edge computing nodes to analyze the temperature field gradient of infrared thermal images and extract thermal entropy feature values ​​that characterize the uniformity of material gelatinization. S3. Based on the thermal entropy characteristic value and the material moisture content, a rheological correction factor is constructed. The rheological correction factor is used to perform non-Newtonian fluid shear compensation calculation on the screw torque time series signal to obtain the equivalent shear stress sequence. S4. Based on the preset geometric parameters of the extrusion die and the equivalent shear stress sequence, the instantaneous apparent viscosity of millet powder is calculated by the capillary rheological inversion equation. S5. The instantaneous apparent viscosity is compared with the preset standard viscosity threshold range. If it exceeds the range, an adjustment command is generated and fed back to the extruder control system to achieve closed-loop real-time monitoring of viscosity.

[0019] It should be noted that this solution completes the synchronous alignment, thermodynamic feature extraction, rheological compensation calculation, and closed-loop control decision-making of multi-source heterogeneous sensor data locally through edge computing nodes. This avoids the network latency and bandwidth pressure caused by uploading raw data to the cloud for processing, and enables rapid online monitoring and real-time control of material viscosity during the extrusion and maturation process of millet powder. The edge computing node hardware platform adopts an industrial embedded computer based on an ARM Cortex-A72 quad-core processor, running a real-time Linux operating system. Sensor data acquisition and processing tasks are scheduled in parallel through multi-threading.

[0020] In an optional embodiment, the screw torque timing signal, the discharge port infrared thermal image, and the material moisture content are simultaneously collected at the millet flour extrusion and cooking outlet, including: A strain gauge torque sensor installed on the screw drive shaft of the extruder continuously acquires screw torque timing signals at a first preset sampling frequency and records a first timestamp sequence. It should be noted that the strain gauge torque sensor measures torque by attaching four resistance strain gauges to the surface of the elastic element on the screw drive shaft and forming a full-bridge Wheatstone bridge. When the drive bearing experiences shear strain due to torque, the resistance of the strain gauges changes proportionally to the strain. The bridge outputs a differential voltage signal, which is amplified by an instrumentation amplifier, filtered by an anti-aliasing low-pass filter (cutoff frequency 100 Hz), and then sampled at equal intervals by a 24-bit analog-to-digital converter at a first preset sampling frequency. Simultaneously, edge computing nodes mark each sampling point with a UTC absolute timestamp via the IEEE 1588 PTP protocol, forming a first timestamp sequence. For example, if the first preset sampling frequency is set to 200 Hz, the sampling interval is 5 ms, generating 200 torque sample values ​​within 1 second, with timestamp accuracy reaching the microsecond level, such as "10:00:00.000000, torque value 52.1 N·m" or "10:00:00.005000, torque value 52.3 N·m". N·m” etc., continuously record the screw torque timing signal and the first timestamp sequence; An infrared thermal imager mounted above the extrusion die outlet is used to continuously capture infrared thermal images of the material discharge flow at a second preset sampling frequency, and a second timestamp sequence is recorded. It should be noted that the infrared thermal imager operates in the long-wave infrared band of 8-14μm. It outputs a two-dimensional temperature matrix (i.e., an infrared thermal image) by detecting the infrared radiance of the material surface and performing built-in non-uniformity correction, ambient temperature compensation, and temperature calibration, resulting in a temperature value corresponding to each pixel. The infrared thermal imager samples frames at a second preset sampling frequency independent of the torque sensor. Edge computing nodes timestamp each frame, forming a second timestamp sequence. The emissivity of the millet powder material is set to 0.94, and the non-uniformity correction uses a two-point correction method, automatically updating the correction parameters every 8 hours using a built-in blackbody baffle. For example, the infrared thermal imager has a resolution of 640×480 pixels, and the second preset sampling frequency is set to 30 Hz, meaning it captures 30 thermal images per second. The second timestamp sequence includes sequences such as "10:00:00.033, frame 001" and "10:00:00.066, frame 002," where each frame is a 640×480 floating-point temperature matrix with a temperature resolution of 0.05℃. The moisture content of the material at the moment of extrusion is obtained online by a near-infrared moisture content measuring probe set at the discharge port, and the third timestamp sequence is recorded. It should be noted that the near-infrared moisture content measurement probe operates based on the principle of frequency doubling and combination absorption of OH bonds in water molecules at wavelengths of 1450 nm and 1940 nm. The probe's built-in halogen tungsten lamp emits near-infrared light at a reference wavelength (1300 nm) and a measurement wavelength (1450 nm) through a filter wheel, illuminating the material surface. The diffuse reflected light is received by a PbS detector. The moisture content of the material is calculated in real time by combining the absorbance ratio of the measurement wavelength to the reference wavelength with a pre-established partial least squares regression model. The partial least squares regression model uses 15 principal components, with a calibration set determination coefficient R² = 0.98 and a root mean square error RMSE = 0.25%. Edge computing nodes mark each moisture content sample with a timestamp, forming a third timestamp sequence. For example, the sampling frequency of the near-infrared moisture content measurement probe is set to 10 Hz, meaning that every 100 Hz... The system outputs a moisture content value per millisecond; a third timestamp sequence, such as "10:00:00.050, moisture content 31.2%", "10:00:00.150, moisture content 31.5%", etc., with a measurement accuracy of ±0.3%. In the edge computing node, the second and third timestamp sequences are interpolated and resampled using the first timestamp sequence as a reference, so that the screw torque timing signal, infrared thermal image and material moisture content are aligned on the time axis, and synchronous acquisition is achieved. It should be noted that the three sensors have independent sampling frequencies and sampling times, resulting in asynchronous data flow on the time axis. The edge computing node uses the first timestamp sequence as the reference time axis, performs temporal cubic spline interpolation pixel-by-pixel on the infrared thermal image sequence corresponding to the second timestamp sequence, and performs one-dimensional cubic spline interpolation on the moisture content sequence corresponding to the third timestamp sequence. At each reference timestamp, resampling is performed to generate the corresponding interpolated infrared thermal image and interpolated moisture content value, thus achieving precise alignment of the three heterogeneous data sets on a unified time axis. For cases where the reference timestamp exceeds the beginning or end boundaries of the source data range, nearest neighbor extrapolation is used. A maximum allowable time is set. The time stamp deviation threshold is 1.5 times the sampling period. If a sensor continuously loses frames, causing the timestamp to jump beyond this threshold, the data in that interval is marked as invalid and a sensor status alarm is triggered. For example, if the reference timestamp is 10:00:00.100, the moisture content sequence is 31.2% at 10:00:00.050 and 31.5% at 10:00:00.150. After cubic spline interpolation, the moisture content at that moment is 31.35%. Similarly, for infrared thermal images, each pixel is interpolated in the time domain between two adjacent frames to generate a complete interpolated frame corresponding to the time 10:00:00.100. If the timestamp of a torque signal is missing for more than 7.5 ms, the corresponding synchronization data is marked as invalid and will not be included in subsequent viscosity calculations.

[0021] In an optional embodiment, edge computing nodes are used to perform temperature field gradient analysis on the infrared thermal image to extract thermal entropy feature values ​​characterizing the uniformity of material gelatinization, including: Edge computing nodes acquire timestamp-aligned infrared thermal images, calculate local temperature gradient magnitudes pixel by pixel for a single frame of infrared thermal image, and obtain a gradient magnitude map of the same size as the original infrared thermal image. It should be noted that the infrared thermal image T(x,y) is convolved with the 3×3 Sobel gradient operator in both the horizontal and vertical directions to obtain the horizontal gradient Gx and the vertical gradient Gy. The local temperature gradient magnitude at each pixel is the Euclidean norm of Gx and Gy, i.e., the square root of the sum of the squares of the two directional gradient components for each pixel. The resulting gradient magnitude map has the exact same width and height as the original infrared thermal image. For example, a 3×3 window consisting of a pixel and its 8 neighboring temperature values ​​is: [[98,100,102],[97,100,103],[96,99,101]] (unit: °C). After convolving with the Sobel horizontal operator, Gx = 8.0, and after convolving with the Sobel vertical operator, Gy = -4.0. Therefore, the gradient magnitude of this pixel is... Perform this operation on all 640×480 pixels of the image to obtain a gradient magnitude map of the same size. Divide the gradient magnitude map into a preset number of equal-area sub-regions, count the pixel distribution of gradient magnitude in each sub-region, and construct a gradient magnitude histogram for each sub-region. It should be noted that the gradient magnitude map is evenly divided into rectangular sub-regions of equal area according to a preset number of rows and columns. The number of sub-regions is determined based on the characteristics of the die head's discharge section and computational resources; a division of 8 rows × 8 columns, totaling 64 sub-regions, is recommended. This division has been experimentally verified to maintain sufficient histogram statistical sample size while ensuring spatial resolution. For each sub-region, the gradient magnitude of all pixels within that region is traversed, and the gradient magnitude range is divided into several consecutive statistical intervals using a preset group interval. The number of pixels falling into each interval is counted, and a gradient magnitude histogram is constructed with the gradient magnitude interval as the horizontal axis and the pixel frequency as the vertical axis. For example, a 640×480 gradient magnitude map can be divided into 64 equal-area sub-regions of 8 rows × 8 columns, with each sub-region containing 80×60 pixels; the group interval is set to 1.0. ℃ / pixel, divide the gradient magnitude range into several intervals such as [0,1), [1,2), [2,3), etc., and count the number of pixels falling into each interval in each sub-region to form 64 gradient magnitude histograms; The gradient magnitude histogram of each sub-region is normalized to obtain the gradient probability density distribution of the sub-region, and the local thermal entropy of the temperature field of the sub-region is calculated based on the gradient probability density distribution. It should be noted that L1 norm normalization is performed on the gradient magnitude histogram of each sub-region, which means dividing the frequency of each interval by the total number of pixels in that sub-region, so that the sum of the probabilities of all intervals equals 1, thus obtaining the gradient probability density distribution. Based on the definition of information entropy, the local thermal entropy value of the temperature field in this sub-region is equal to the probability density of each gradient amplitude interval. Multiply by the logarithm of the probability density with base 2 The sum of the products takes a negative sign, that is... The lower the local thermal entropy value, the more concentrated and orderly the temperature gradient distribution within the sub-region; the higher the local thermal entropy value, the more diffuse and chaotic the temperature gradient distribution within the sub-region. For example, a sub-region has 4800 pixels, and the histogram statistics show that there are 800 pixels in the gradient interval [0,1), 1600 pixels in [1,2), 1200 pixels in [2,3), 800 pixels in [3,4), and 0 pixels in the remaining intervals; the normalized probability densities are 0.167, 0.333, 0.250, and 0.167, respectively; the local thermal entropy value... If the gradient magnitudes of all pixels in the sub-region are concentrated within the same interval, the probability is 1. =0, then the thermal entropy value is 0 bit, indicating that the gradient distribution is completely ordered; Arrange the local thermal entropy values ​​of the temperature field of all sub-regions according to their spatial location to form the thermal entropy distribution matrix corresponding to the infrared thermal image frame. It should be noted that the local thermal entropy values ​​of the temperature field calculated from the 64 sub-regions are arranged sequentially according to the row and column positions of the sub-regions in the original gradient magnitude map, forming an 8-row × 8-column thermal entropy distribution matrix. Each element in this matrix records the degree of thermodynamic disorder of the temperature field in the corresponding spatial sub-region, and the spatial resolution depends on the sub-region division density. For example, the thermal entropy distribution matrix is ​​an 8×8 floating-point matrix, where the value of the element in the 3rd row and 5th column is 2.13 bits, indicating that the temperature gradient information entropy of the sub-region at that spatial location is 2.13 bits, reflecting a high degree of disorder in heat transfer in this local region, which may correspond to a region of insufficiently gelatinized raw starch clumps. The coefficient of variation is calculated pixel by pixel for multiple thermal entropy distribution matrices corresponding to consecutive frames of infrared thermal images within a time window to obtain the thermal entropy coefficient of variation matrix; where each element in the coefficient of variation matrix represents the temporal stability of the temperature field order at the corresponding spatial location. It should be noted that a sliding time window containing K consecutive infrared thermal images is set (the time window length is set to 5 seconds based on the dynamic characteristics of the process, and K=150 frames at a frame rate of 30 Hz). Within the window, K thermal entropy values ​​are extracted for each spatial location to form a time series; the standard deviation of this series is calculated. with the mean The ratio of these values ​​is the coefficient of variation.

[0022] A coefficient of variation matrix of the same size as the thermal entropy distribution matrix is ​​constructed. A small coefficient of variation indicates small temporal fluctuations in thermal entropy and a stable gelatinization state at that location, while a large coefficient of variation indicates drastic fluctuations in thermal entropy and an unstable gelatinization state at that location. For example, with a sliding time window set to 5 seconds and a sampling frequency of 30 Hz, the window contains 150 frames of the thermal entropy distribution matrix. 150 thermal entropy values ​​are extracted from the 3rd row, 5th column position, with a mean of 2.10 bits and a standard deviation of 0.42 bits. The coefficient of variation at that position is... =0.42 / 2.10=0.20; If the mean at a certain position is 1.50 bits and the standard deviation is 0.15 bits, then the coefficient of variation is 0.10, indicating that the gelatinization state at that position is more stable. Weighted pooling of the thermal entropy variation coefficient matrix yields a scalar value as the thermal entropy characteristic value characterizing the uniformity of material gelatinization. It should be noted that the spatial weight matrix is ​​determined based on the flow velocity distribution at the extrusion die's discharge section. The normalized axial velocity field of the material at the die exit is obtained using computational fluid dynamics simulation, and this velocity field is downsampled to the same size as the thermal entropy variation coefficient matrix, then normalized and used as the weight matrix. Typically, the central high-velocity region is assigned a larger weight, while the peripheral low-velocity region is assigned a smaller weight. The weighted pooling formula is: ; in For normalized weights, satisfying This characteristic value comprehensively reflects the spatially weighted average time stability of the gelatinization uniformity across the entire discharge section. For example, an 8×8 weight matrix is ​​obtained through CFD simulation, with the four central cells each having a weight of 0.08, the innermost ring having a weight of 0.05, and the edge regions having a weight of 0.02. The sum of all weights is 1. The thermal entropy eigenvalue is obtained by multiplying the current frame's coefficient of variation matrix element-wise with the weight matrix and then summing the results. The smaller the value, the more uniform and stable the gelatinization is across the entire cross-section; the larger the value, the more uneven and unstable the gelatinization is in some areas.

[0023] In an optional embodiment, a rheological correction factor is constructed based on the thermal entropy characteristic value and the material moisture content, including: Obtain the timestamp-aligned material moisture content sequence and thermal entropy feature value sequence, and perform a sliding window average on the material moisture content sequence to obtain the material moisture content trend value; It should be noted that the material moisture content sequence is subject to random noise and high-frequency fluctuations due to online measurement. Therefore, a sliding window averaging filter is used for smoothing. The window length is selected as 1 second (10 sampling points) based on the moisture content measurement frequency and process dynamics. The arithmetic mean of all sampling points within the window is taken as the moisture content trend value at the center of the window to eliminate measurement noise and retain the process change trend. For example, if the moisture content sampling frequency is 10 Hz and the sliding window length is set to 10 sampling points, i.e., a 1-second time span, the 10 consecutive moisture content values ​​are 31.2%, 31.5%, 31.3%, 31.6%, 31.4%, 31.5%, 31.7%, 31.5%, 31.6%, and 31.4%, and their arithmetic mean is 31.47%, which is used as the moisture content trend value at the center of the time window. The thermal entropy feature value sequence is matched one-to-one with the material moisture content trend value according to the timestamp to form a time-varying two-dimensional state vector. It should be noted that, on a unified timeline, the thermal entropy characteristic value and the material moisture content trend value corresponding to each timestamp are taken to form a state vector containing two-dimensional components. This vector represents a comprehensive description of the material's gelatinization state at the current moment in a two-dimensional state space, where the first dimension is the moisture content trend value (unit: %) and the second dimension is the thermal entropy characteristic value (dimensionless). For example, at timestamp 10:00:05.000, the moisture content trend value is 31.47% and the thermal entropy characteristic value is 0.35, then the two-dimensional state vector at this time is represented as [31.47, 0.35]. From the pre-defined gelatinized state reference library, retrieve the reference state point with the smallest Euclidean distance to the time-varying two-dimensional state vector; wherein, the gelatinized state reference library contains a reference record consisting of multiple reference state points and the standard rheological coefficients corresponding to each reference state point. It should be noted that the gelatinization state reference library was pre-established through offline system identification experiments, covering a moisture content range of 24%–36% (step size 0.5%) and a thermal entropy characteristic value range of 0.15–0.55 (step size 0.02), totaling approximately 1050 reference records. Under laboratory conditions, using a high-pressure capillary rheometer, the rheological properties of millet flour gelatinized products were tested under different moisture contents and different gelatinization uniformities covering the process range from low to high, obtaining the corresponding non-Newtonian rheological parameters, i.e., standard rheological coefficients, for each condition. (Dimensionless corrected power exponent, typical range 0.80–1.50); ​​To accelerate online retrieval, the data in the reference library is pre-organized into a KD-tree spatial index structure, making the average time complexity of finding the nearest neighbor among 1050 records O(log N), with a single retrieval time of less than 50μs, meeting the millisecond-level closed-loop requirement. During online retrieval, the current state vector is first Z-score standardized for each dimension, and then the Euclidean distance with each reference point is calculated, selecting the record corresponding to the minimum value; for example, the fuzzy state reference library contains 1050 reference records. After standardization, the standardized Euclidean distance between the current state vector [31.47, 0.35] and the reference point [32.0, 0.40] is 0.53, which is the minimum; the standard rheological coefficient corresponding to this reference point is... ; Extract the standard rheological coefficient corresponding to the reference state point with the smallest distance as the reference rheological coefficient; It should be noted that the standard rheological coefficient associated with the reference record with the smallest Euclidean distance is extracted from the retrieved reference record and used as the baseline rheological coefficient under the current operating conditions. This coefficient is a dimensionless exponent, characterizing the degree of nonlinearity in stress response caused by the non-Newtonian rheological properties of the material at the current moisture content and gelatinized state; for example, if the most recent reference record shows a standard rheological coefficient of 1.12 corresponding to the state point [32.0, 0.40], then the reference rheological coefficient is... ; Calculate the deviation rate of the characteristic value of thermal entropy from its ideal homogeneous state thermal entropy value, and at the same time calculate the deviation rate of the material moisture content from its nominal moisture content; It should be noted that the ideal homogeneous state thermal entropy value Based on process calibration experiments, the recommended moisture content for this millet flour formula is 0.30; the nominal moisture content is... The target moisture content set for the product formula is 30.0%; the deviation rate is calculated using the following formula: ; The two bias rates are weighted and fused to obtain the time-varying dynamic bias. ; For example, the ideal homogeneous thermal entropy value is calibrated to 0.30, and the nominal water content is 30.0%; the current thermal entropy characteristic value is 0.35, and the deviation rate is... The current moisture content trend value is 31.47%, with a deviation rate of... If the weighting of thermal entropy deviation α = 0.60 and the weighting of moisture content deviation β = 0.40, then the time-varying dynamic bias is... ; The rheological correction factor is obtained by superimposing the reference rheological coefficient with the time-varying dynamic bias. It should be noted that the rheological correction factor Equal to the reference rheological coefficient With time-varying dynamic bias The algebraic sum, that is:

[0024] The baseline rheological coefficient provides a basic correction based on historical empirical data, while the time-varying dynamic bias provides adaptive fine-tuning based on real-time state deviations. The combination of the two makes the rheological correction factor both accurate and real-time. The theoretical range is 0.80 to 1.80, when The nominal shear stress decreases over time. This amplifies the nominal shear stress. In an optional embodiment, the two bias rates are weighted and fused to obtain a time-varying dynamic bias, including: The first weighting coefficient of the thermal entropy deviation rate is dynamically determined based on the rate of change of the thermal entropy characteristic value sequence within the sliding time window. It should be noted that within the 5-second sliding time window, the mean of the absolute first-order differences of the thermal entropy characteristic value sequence is used as the rate of change. The first weighting coefficient is determined according to the preset linear mapping function. : ; in The upper limit reference value for the rate of change of thermal entropy is set to 0.50 based on statistics; this mechanism means that the more drastic the fluctuation of the gelatinization uniformity, the higher the weight of its deviation in the fusion process. For example, the rate of change of thermal entropy characteristic value within 5 seconds is calculated as follows: , =0.50, then After clamping However, to make the example consistent with the previous ones, we will use a different set of numbers here: if ,but If the aforementioned example weight of 0.60 is actually used, then the corresponding rate of change is... ; The second weighting coefficient of the moisture content deviation rate is dynamically determined based on the rate of change of the material moisture content sequence within the sliding time window; wherein the sum of the first weighting coefficient and the second weighting coefficient is 1. It should be noted that, similarly, the absolute first-order difference mean of the material moisture content series within the same 5-second sliding time window is calculated as the rate of change. Second weighting coefficient Determined by the following formula;

[0025] For example, rate of change in moisture content ,at this time Since the rate of change of thermal entropy has been determined to be 0.60, then... If the moisture content of the material changes drastically, leading to... Much larger The system will dynamically adjust in the next sliding window. Increase; The time-varying dynamic bias is obtained by weighting and summing the deviation rates of the thermal entropy characteristic value and the material moisture content using the first and second weighting coefficients. It should be noted that the formula for calculating the time-varying dynamic bias Δ is: ; The weighted summation result serves as the adaptive dynamic correction term for the baseline rheological coefficients; for example, , , , ,but .

[0026] In an optional embodiment, a non-Newtonian fluid shear compensation calculation is performed on the screw torque time series signal using a rheological correction factor to obtain an equivalent shear stress sequence, including: Obtain the screw torque timing signal and rheological correction factor sequence after timestamp alignment; Baseline drift removal and power frequency notch filtering are performed on the screw torque timing signal to obtain a preprocessed torque signal; It should be noted that baseline drift is caused by sensor temperature drift and circuit zero drift, manifested as a slowly changing DC offset superimposed on the torque signal. A fourth-order Butterworth high-pass digital filter with a cutoff frequency of 0.5 Hz is used to remove baseline drift. Power frequency interference is introduced by electromagnetic coupling of the 50 Hz AC power supply, manifested as a 50 Hz sinusoidal noise superimposed on the torque signal. This is suppressed using an IIR notch filter with a center frequency of 50 Hz and a quality factor of 30. The pre-processed torque signal is obtained after two stages of filtering. For example, the original torque signal is superimposed with a slowly rising baseline drift of 0.2 N·m and a 50 Hz power frequency interference with an amplitude of 0.1 N·m. After removing the baseline drift by high-pass filtering and attenuating the power frequency component by 40 dB by the notch filter, a clean pre-processed torque signal is obtained, with a signal-to-noise ratio improvement of approximately 20 dB. The preprocessed torque signal is substituted into the preset screw torque-shear stress conversion equation to obtain the nominal shear stress sequence; the screw torque-shear stress conversion equation is pre-established based on the extruder screw geometric parameters and the power-law fluid constitutive equation; It should be noted that the screw torque-shear stress conversion equation is an analytical expression derived from the rectangular screw channel parallel plate model in single-screw extrusion theory. It assumes the material is completely filled in the metering section, the flow is a fully developed steady-state laminar flow, and the power-law fluid constitutive equation is... ,in The screw torque is a non-Newtonian exponent, assuming simplified conditions that neglect pressure flow and leakage. Primarily contributed by drag flow shear, which is related to the nominal shear stress. After integrating the force balance and velocity distribution equations, the relationship can be simplified into the following explicit equation: ; in, The helix angle, The outer diameter of the screw. For the screw groove depth, the equation gives the torque value measured in real time. Direct calculation of nominal shear stress The complete mapping relationship has no undetermined empirical coefficients. If the influence of pressure flow needs to be considered, a correction coefficient determined by the screw characteristic experiment can be introduced. Multiply by the right side of the expression, The value is calibrated to be 0.95–1.05 under standard operating conditions; A twin-screw extruder is equivalent to a single-screw model, with a helix angle of... screw outer diameter mm, threaded groove depth Non-Newtonian exponent , to torque Substitute into the above formula: ;molecular ; denominator ; That is, 2685 kPa; Note: In actual production, due to differences in screw configuration, this nominal value may be scaled by calibration, but its functional relationship with torque is determined by this formula; Obtain the rheological correction factor value that matches the timestamp of each nominal shear stress value in the rheological correction factor sequence. Use the rheological correction factor value as the exponent to perform power exponent compensation operation on the nominal shear stress sequence point by point to obtain the equivalent shear stress sequence. It should be noted that the explicit equation for the power compensation operation is:

[0027] in As a rheological correction factor (dimensionless), this calculation utilizes power-law transformation to uniformly compensate for deviations in the gelatinization state and the nonlinear torque-stress relationship caused by non-Newtonian characteristics, thus ensuring the equivalent shear stress... Better linearly proportional to the actual flow resistance of the material; when hour, = ;when hour, Amplified, this indicates that the actual flow resistance of the material is higher than the nominal value due to uneven gelatinization or moisture content deviation; when The opposite is true at other times.

[0028] In an optional embodiment, the instantaneous apparent viscosity of millet flour is calculated using a capillary rheological inversion equation based on preset extrusion die geometry parameters and an equivalent shear stress sequence, including: Obtain the die hole diameter and die hole length of the extrusion die as geometric parameters of the extrusion die; It should be noted that the die hole diameter The diameter of the circular cross-section of the capillary die outlet is denoted by L; the die orifice length L is the axial length of the shaping section of the capillary die. These two parameters are the basic geometric parameters for capillary rheological measurement, directly affecting the conversion relationship between wall shear stress and flow pressure drop. In this embodiment, the die uses a length-to-diameter ratio of L / 100mm. This is a die with a large length-to-diameter ratio. Experiments have verified that the inlet pressure loss (Bagley correction term) accounts for less than 3% of the total pressure drop. Therefore, the Bagley correction is ignored in the online inversion to simplify real-time calculations. For example, the die orifice diameter... mm, die hole length mm, aspect ratio ; Multiply each equivalent shear stress value in the equivalent shear stress sequence by the orifice diameter and divide by four times the orifice length to obtain the wall shear stress sequence. It should be noted that for a fully developed steady-state laminar flow within a capillary, the wall shear stress is derived from the force equilibrium relationship as follows: ; Here Using the equivalent value after compensation in the previous step implies a partial correction of the non-Newtonian characteristics; this operation is performed on each sampling point in the equivalent shear stress sequence to obtain the wall shear stress sequence. For example, the equivalent shear stress at a certain moment is 7.32 × 10⁻⁶. 7 Pa (order of magnitude before normalization; scaled-down engineering units are used in actual algorithms), based on the calibrated actual order of magnitude. Given that Pa is the diameter of the die hole (3.0 mm) and the length of the die hole (90.0 mm), the wall shear stress is... (Note: This value is consistent with 1220 Pa in the previous example, and both are in engineering units.)

[0029] Obtain the extrusion volume of the extruder per unit time, and divide the extrusion volume by the die cross-sectional area to obtain the apparent shear rate. It should be noted that the extrusion volume Q is obtained in real time through a feeding metering device or a weighing sensor, and the unit is cm³ / s; the die cross-sectional area... The apparent shear rate is defined as: ; This parameter serves as the characteristic rate of capillary flow; for the power-law fluid in this scheme, the actual wall shear rate is... Should be corrected by Rabinowitsch: However, given the equivalent shear stress Rheological correction factor has been passed Non-Newtonian exponents included Indirect compensation was performed, and to maintain the simplicity of online calculations, the apparent shear rate was directly adopted. Systematic biases have been covered by calibration using a gelatinized state reference library; For example, the volume of extrusion per unit time by an extruder Mold hole diameter Cross-sectional area Apparent shear rate If Rabinowitsch correction is considered, Time correction factor The actual shear rate is approximately 298.8 s⁻¹, but the rheological correction factor in this scheme has already absorbed this difference, so the apparent shear rate is used directly. Divide each wall shear stress value in the wall shear stress sequence by the apparent shear rate to obtain the initial apparent viscosity sequence. It should be noted that the formula for calculating the initial apparent viscosity is: ; This ratio follows the definition of Newtonian fluids, and the resulting sequence includes dynamic inertial effects; For example, , The initial apparent viscosity ; Divide the difference in equivalent shear stress between adjacent sampling points in the equivalent shear stress sequence by the sampling time interval to obtain the shear stress change rate sequence. It should be noted that the rate of change of shear stress is calculated using backward difference: ; in, The sampling period is 5 ms; this rate of change reflects the dynamic speed of stress caused by changes in the working state of the extruder or sudden changes in material properties. For example, the previous sampling point ,current Pa, difference +20 Pa, rate of change ; Based on the absolute value of the shear stress change rate sequence, time-varying inertia compensation is performed on the initial apparent viscosity sequence to obtain the instantaneous apparent viscosity sequence. It should be noted that the complete calculation formula for time-varying inertia compensation is as follows: ; in, The inertial compensation coefficient is determined by the system identification experiment; the calibration value of this scheme is [value missing]. Pa·s / (Pa / s); This formula means that when the shear stress changes rapidly (whether positive or negative), the initial apparent viscosity will include an additional inertial drag component caused by the acceleration or deceleration of the material. This component is proportional to the absolute value of the rate of change of stress and needs to be subtracted from the initial value to obtain the instantaneous apparent viscosity that better reflects the true rheological properties of the material. For example, , , Compensation items , When the rate of change is extremely high (e.g., 260,000 Pa / s), the compensation amount can reach 0.312 Pa·s, which has a significant effect on correcting the viscosity value.

[0030] In an optional embodiment, the instantaneous apparent viscosity is compared with a preset standard viscosity threshold range. If the viscosity exceeds the range, a control command is generated and fed back to the extruder control system to achieve closed-loop real-time viscosity monitoring, including: Obtain the instantaneous apparent viscosity sequence, and compare the instantaneous apparent viscosity value of each sampling point in the sequence with the preset standard viscosity threshold range point by point; It should be noted that the standard viscosity threshold range is determined by the product process specifications and includes the lower limit. and upper limit value These represent the minimum and maximum viscosity values ​​allowed for product qualification, respectively; point-by-point comparison involves determining whether each sample value in the instantaneous apparent viscosity sequence meets the requirements. ≤ ≤ For example, the standard viscosity threshold range for millet flour products is [5200, 6000] mPa·s; at a certain moment, the instantaneous apparent viscosity is 5438 mPa·s, which falls within the range and is considered normal; at another moment, it is 6150 mPa·s, which exceeds the upper limit and is considered abnormal. In response to the instantaneous apparent viscosity value at the current sampling point exceeding the standard viscosity threshold range, a timer is triggered to start accumulating the duration of the anomaly. It should be noted that, to distinguish between transient interference and continuous process deviation, a timer is started when a sampling point first exceeds the threshold range. The timer continues to accumulate during the period of continuous exceedance, and the timer resets to zero once a sampling point returns to the normal range. Only when the duration of the abnormality exceeds the preset continuous abnormality duration threshold of 1.0 second is it identified as a real process abnormality event, triggering subsequent control procedures. For example, if the preset continuous abnormality duration threshold is 1.0 second, at a sampling frequency of 200 Hz, if 200 consecutive sampling points exceed the threshold range, the accumulated time = 200 × 0.005 = 1.0 second, triggering control. If the process returns to normal after only 50 consecutive sampling points (0.25 seconds), the timer resets to zero, and control is not triggered. When the duration of the anomaly reaches the preset threshold for continuous anomaly duration, the mean of all instantaneous apparent viscosity values ​​that exceed the interval within the anomaly duration is calculated and used as the deviation viscosity value. It should be noted that the abnormal period may contain multiple consecutive out-of-limit sampling points. The arithmetic mean of the instantaneous apparent viscosity values ​​of all out-of-limit sampling points is used to obtain the deviation viscosity value for this abnormal event. Mean value calculation can suppress the influence of individual extreme noise sampling values ​​on control decisions; for example, there are 200 sampling points exceeding the limit within an abnormal period of 1.0 seconds, the cumulative viscosity value is 1250000 mPa·s, and the deviation viscosity value is 1250000 / 200=6250 mPa·s. The difference between the deviation viscosity value and the center value of the standard viscosity threshold range is calculated to obtain the viscosity deviation amount; It should be noted that the center value of the standard viscosity threshold range This represents the optimal target viscosity for the process; viscosity deviation. The calculation formula is:

[0031] Positive deviation indicates that the material is too thick, while negative deviation indicates that the material is too thin; for example, the standard range [5200, 6000] has a center value of 5600 mPa·s, and the viscosity deviation is 6250 mPa·s. The viscosity deviation is... This indicates that the current viscosity of the material is too high; The direction of adjustment is determined based on the sign of the viscosity deviation; the direction of adjustment includes increasing or decreasing the viscosity. It should be noted that, The viscosity needs to be adjusted by increasing the amount of water or decreasing the amount of powder. The viscosity needs to be adjusted by reducing the amount of water or increasing the amount of powder; heating power is used as an auxiliary adjustment method, usually... Reduce the heating temperature appropriately. Increase the heating temperature appropriately; The control step size level is determined based on the preset grade range in which the absolute value of the viscosity deviation falls; It should be noted that the absolute value of the viscosity deviation... The system is divided into three preset intervals, each corresponding to a different control step size intensity. The specific hierarchical logic is as follows: Adjust step size =

[0032] The feeding and heating adjustment percentages for each level are determined by the preset gain table: Level I fine-tuning: Feeding rate adjustment ±1%, heating power adjustment ±0.5%; Level II intermediate adjustment: feed rate adjustment ±3%, heating power adjustment ±1.5%; Level III emphasizes: feed rate adjustment ±5%, heating power adjustment ±3%; This grading table is calibrated through a step response test to ensure that adjustments do not overshoot under normal operating conditions and that rapid correction is possible in case of severe deviations. All adjustments are based on the current operating parameters and are superimposed with actuator limits (feeding rate not exceeding ±15% of the rated value, heating power not exceeding ±10% of the rated power) to prevent exceeding the equipment's capacity. For example, currently If it falls into interval III, the adjustment step size level is determined to be Level III for emphasis; The control direction and control step size are combined into a control command code, which is fed back to the feed rate control unit and heating power control unit of the extruder control system to adjust the extruder operating parameters in a closed loop so that the subsequent instantaneous apparent viscosity returns to the standard viscosity threshold range. It should be noted that the control command code adopts "direction-level" encoding, such as "DOWN-3" indicating a decrease in viscosity and a level III step size. The command is sent to the PLC via the EtherCAT industrial Ethernet bus at a cycle of 1 kHz. The control logic in the PLC calculates the new feed screw speed setpoint and heater duty cycle based on the current operating parameters, and writes them to the frequency converter and SCR power regulator respectively. The system also has built-in actuator delay compensation, which pauses the generation of new control commands within a 1.5-second window after the command is issued (identified by the object response time) to prevent oscillation. If the viscosity remains abnormal after adjustment and exceeds the safety upper limit of 7500 mPa·s or the lower limit of 4500 mPa·s, the safety interlock is triggered, forcibly switching to bypass circulation and triggering an alarm.

[0033] Example 2, please refer to Figure 2 This invention provides a technical solution: an online real-time viscosity monitoring system for millet flour processing based on edge computing, which is applicable to the aforementioned online real-time viscosity monitoring method for millet flour processing based on edge computing, comprising: Multimodal acquisition module 1 is used to simultaneously acquire screw torque timing signal, discharge port infrared thermal image and material moisture content at the millet flour extrusion and cooking outlet; The thermal entropy feature extraction module 2 is used to perform temperature field gradient analysis on infrared thermal images using edge computing nodes, and extract thermal entropy feature values ​​that characterize the uniformity of material gelatinization. Rheological stress compensation module 3 is used to construct a rheological correction factor based on the thermal entropy characteristic value and the material moisture content, and to perform non-Newtonian fluid shear compensation calculation on the screw torque time series signal using the rheological correction factor to obtain the equivalent shear stress sequence. The viscosity inversion calculation module 4 is used to calculate the instantaneous apparent viscosity of millet powder based on the preset extrusion die geometric parameters and equivalent shear stress sequence through the capillary rheological inversion equation. The closed-loop control instruction module 5 is used to compare the instantaneous apparent viscosity with the preset standard viscosity threshold range. If it exceeds the range, a control instruction is generated and fed back to the extruder control system to realize closed-loop real-time monitoring of viscosity.

[0034] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. A method for real-time online viscosity monitoring in millet flour processing based on edge computing, characterized in that, include: Simultaneously collect screw torque timing signals, discharge port infrared thermal images, and material moisture content at the millet flour extrusion and cooking outlet; The infrared thermal image is analyzed by edge computing nodes to extract thermal entropy feature values ​​that characterize the uniformity of material gelatinization. Based on the thermal entropy characteristic value and the moisture content of the material, a rheological correction factor is constructed. The rheological correction factor is used to perform non-Newtonian fluid shear compensation calculation on the screw torque time series signal to obtain an equivalent shear stress sequence. The instantaneous apparent viscosity of millet powder is calculated using the capillary rheological inversion equation based on the preset extrusion die geometry parameters and the equivalent shear stress sequence. The instantaneous apparent viscosity is compared with a preset standard viscosity threshold range. If the viscosity exceeds the range, an adjustment command is generated and fed back to the extruder control system to achieve closed-loop real-time monitoring of viscosity.

2. The method for real-time online viscosity monitoring of millet flour processing based on edge computing according to claim 1, characterized in that, Simultaneously collect screw torque timing signals, discharge port infrared thermal images, and material moisture content at the millet flour extrusion and cooking outlet, including: A strain gauge torque sensor installed on the screw drive shaft of the extruder continuously acquires screw torque timing signals at a first preset sampling frequency and records a first timestamp sequence. An infrared thermal imager mounted above the extrusion die outlet is used to continuously capture infrared thermal images of the material discharge flow at a second preset sampling frequency, and a second timestamp sequence is recorded. The moisture content of the material at the moment of extrusion is obtained online by a near-infrared moisture content measuring probe set at the discharge port, and the third timestamp sequence is recorded. In the edge computing node, the second and third timestamp sequences are interpolated and resampled using the first timestamp sequence as a reference, so that the screw torque timing signal, infrared thermal image and material moisture content are aligned on the time axis, and synchronous acquisition is achieved.

3. The method for real-time online viscosity monitoring of millet flour processing based on edge computing according to claim 2, characterized in that, The infrared thermal image is analyzed using edge computing nodes to extract temperature field gradient features that characterize the uniformity of material gelatinization, including: The edge computing node acquires the timestamp-aligned infrared thermal image, calculates the local temperature gradient amplitude pixel by pixel for a single frame of infrared thermal image, and obtains a gradient amplitude map of the same size as the original infrared thermal image. The gradient magnitude map is divided into a preset number of equal-area sub-regions. The pixel distribution of gradient magnitude in each sub-region is statistically analyzed, and a gradient magnitude histogram of each sub-region is constructed. The gradient magnitude histogram of each sub-region is normalized to obtain the gradient probability density distribution of the sub-region, and the local thermal entropy of the temperature field of the sub-region is calculated based on the gradient probability density distribution. Arrange the local thermal entropy values ​​of the temperature field of all sub-regions according to their spatial location to form the thermal entropy distribution matrix corresponding to the infrared thermal image frame. The coefficient of variation is calculated pixel by pixel for multiple thermal entropy distribution matrices corresponding to multiple consecutive frames of infrared thermal images within a time window to obtain the thermal entropy coefficient of variation matrix; wherein, each element in the coefficient of variation matrix represents the temporal stability of the temperature field order at the corresponding spatial location. The thermal entropy variation coefficient matrix is ​​weighted and pooled to obtain a scalar value as the thermal entropy characteristic value representing the uniformity of material gelatinization.

4. The method for real-time online viscosity monitoring of millet flour processing based on edge computing according to claim 3, characterized in that, A rheological correction factor is constructed based on the thermal entropy characteristic value and the material moisture content, including: Obtain the timestamp-aligned material moisture content sequence and the thermal entropy feature value sequence, and perform a sliding window average on the material moisture content sequence to obtain the material moisture content trend value; The thermal entropy feature value sequence is matched one-to-one with the material moisture content trend value according to the timestamp to form a time-varying two-dimensional state vector; From a preset gelatinization state reference library, retrieve the reference state point with the smallest Euclidean distance to the time-varying two-dimensional state vector; wherein, the gelatinization state reference library contains a reference record consisting of multiple reference state points and the standard rheological coefficients corresponding to each reference state point; Extract the standard rheological coefficient corresponding to the reference state point with the smallest distance as the reference rheological coefficient; The deviation rate between the thermal entropy characteristic value and its ideal homogeneous state thermal entropy value is calculated, and the deviation rate between the material moisture content and its nominal moisture content is calculated. The two deviation rates are weighted and fused to obtain the time-varying dynamic bias. The rheological correction factor is obtained by superimposing the reference rheological coefficient with the time-varying dynamic bias.

5. The method for real-time online viscosity monitoring of millet flour processing based on edge computing according to claim 4, characterized in that, The two bias rates are weighted and fused to obtain the time-varying dynamic bias, including: Based on the rate of change of the thermal entropy feature value sequence within the sliding time window, the first weighting coefficient of the thermal entropy deviation rate is dynamically determined. The second weighting coefficient of the moisture content deviation rate is dynamically determined based on the rate of change of the material moisture content sequence within the sliding time window; wherein the sum of the first weighting coefficient and the second weighting coefficient is 1. The time-varying dynamic bias is obtained by weighting and summing the deviation rates of the thermal entropy characteristic value and the material moisture content using the first weighting coefficient and the second weighting coefficient.

6. The method for real-time online viscosity monitoring of millet flour processing based on edge computing according to claim 5, characterized in that, Using the rheological correction factor, non-Newtonian fluid shear compensation calculations are performed on the screw torque time series signal to obtain an equivalent shear stress sequence, including: Obtain the timestamp-aligned screw torque timing signal and the rheological correction factor sequence; The screw torque timing signal is subjected to baseline drift removal and power frequency notch filtering to obtain a preprocessed torque signal; The preprocessed torque signal is substituted into the preset screw torque-shear stress conversion equation to obtain the nominal shear stress sequence; the screw torque-shear stress conversion equation is pre-established based on the extruder screw geometric parameters and the power-law fluid constitutive equation; Obtain the rheological correction factor value that matches the timestamp corresponding to each nominal shear stress value in the rheological correction factor sequence, and use the rheological correction factor value as the exponential term to perform power exponential compensation operation on the nominal shear stress sequence point by point to obtain the equivalent shear stress sequence.

7. The method for real-time online viscosity monitoring of millet flour processing based on edge computing according to claim 6, characterized in that, The screw torque-shear stress conversion equation is pre-established based on the extruder screw geometry parameters and the power-law fluid constitutive equation, including: Obtain the helix angle, groove depth, and groove width of the extruder screw; Determine the non-Newtonian exponents based on the constitutive equation of power-law fluids; Substituting the helix angle, groove depth, groove width, and non-Newtonian exponent into the analytical expression for screw extrusion shear stress, we obtain the screw torque-shear stress conversion equation with torque as input and nominal shear stress as output.

8. The method for real-time online viscosity monitoring of millet flour processing based on edge computing according to claim 7, characterized in that, Based on the preset extrusion die geometry parameters and the equivalent shear stress sequence, the instantaneous apparent viscosity of millet flour is calculated using the capillary rheological inversion equation, including: Obtain the die hole diameter and die hole length of the extrusion die head as the geometric parameters of the extrusion die head; Multiply each equivalent shear stress value in the equivalent shear stress sequence by the diameter of the die hole, and divide by four times the length of the die hole to obtain the wall shear stress sequence. The extrusion volume of the extruder per unit time is obtained, and the extrusion volume is divided by the cross-sectional area of ​​the die orifice to obtain the apparent shear rate. Divide each wall shear stress value in the wall shear stress sequence by the apparent shear rate to obtain the initial apparent viscosity sequence. Divide the difference in equivalent shear stress between adjacent sampling points in the equivalent shear stress sequence by the sampling time interval to obtain the shear stress change rate sequence. Based on the absolute value of the shear stress change rate sequence, time-varying inertia compensation is performed on the initial apparent viscosity sequence to obtain the instantaneous apparent viscosity sequence.

9. The method for real-time online viscosity monitoring of millet flour processing based on edge computing according to claim 8, characterized in that, The instantaneous apparent viscosity is compared with a preset standard viscosity threshold range. If it exceeds the range, a control command is generated and fed back to the extruder control system to achieve closed-loop real-time viscosity monitoring, including: The instantaneous apparent viscosity sequence is obtained, and the instantaneous apparent viscosity value of each sampling point in the sequence is compared with the preset standard viscosity threshold range point by point. In response to the instantaneous apparent viscosity value at the current sampling point exceeding the standard viscosity threshold range, a timer is triggered to start accumulating the abnormal duration; When the duration of the abnormality reaches a preset threshold for continuous abnormality duration, the average of all instantaneous apparent viscosity values ​​that exceed the interval within the duration of the abnormality is calculated as the deviation viscosity value. The difference between the deviation viscosity value and the center value of the standard viscosity threshold range is calculated to obtain the viscosity deviation amount; The direction of adjustment is determined based on the sign of the viscosity deviation; wherein, the direction of adjustment includes increasing or decreasing the viscosity. The adjustment step size level is determined based on the preset grade range in which the absolute value of the viscosity deviation falls; The control direction and the control step size are combined into a control command code, which is fed back to the feed rate control unit and heating power control unit of the extruder control system to adjust the extruder operating parameters in a closed loop, so that the subsequent instantaneous apparent viscosity returns to the standard viscosity threshold range.

10. A real-time online viscosity monitoring system for millet flour processing based on edge computing, applicable to the real-time online viscosity monitoring method for millet flour processing based on edge computing as described in any one of claims 1-9, characterized in that, include: The multimodal acquisition module is used to simultaneously acquire screw torque timing signals, discharge port infrared thermal images, and material moisture content at the millet flour extrusion and cooking outlet. The thermal entropy feature extraction module is used to perform temperature field gradient analysis on the infrared thermal image using edge computing nodes, and extract thermal entropy feature values ​​that characterize the uniformity of material gelatinization. The rheological stress compensation module is used to construct a rheological correction factor based on the thermal entropy characteristic value and the material moisture content, and to perform non-Newtonian fluid shear compensation calculation on the screw torque time series signal using the rheological correction factor to obtain an equivalent shear stress sequence. The viscosity inversion calculation module is used to calculate the instantaneous apparent viscosity of millet powder based on the preset extrusion die geometry parameters and the equivalent shear stress sequence through the capillary rheological inversion equation. The closed-loop control command module is used to compare the instantaneous apparent viscosity with a preset standard viscosity threshold range. If the viscosity exceeds the range, a control command is generated and fed back to the extruder control system to achieve closed-loop real-time monitoring of viscosity.

Images