Intelligent control method for cell disruption based on real-time estimation of Dendrobium cell wall disruption rate
By synchronously acquiring broadband vibration and micro-gap signals, and combining the frequency band energy spectrum characteristics and real-time gap value intelligent control method, the problem of the inability to detect the cell wall breakage rate in Dendrobium ultrafine pulverization in real time has been solved. Real-time online estimation and closed-loop control of the cell wall breakage rate have been realized, improving control stability and product quality.
Patent Information
- Application Number
- CN202611280156.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-22
- Publication Date
- 2026-11-17
AI Technical Summary
The current method for ultrafine grinding of Dendrobium cannot detect the cell wall breakage rate in real time online. The single-variable monitoring scheme has insufficient estimation accuracy and generalization ability, and the lack of closed-loop control means leads to frequent over-grinding or under-grinding.
By synchronously acquiring broadband vibration acceleration signals and micro-gap capacitor voltages in the conical grinding disc bearing housing, energy spectrum features of the 5-8kHz frequency band are extracted. A time window feature sequence is constructed by combining the real-time gap value, which is then input into a convolutional neural network model to estimate the cell wall breakage rate. Stable feedback is generated through first-order low-pass filtering and adaptive normalization correction to drive a servo motor to adjust the grinding gap and achieve closed-loop control.
It enables real-time online estimation and closed-loop precise control of Dendrobium cell wall breakage rate, improving control stability and product quality consistency between different batches of materials.
Smart Images

Figure CN123102481A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent detection and control technology for the ultrafine grinding process of Dendrobium, and in particular to an intelligent control method for crushing based on real-time estimation of Dendrobium cell wall breakage rate. Background Technology
[0002] Currently, in the field of ultrafine grinding and cell wall breakage rate detection and control of Dendrobium, the mainstream technical approach generally adopts manual timed shutdown sampling, using offline detection methods such as microscopic counting, chemical titration, or infrared spectroscopy to obtain the actual cell wall breakage rate of the material. The production process relies on operators manually sampling during fixed processing periods, and the cell wall breakage rate result can only be obtained after laboratory sample preparation and testing, resulting in a long detection cycle. If the cell wall breakage rate deviates from the target range, parameters need to be adjusted and resampling and testing are required, leading to a lag in the control process and making it prone to quality problems such as over-grinding or under-grinding.
[0003] In recent years, the industry has gradually explored indirect cell wall breakage rate criteria based on process signals such as mechanical vibration. However, the method of using a single vibration sensor to obtain the overall amplitude information of the grinding disc and inferring the degree of cell wall breakage through a simple energy thresholding strategy is limited by complex process noise and interference from a large number of irrelevant components in the vibration signal, making it difficult to accurately distinguish changes in cell wall breakage rate. In addition, existing control systems mostly adopt an open-loop approach, that is, they rely solely on preset process curves and empirical parameters for operation, lacking real-time closed-loop feedback, making it difficult to guarantee product consistency and the dissolution rate of effective components.
[0004] The existing solutions still have the following shortcomings: On the one hand, in the process of extraction and crushing separation, the correlation between complex process parameters and material changes is weak, making it difficult to achieve dynamic adjustment and control; on the other hand, in an environment that only monitors single variables (such as temperature, current or unidirectional vibration energy), it is impossible to capture the spectral characteristics of cell rupture and its coupling changes with physical shear stress, which cannot support accurate online cell rupture rate inference, and the material sources vary greatly from batch to batch, so the generalization ability and reliability of the single physical quantity acquisition solution are insufficient. Summary of the Invention
[0005] This application provides a method for intelligent control of cell wall disruption based on real-time estimation of Dendrobium cell wall disruption rate. It aims to solve the technical problems of the inability to detect cell wall disruption rate in real time during the existing Dendrobium ultrafine grinding process, the insufficient estimation accuracy and generalization ability of existing single-variable monitoring schemes, and the lack of closed-loop control means that leads to frequent over-grinding or under-grinding.
[0006] The intelligent control method for cell wall disruption based on real-time estimation of Dendrobium cell wall disruption rate provided in this application specifically includes:
[0007] Wideband vibration acceleration signal and micro-gap capacitor voltage are synchronously acquired in the conical grinding disc bearing housing. The acceleration is sampled to obtain the original vibration time sequence data, and the micro-gap capacitor voltage is calibrated and converted to obtain the real-time gap value.
[0008] The original vibration time series data is bandpass filtered to retain the 5-8kHz cell wall breaking characteristic frequency band and suppress the interference frequency band. The filtered signal is then subjected to short-time Fourier transform to extract the frequency band energy spectrum feature vector.
[0009] The frequency band energy spectrum feature vector and the real-time gap value are combined into a continuous time window feature sequence, which is then input into a pre-trained convolutional neural network soft measurement model to output an online estimate of the current cell wall breakage rate.
[0010] The online estimated value is subjected to first-order low-pass filtering and adaptive normalization correction is performed in combination with the statistical distribution characteristics of the previous batch data to generate a stable cell wall breakage rate feedback.
[0011] The stable cell wall breakage rate feedback is compared with the first preset threshold for the lower limit of the cell wall breakage rate and the second preset threshold for the upper limit of the cell wall breakage rate. If it is lower than the first threshold, a gap reduction command is generated; if it is higher than the second threshold, a gap increase command is generated; if it is between the two, a zero position hold command is generated. Based on the comparison result, the dynamic adjustment target value of the axial position of the small end of the conical grinding disc is output.
[0012] Based on the dynamically adjusted target value, the servo motor is driven to perform displacement, thereby changing the axial position of the cutter at the small end of the conical grinding disc and adjusting the grinding gap to achieve real-time closed-loop control of the cell wall breaking rate.
[0013] Furthermore, after adjusting the grinding gap as described above, the following also includes:
[0014] Record the vibration spectrum, real-time gap value, and stable cell breakage rate feedback of the current operating cycle to form an operating log;
[0015] The actual cell wall breakage rate measured by offline microscopic counting method after the current batch ends is used as the calibration value. The deviation between the online estimated value and the calibration value of each sampling period in the batch is calculated, and the arithmetic mean of the deviations of all sampling periods in the batch is taken as the average deviation.
[0016] When the average deviation of multiple consecutive batches accumulates to exceed a preset threshold, an incremental update of the model is triggered to adjust the weight parameters of the convolutional neural network soft measurement model in order to optimize its cross-batch generalization ability.
[0017] The intelligent control method for cell wall disruption based on real-time estimation of Dendrobium cell wall disruption rate provided in this application has the following beneficial effects:
[0018] By synchronously acquiring broadband vibration acceleration and micro-gap capacitance voltage through a conical grinding disc bearing seat and extracting energy spectrum features in the 5-8kHz frequency band, and combining real-time gap values to construct a time window feature sequence input to a convolutional neural network soft measurement model, a stable cell wall breakage rate feedback quantity is generated after first-order low-pass filtering and adaptive normalization correction. The servo motor is driven to perform step displacement through threshold comparison to adjust the grinding gap in a closed loop. This solves the technical problems in existing technologies, such as the inability to detect cell wall breakage rate online in real time, insufficient estimation accuracy and generalization ability of single-variable monitoring schemes, and the lack of closed-loop control methods leading to frequent over-grinding or under-grinding. It realizes real-time online estimation and closed-loop precise control of cell wall breakage rate, effectively improving the control stability and product quality consistency between different batches of materials. Attached Figure Description
[0019] Figure 1 This is the main flowchart of an intelligent control method for the disruption of Dendrobium officinale cells based on real-time estimation of cell wall disruption rate.
[0020] Figure 2 This is a sub-flowchart of a method for intelligent control of cell disruption based on real-time estimation of Dendrobium cell wall disruption rate.
[0021] Figure 3 This is another sub-flowchart of the intelligent control method for cell disruption based on real-time estimation of Dendrobium cell wall disruption rate. Detailed Implementation
[0022] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.
[0023] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.
[0024] like Figure 1 As shown, this application provides an intelligent control method for dendrobium cell wall disruption based on real-time estimation of cell wall disruption rate, specifically including:
[0025] S1: Simultaneously acquire broadband vibration acceleration signals and micro-gap capacitor voltage signals at the conical grinding disc bearing seat, and perform calibration and conversion processing on the micro-gap capacitor voltage signals to obtain the original vibration timing data and real-time gap values containing timestamp alignment information.
[0026] S2: Bandpass filtering is performed on the original vibration time series data to highlight the characteristic frequency band of cell wall rupture, and short-time Fourier transform calculation is performed to obtain the 5-8kHz frequency band energy spectrum feature vector characterizing the degree of Dendrobium fiber breakage.
[0027] S3: Construct a continuous time window feature sequence based on the energy spectrum feature vector of the 5-8kHz frequency band and the real-time gap value, and input the continuous time window feature sequence into a predefined deep convolutional neural network soft measurement model to obtain the online estimate of the current cell wall breakage rate at the current moment.
[0028] S4: Perform first-order low-pass filtering on the current online estimate of cell wall breakage rate to eliminate high-frequency fluctuation interference, and perform adaptive normalization correction in combination with the statistical distribution characteristics of previous batch data to generate a stable cell wall breakage rate feedback quantity for closed-loop control decision.
[0029] S5: Compare the stable cell wall breakage rate feedback with the first preset threshold for the lower limit of the cell wall breakage rate and the second preset threshold for the upper limit of the cell wall breakage rate. If it is lower than the first threshold, a gap reduction command is generated. If it is higher than the second threshold, a gap increase command is generated. If it is between the two, a zero position hold command is generated. Based on the comparison result, the dynamic adjustment target value of the axial position of the small end of the conical grinding disc is output.
[0030] S6: Based on the dynamic adjustment of the target value, the servo motor is driven to perform displacement, thereby changing the axial position of the tool at the small end of the conical grinding disc and adjusting the grinding gap. In turn, the cell wall breakage rate of Dendrobium officinale is controlled in real time by changing the shear stress distribution.
[0031] S7: Record the vibration spectrum, real-time gap value, and stable cell breakage rate feedback of the current operating cycle to form an operating log; use the actual cell breakage rate measured by offline microscopic counting method after the end of the current batch as the calibration value, calculate the deviation between the online estimated value and the calibration value of each sampling cycle in the batch, and take the arithmetic mean of the deviations of all sampling cycles in the batch as the average deviation; when the average deviation of multiple consecutive batches accumulates to exceed the preset threshold, trigger the incremental update of the model, adjust the weight parameters of the convolutional neural network soft measurement model to optimize its cross-batch generalization ability.
[0032] It should be noted that, in this invention, the cell wall breakage rate refers to the percentage of Dendrobium cells whose cell walls have ruptured after grinding, out of the total number of cells in the material. Its true value is determined using an offline microscopic counting method—that is, collecting Dendrobium powder samples and counting intact and broken cells under a microscope, using the proportion of broken cells to the total number of cells as the true calibrated value of the cell wall breakage rate. The online estimation value of this invention is a real-time calculation of this percentage.
[0033] Step S1: Simultaneously acquire broadband vibration acceleration signals and micro-gap capacitor voltage signals at the conical grinding disc bearing housing, and perform calibration and conversion processing on the micro-gap capacitor voltage signals to obtain raw vibration timing data and real-time gap values containing timestamp alignment information. Specifically, this includes:
[0034] S1.1: Based on the radial and axial intersection of the conical grinding disc bearing seat, a broadband vibration sensor and a micro-gap capacitive sensor are fixed using a rigid mounting base to eliminate the phase shift caused by the difference in mechanical vibration transmission paths, thereby obtaining a sensor physical mounting structure with a fixed relative positional relationship.
[0035] At the point where the radial and axial directions of the conical grinding disc bearing housing intersect, a rigid mounting area with a surface flatness tolerance of less than 0.02 mm is selected as the sensor base fixing surface.
[0036] The base is made of high-rigidity aluminum alloy and is an integrated L-shaped rigid mounting base. The bottom of the base is designed with precision positioning pin holes to match the pre-drilled mounting holes of the bearing housing, ensuring that the mechanical coupling rigidity between the base and the bearing housing is maximized.
[0037] The piezoelectric ceramic sensing element of the broadband vibration sensor is fastened to the vertical surface of the L-shaped base with M6 high-strength bolts, so that the sensor sensing axis is strictly parallel to the radial direction of the grinding disc, in order to directly capture the radial impact vibration component caused by cell wall rupture.
[0038] Embed the probe end face of the micro-gap capacitive sensor into the end of the horizontal extension arm of the L-shaped base, and adjust the probe axis so that it is perpendicular to the metal reflective surface of the small end of the conical grinding disc, ensuring that the measuring optical axis and the grinding disc rotation axis maintain a strict geometric orthogonal relationship.
[0039] The spatial relative coordinates of the sensitive center of the vibration sensor and the tip of the capacitive sensor probe are calibrated using a laser interferometer. The fixed offset vectors of the two in three-dimensional space are recorded to eliminate the signal propagation time difference caused by different installation positions.
[0040] Modal analysis tests were performed on the installed base to confirm that the first natural frequency of the base is higher than 15kHz, which is far above the upper limit of the characteristic frequency band of dendrobium cell rupture, thus avoiding the modulation interference of base resonance on high-frequency vibration signals.
[0041] Through the aforementioned rigid connection and precision alignment process, a sensor physical installation structure with a fixed spatial topology is constructed, ensuring the spatiotemporal consistency of vibration signals and gap signals at the physical source.
[0042] By using a rigid base fixation and spatial coordinate calibration, the loose installation requirements of the previous step are transformed into a physical installation structure for sensors with a fixed relative position and no phase shift, thus achieving the expected spatiotemporal synchronization of multimodal signal sources.
[0043] For example, an IEPE wideband vibration sensor with a sensitivity of 100mV / g and an eddy current displacement sensor with a range of 2mm and a resolution of 1μm were selected. The L-shaped base was CNC machined from 7075 aluminum alloy with a thickness of 20mm and fixed to the bearing housing with four M8 bolts at a torque of 50N·m. The vibration sensor was installed 50mm away from the bearing centerline, and the initial distance between the capacitive probe and the side of the grinding disc was set to 1mm. After laser calibration, the spatial distance error between the sensitive points of the two sensors was controlled within ±0.05mm. Modal testing showed that the first-order resonant frequency of the base was 18.5kHz, meeting the requirement of greater than 15kHz. This installation structure allows the acquired vibration signal and gap signal to have physical synchronization without software-level time alignment compensation, significantly improving the signal-to-noise ratio of subsequent feature fusion.
[0044] S1.2: Based on the sensor's physical installation structure, the multi-channel synchronous sampling mode of the data acquisition card is used to perform synchronous acquisition operations on the triaxial acceleration signal and the gap voltage signal to generate the original vibration acceleration sequence and the original gap voltage sequence containing strictly aligned timestamp information.
[0045] Based on the rigid mounting base physical structure constructed using S1.1, the three-dimensional analog output terminal of the broadband vibration sensor and the voltage output terminal of the micro-gap capacitive sensor are respectively connected to the high-impedance differential input channel of the data acquisition card. The internal clock source of the data acquisition card is configured to be a unified master clock signal, and the sampling frequency is set to 10kHz to meet the Nyquist sampling theorem requirement for aliasing-free acquisition of 0-10kHz broadband vibration signals.
[0046] The multi-channel synchronous sampling command is initiated, and the data acquisition card simultaneously triggers the analog-to-digital converters of all channels to read voltage values within each clock cycle, ensuring that the vibration acceleration signal and the gap voltage signal have strict alignment characteristics at the nanosecond level on the time axis. The acquired discrete digital signals are stored in a circular buffer according to the timestamp sequence to form the original vibration acceleration sequence and the original gap voltage sequence.
[0047] By using a hardware-level synchronous sampling mechanism, the phase delay error caused by channel switching in the traditional serial acquisition method is eliminated, and accurate registration of multimodal sensing data in the time domain is achieved, providing a high-fidelity data foundation for subsequent feature fusion.
[0048] For example, an NIPXIe-4499 data acquisition card was selected, configured with four synchronous analog input channels. CH0-CH2 were connected to an IEPE triaxial accelerometer, and CH3 was connected to a capacitive displacement sensor. The sampling rate was set to 10000Hz, and the sampling depth to 1024 points / frame. During the Dendrobium officinale crushing process, the acquisition card synchronously read the voltage values of each channel at 100-microsecond intervals. If at a certain time t0, the CH0 reading is 0.5V (corresponding to 5g acceleration), and the CH3 reading is 2.1V (corresponding to a 0.5mm gap), then these two data points are marked as the same timestamp t0. Data was continuously acquired for 4 seconds, generating a synchronous sequence containing 40,000 time points. This synchronous sequence ensures a strict temporal correspondence between the vibration energy peak and the gap change, avoiding model training bias caused by timing misalignment and significantly improving the feature extraction accuracy of the soft measurement model.
[0049] S1.3: Based on the original gap voltage sequence, a nonlinear mapping conversion process from voltage to distance is performed using a preset sensor calibration curve to obtain the real-time dynamic gap original value that characterizes the actual physical distance between the small end of the grinding disc and the tool.
[0050] Obtain the raw gap voltage sequence output from step S1.2, which contains strictly timestamp aligned information. This sequence reflects the real-time potential changes sensed by the micro-gap capacitive sensor during the Dendrobium pulverization process.
[0051] The sensor static calibration curve data table, which is pre-stored in the embedded industrial computer's memory, is called. This data table establishes a nonlinear mapping relationship between the standard gap distance and the output voltage, which is used to compensate for the edge effect and nonlinear response characteristics of the capacitive sensor.
[0052] For each sampling point voltage value in the original gap voltage sequence, a lookup table interpolation operation is performed. A cubic spline interpolation algorithm is used to smoothly fit the discrete nodes of the calibration curve to eliminate quantization errors and improve conversion accuracy.
[0053] The calculated instantaneous physical distance values are reassembled in chronological order to generate a real-time dynamic gap original value sequence that is completely consistent with the vibration acceleration sequence time reference.
[0054] Through the above-mentioned nonlinear mapping transformation of voltage to distance, the electrical signal result of the previous step is transformed into real-time dynamic gap raw value data that characterizes the actual physical distance between the small end of the grinding disc and the cutting tool. This achieves accurate quantification of the microscopic changes in the crushing gap and provides accurate geometric state input for subsequent filtering and soft measurement of the cell wall breakage rate.
[0055] For example, the calibration curve of the micro-gap capacitive sensor is obtained through static calibration using a high-precision displacement stage within a range of 0 to 2 mm in 0.1 mm steps. For a sample point voltage value of 1.532 V in the original voltage sequence, a query of the calibration database reveals that it lies between two calibration points: 1.50 V (1.20 mm) and 1.60 V (1.10 mm). Using a cubic spline interpolation algorithm, the actual physical gap value at that moment is calculated to be 1.168 mm. The system continuously processes the voltage sequence at a 10 kHz sampling rate, converting all voltage points into a gap value sequence with micron-level precision. This conversion process eliminates measurement distortion caused by sensor nonlinearity, reducing the linearity error of the gap data to within 0.5%, significantly improving the execution accuracy of gap adjustment commands in subsequent closed-loop control.
[0056] S1.4: Based on the original value of the real-time dynamic gap, a moving average filtering algorithm is used to perform high-frequency noise suppression processing to obtain a smooth real-time gap value after eliminating random interference.
[0057] Obtain the real-time dynamic gap raw value sequence after nonlinear mapping transformation. This sequence contains random interference components caused by mechanical high-frequency jitter and electrical noise.
[0058] The window length parameter of the moving average filter is set to 10 sampling points to match the signal smoothing requirements at the current 10kHz sampling rate while also taking into account the real-time response speed.
[0059] Construct a first-in-first-out (FIFO) data queue of length 10, push the latest collected inter-gap raw value to the tail of the queue, and remove the oldest data from the head of the queue to keep the total amount of data in the queue constant.
[0060] Perform an arithmetic mean operation on the 10 existing gap values in the queue to calculate the estimated smooth gap value at the current time.
[0061] The calculated smoothing gap estimate is used as the valid output of the current sampling period, replacing the original fluctuation data and storing it in a temporary buffer.
[0062] By using a moving average filtering method, the original real-time dynamic gap value containing high-frequency noise in the previous step is transformed into a smooth real-time gap value after eliminating random interference, thereby achieving the expected technical effect of improving the signal-to-noise ratio of gap measurement and ensuring the stability of subsequent control commands.
[0063] For example, during the ultrafine grinding process of Dendrobium, a micro-gap capacitive sensor collects the gap voltage at a frequency of 10kHz and converts it into the raw gap value. Due to the mechanical vibration generated by the high-speed rotation of the grinding disc, the raw gap value exhibits high-frequency fluctuations of ±0.02mm around a baseline of 0.50mm. The system initializes a sliding window of length 10. When the raw gap value of 0.512mm is collected at time t, it is added to the window, and the earliest time data is discarded. The sum of the 10 data points within the window (e.g., 0.501, 0.498, ..., 0.512) yields 5.005mm, which is divided by 10 to obtain a smoothed gap value of 0.5005mm. This processing effectively filters out high-frequency spikes caused by bearing micro-movements, making the output gap curve smooth and continuous, providing stable gap feature input for the CNN model, and avoiding errors in cell wall breakage rate estimation caused by gap jumps.
[0064] S1.5: Based on the original vibration acceleration sequence and smoothed real-time gap values, a data fusion encapsulation protocol is used to perform time-series data pairing and formatted encapsulation processing to obtain a combined dataset of original vibration time-series data and real-time gap values containing timestamp alignment information for subsequent feature extraction.
[0065] Step S2: Bandpass filtering is performed on the original vibration time series data to highlight the characteristic frequency band of cell wall rupture, and short-time Fourier transform calculation is performed to obtain the 5-8 kHz frequency band energy spectrum feature vector characterizing the degree of Dendrobium fiber breakage. Specifically, this includes:
[0066] S2.1: Perform DC component removal processing on the original vibration time series data to eliminate sensor zero bias error, and perform filtering operation on the DC-removed vibration signal based on the preset Butterworth bandpass filter parameters to obtain a bandpass filtered vibration signal that highlights the cell wall rupture characteristic frequency band.
[0067] Read the original triaxial acceleration sequence containing timestamp alignment information, and extract the time-domain waveform data of a single axis or a composite vector as the signal source to be processed.
[0068] Calculate the arithmetic mean of the time-domain waveform data within a preset sampling window to determine the value of the DC offset component in the signal.
[0069] The DC offset component value is subtracted from each sampling point of the original time-domain waveform data to generate a DC-de-vibration signal sequence with a mean of zero, thus eliminating the baseline deviation caused by sensor installation stress or circuit drift.
[0070] A fourth-order Butterworth digital bandpass filter was constructed, with passband cutoff frequencies set at 4kHz and 9kHz, and stopband attenuation of no less than 40dB, to ensure effective suppression of low-frequency mechanical resonance and high-frequency electronic noise while preserving the 5-8kHz cell wall rupture characteristic frequency band.
[0071] The DC vibration signal sequence is input into a Butterworth digital bandpass filter, and a recursive convolution operation is performed to filter out frequency components outside the passband.
[0072] The output is a bandpass filtered vibration signal with minimized phase distortion, which significantly enhances the proportion of high-frequency impact energy generated at the moment of dendrobium fiber breakage.
[0073] By combining DC removal with bandpass filtering in specific frequency bands, the original vibration data containing complex background noise is transformed into a bandpass filtered vibration signal with significantly improved signal-to-noise ratio and concentrated energy in the characteristic frequency band, thus achieving the expected technical effect of macroscopic signal enhancement for the microscopic physical event of Dendrobium cell wall rupture.
[0074] For example, a raw vibration acceleration sequence with a sampling rate of 10kHz was acquired, and the window length was set to 2048 points. The DC component was calculated to be 0.02g, and 0.02g was subtracted from each sampling point to remove the DC. A fourth-order Butterworth bandpass filter was configured with a lower cutoff frequency of 4kHz and an upper cutoff frequency of 9kHz. Passing the de-DC signal through the filter suppressed the fundamental frequency interference of the grinding disc rotation below 3kHz and the electrical noise above 9kHz. The energy amplitude of the output bandpass filtered signal in the 5-8kHz range was significantly improved compared to the original signal, clearly reflecting the transient vibration characteristics induced by cell wall rupture, providing high-purity input data for subsequent short-time Fourier transform.
[0075] S2.2: Apply the Hanning window function to the bandpass filtered vibration signal for windowing and truncation, and perform a sliding window segmentation operation according to the preset overlap rate to generate a discrete vibration signal frame sequence for short-time spectrum analysis.
[0076] The bandpass filtered vibration signal processed by step S2.1 is received. The DC component has been removed and the cell wall rupture characteristic information in the 4-9kHz frequency band has been retained.
[0077] A sequence of Hanning window functions with a length of 2048 sampling points is constructed, and the cosine weighting principle is used to smooth the attenuation at both ends of the time-domain signal to suppress the spectral leakage effect.
[0078] The Hanning window function is multiplied point-by-point in the time domain with the bandpass filtered vibration signal to generate a single frame vibration data segment after windowing and truncation, ensuring the continuity of the signal edges.
[0079] Set the sliding window overlap rate to 50%, i.e., the step size is 1024 sampling points, and move along the time axis to capture the next frame of data.
[0080] Repeat the windowing and sliding interception operations until the entire effective signal length within the current sampling period is covered, forming a discretized vibration signal frame sequence.
[0081] By using Hanning window truncation and sliding segmentation with 50% overlap, the continuous filtered signal from the previous step is transformed into a discrete vibration signal frame sequence with low spectral leakage characteristics. This achieves the structuring of the standard input data required for short-time Fourier transform, laying the foundation for time-frequency analysis for the subsequent accurate extraction of energy spectrum features in the 5-8kHz frequency band.
[0082] For example, for a bandpass filtered vibration signal with a sampling rate of 10kHz, 2048 points are selected as the number of points for FFT calculation, corresponding to a time window length of 204.8ms. The Hanning window function w(n) = 0.5 * (1 - cos(2pin / (N-1))) is applied, where N = 2048 and n is the sampling index from 0 to 2047. The window function is multiplied by the original signal x(n) to obtain x_w(n). With an overlap rate of 50%, the frame shift is 1024 points. For a signal segment lasting 1 second, approximately 9 discrete vibration signal frames are generated. This parameter configuration significantly reduces sidelobe interference, greatly improving the accuracy of energy peak extraction in the 5-8kHz frequency band and effectively avoiding the estimation deviation of the cell wall breakage rate caused by spectral leakage.
[0083] S2.3: Perform a fast Fourier transform algorithm based on the discrete vibration signal frame sequence to complete the time-frequency domain mapping, and calculate the squared modulus of the transformed complex sequence to obtain instantaneous power spectral density data characterizing the intensity distribution of each frequency component.
[0084] The system receives a sequence of discrete vibration signal frames processed by Hanning windowing and sliding segmentation. This sequence contains transient impact information caused by Dendrobium fiber fracture. A Fast Fourier Transform (FFT) algorithm is performed on each frame of the discrete time-domain signal to map the time-domain amplitude sequence to the complex frequency domain, generating a complex spectrum sequence containing real and imaginary parts. Based on the principle of Discrete Fourier Transform, the square of the magnitude at each frequency point in the complex spectrum sequence is calculated to eliminate phase effects and obtain energy distribution characteristics. The instantaneous power spectral density data is calculated using the following formula:
[0085]
[0086] in, For the first Power spectral density values at each frequency point These are the complex spectral coefficients after the Fast Fourier Transform. The FFT points are 2048.
[0087] By iterating through all frequency points, a complete power spectral density array covering half of the sampling frequency range from 0 to 5 kHz is constructed. This array accurately characterizes the energy intensity distribution of the vibration signal at each frequency component within the current time window, particularly highlighting the high-frequency energy peaks generated by cell wall rupture in the 5-8 kHz band.
[0088] By using Fast Fourier Transform and Modulus Squared Operation, the time-domain discrete signal frame from the previous step is transformed into frequency-domain instantaneous power spectral density data, realizing the feature conversion from time-domain waveform to frequency-domain energy distribution, and providing a standardized data foundation for subsequent extraction of energy features in specific frequency bands.
[0089] For example, a radix-2 Fast Fourier Transform is performed on a discrete vibration signal frame of length 2048 points, outputting 1024 complex spectral coefficients. The 5-8kHz frequency band corresponding to the 500th to 800th coefficients is selected, and the squared modulus of each coefficient is calculated and divided by 2048 to obtain the power spectral density distribution for that frequency band. If the complex coefficient at a certain frequency point is 32+24j, its squared modulus is 1600, and its power spectral density value is 0.78125. Finally, a power spectral density array of length 1024 is generated, where the values in the 5-8kHz interval are significantly higher than the background noise level, accurately reflecting the cell wall disruption state.
[0090] S2.4: Perform integral accumulation on the instantaneous power spectral density data within the target frequency band of 5 to 8 kHz to extract the energy spectral characteristic values of the 5 to 8 kHz frequency band that characterize the degree of dendrobium fiber breakage.
[0091] Receive the instantaneous power spectral density data output from step S2.3, which characterizes the energy distribution intensity of discrete vibration signal frames at each frequency point.
[0092] Determine the target frequency band boundary of 5 to 8 kHz, which corresponds to the characteristic resonance frequency range generated by the brittle fracture of Dendrobium cell walls.
[0093] The power spectral density values of all discrete frequency points within the target frequency band are traversed, and the energy amplitude corresponding to each frequency point is extracted.
[0094] The extracted energy amplitudes are accumulated using the trapezoidal numerical integration method to calculate the total energy integral value within the frequency band.
[0095] The specific calculation formula is as follows:
[0096]
[0097] in, These are the characteristic values of the energy spectrum in the 5-8kHz frequency band. For frequency Power spectral density at that point This refers to the frequency resolution.
[0098] The calculated frequency band energy integral value is used as a characteristic scalar of the current sampling period. This scalar directly reflects the total vibrational energy released by cell wall rupture per unit time.
[0099] By using the above integral accumulation processing method, the spectral distribution data of the previous step is transformed into a single-dimensional energy intensity index, which realizes the quantitative characterization of the degree of dendrobium fiber breakage, effectively suppresses non-characteristic frequency band noise interference, and improves the signal-to-noise ratio of cell wall breakage estimation.
[0100] For example, the sampling rate is set to 10kHz, the FFT points are 2048, and the frequency resolution Δf is approximately 4.88Hz. 5kHz corresponds to index k5 of 1024, and 8kHz corresponds to index k8 of 1638. The power spectral density values P(fi) between indices 1024 and 1638 are iterated. Assuming the average power spectral density in this frequency band at a certain moment is 0.05g² / Hz, the calculated energy characteristic value E is approximately 150g². This value shows a monotonically increasing trend with the increase of the cell wall breakage rate. When the cell wall breakage rate increases from 60% to 90%, the E value significantly increases from 8.2g² to 15.1g², verifying the high sensitivity of this characteristic value to the cell wall breakage state.
[0101] S2.5: Construct a multi-dimensional feature vector based on the energy spectrum feature values of the 5 to 8 kHz frequency band based on the current sampling period and a preset number of historical periods, so as to generate the 5 to 8 kHz frequency band energy spectrum feature vector that is finally used as input to the deep convolutional neural network soft measurement model.
[0102] The system receives the 5-8 kHz frequency band energy spectrum feature values calculated during the current sampling period and stores them as the latest element of the time series in a pre-defined first-in-first-out (FIFO) circular buffer. It then retrieves the frequency band energy spectrum feature values from a pre-defined number of historical periods stored in the buffer, ensuring that the data time window covers the dynamic evolution stage of the Dendrobium fiber fracture process. The feature values in the buffer are arranged chronologically to construct a one-dimensional time series feature array containing information from both the current and historical moments. This one-dimensional time series feature array is then normalized to eliminate dimensional inconsistencies caused by sensor gain fluctuations or differences in the initial state of the material. Finally, the normalized time series feature array is converted into a standardized input tensor format required by a deep convolutional neural network, and the data dimensions are adjusted to match the shape requirements of the model's input layer.
[0103] By constructing a multidimensional feature vector, discrete single-point energy features are transformed into high-dimensional input vectors containing temporal evolution laws, thus achieving a complete characterization of the dynamic characteristics of the Dendrobium cell wall disruption process and providing a high-quality data foundation for subsequent CNN models to extract deep spatiotemporal features.
[0104] For example, the preset historical sampling number is set to 80 sampling periods, corresponding to a 4-second time window. The characteristic value of the 5-8kHz frequency band energy spectrum acquired in the current sampling period t is E. t =12.5mV 2 / Hz. The circular buffer stores E sequentially. t-79 To E t-1 Historical data. The constructed one-dimensional time-series feature array is X=[E t-79 E t-79 ,...,E t The dimension is 80x1. Calculate the mean of the data within this window, μ = 10.2mV. 2 / Hz and standard deviation σ = 2.1mV 2 2 / Hz. Z-score normalization is performed on each element in the array to obtain a normalized feature vector. The normalized feature vector is then reshaped into an 80x1x1 three-dimensional tensor and used as input to the CNN. This process eliminates inter-batch baseline drift, allowing the model to focus on energy change trends and significantly improving the robustness of cell wall breakage rate estimation.
[0105] Step S3: Construct a continuous time window feature sequence based on the 5-8kHz frequency band energy spectrum feature vector and the real-time gap value, and input the continuous time window feature sequence into a predefined deep convolutional neural network soft measurement model to obtain an online estimate of the current cell wall breakage rate at the current moment. Specifically, this includes:
[0106] S3.1: Based on the energy spectrum feature vector of the 5-8kHz frequency band and the real-time gap value, perform multi-period sliding window splicing processing to generate a continuous time window feature sequence containing temporal evolution information, which serves as the standardized input tensor of the deep convolutional neural network soft measurement model.
[0107] The current period 5-8kHz frequency band energy spectrum characteristic value output from S2.5 and the smooth real-time gap value output from S1.4 are combined into a two-dimensional basic feature vector.
[0108] Construct a first-in-first-out circular buffer of length 80 to store the basic feature vectors of 80 consecutive sampling periods, forming a historical data sequence with a time span of 4 seconds.
[0109] When a new base feature vector is generated, it is appended to the end of the buffer, and the oldest feature vector at the beginning of the buffer is removed, thus maintaining the sliding update property of the data window.
[0110] The 80 sets of data in the buffer are reorganized in terms of dimensions, with the time step as the first dimension and the feature type as the second dimension, to construct a two-dimensional matrix structure with a shape of 80×2.
[0111] Z-Score normalization is performed on the two-dimensional matrix to calculate the mean and standard deviation of each feature channel within the time window. The normalized two-dimensional matrix is then converted into a floating-point tensor, and the batch dimension is increased to form a three-dimensional input tensor with a shape of 1×80×2 to match the input interface requirements of the CNN model.
[0112] By using multi-period sliding window splicing and standardization, discrete single-point features are transformed into a continuous time window feature sequence containing temporal evolution laws, thus realizing a complete characterization of the dynamic characteristics of the Dendrobium pulverization process and providing high-dimensional spatiotemporal correlation input for deep neural networks.
[0113] S3.2: Input the continuous time window feature sequence into the pre-built deep convolutional neural network soft measurement model, and use multi-layer one-dimensional convolutional kernels to perform local feature extraction operations to obtain a high-dimensional abstract feature map characterizing the fiber fracture mode of Dendrobium.
[0114] Receive the continuous time window feature sequence generated in step S3.1. This sequence contains 5-8kHz frequency band energy spectrum feature values and real-time gap values for 80 time steps, forming an input tensor with a dimension of 80×2.
[0115] The input tensor is fed into the first one-dimensional convolutional layer of the deep convolutional neural network. This layer is configured with 32 convolutional kernels of size 7, with a stride of 1 and padding mode of SAME to cover local temporal dependencies.
[0116] Convolution operations are performed, and the sliding window mechanism is used to extract the coupling features of local vibration energy change patterns and gap fluctuations on the time axis, generating a preliminary feature map with a dimension of 80×32.
[0117] Batch normalization is applied to the initial feature maps to calculate the mean and variance of each channel feature and perform standardization transformation to accelerate network convergence and eliminate the influence of dimensional differences.
[0118] The standardized features are nonlinearly mapped by the ReLU activation function, and negative values are set to zero. This preserves the positive excitation signal that characterizes the fracture strength of Dendrobium fibers and enhances the model's sensitivity to key fracture events.
[0119] The activated feature map is fed into the second one-dimensional convolutional layer, which is configured with 64 convolutional kernels of size 5 to further extract higher-order abstract features and capture the transient frequency modulation rules in the cell wall rupture process.
[0120] Batch normalization and ReLU activation are performed again to suppress noise interference, enhance the response amplitude of effective feature channels, and generate a medium-level feature map with a dimension of 80×64.
[0121] The intermediate feature map is fed into the third one-dimensional convolutional layer, which is configured with 128 convolutional kernels of size 3. This layer focuses on the subtle structural changes in the extremely short time domain and accurately identifies the spectral texture differences corresponding to different cell wall breaking stages.
[0122] After the third batch normalization and ReLU activation, a high-dimensional abstract feature map with a dimension of 80×128 is output. This feature map fully preserves the nonlinear rupture mode in the temporal evolution information.
[0123] By using cascaded extraction operations of multiple one-dimensional convolutional kernels, the original low-dimensional time-series data is transformed into a high-dimensional abstract feature map rich in semantic information, realizing a deep feature representation of the broken state of Dendrobium cells, laying a data foundation for subsequent global pooling and regression prediction.
[0124] For example, the input tensor is an 80×2 matrix. The first convolutional layer has a kernel size of 7×1×2 and a number of 32, outputting an 80×32 feature map; the second convolutional layer has a kernel size of 5×1×32 and a number of 64, outputting an 80×64 feature map; the third convolutional layer has a kernel size of 3×1×64 and a number of 128, outputting an 80×128 feature map. Each convolutional layer is followed by a BatchNorm layer and a ReLU layer. Testing shows that this structure can effectively capture energy mutation points caused by cell wall rupture in the 5-8kHz frequency band, significantly improving the signal-to-noise ratio of the feature map. It can distinguish subtle spectral morphological differences between cell wall rupture rates of 85% and 95%, ensuring that subsequent estimation accuracy meets the ±3% requirement.
[0125] S3.3: Based on high-dimensional abstract feature maps, batch normalization and non-linear activation processing are performed, and spatial dimension compression is performed through a global average pooling layer to generate a cell wall breakage rate state feature vector with strong generalization ability.
[0126] It receives a high-dimensional abstract feature map output from the preceding convolutional layer, which contains nonlinear distribution information of Dendrobium fiber fracture patterns in the deep latent space.
[0127] Batch normalization is performed on the high-dimensional abstract feature map, and the mean and variance of the current mini-batch data in each channel are calculated to eliminate the training instability caused by internal covariate bias. The normalized features are then subjected to linear transformation.
[0128] The normalized result is input into the ReLU activation function, and all negative values are set to zero through the threshold truncation mechanism, thus preserving the positive excitation signal that characterizes the intensity of cell wall rupture and enhancing the nonlinear expression ability of the model.
[0129] A global average pooling operation is performed on the activated feature map, and the feature values of each channel are arithmetically averaged along the time series dimension to compress the multidimensional feature map into a one-dimensional cell wall breakage rate state feature vector.
[0130] Through the above processing method, the high-dimensional abstract features of the previous step are transformed into low-dimensional state vectors with strong generalization ability, realizing the dimensionality reduction mapping from complex vibration modes to key wall breaking state indicators, and providing stable input for accurate regression prediction of subsequent fully connected layers.
[0131] S3.4: The fully connected layer performs a nonlinear mapping transformation on the cell wall breakage rate state feature vector, projecting the high-dimensional hidden space features onto the physical dimension space to output an online estimate of the initial current cell wall breakage rate without filtering.
[0132] Receive the cell breakage rate state feature vector output by the global average pooling layer. This vector contains implicit information about the cell breakage pattern of Dendrobium officinale after high-dimensional abstraction.
[0133] The cell wall breakage rate state feature vector is input into the first fully connected layer, and a linear transformation operation is performed using a preset weight matrix to map the high-dimensional hidden features to the intermediate hidden space.
[0134] The ReLU activation function is applied to the intermediate hidden layer data after linear transformation, and nonlinear factors are introduced through threshold truncation to enhance the model's ability to fit complex crushing conditions.
[0135] The activated intermediate layer features are passed to the second fully connected layer, where weighted summation and bias addition operations are performed again to further compress the feature dimensions and focus on the single physical quantity indicator of cell wall breakage rate.
[0136] The output value of the second fully connected layer is normalized by using the Sigmoid activation function, and the numerical range is strictly constrained to be between 0 and 1 to match the physical definition of the percentage of cell wall breakage rate.
[0137] Multiply the normalized probability value by 100 to convert it into an online estimate of the initial current cell wall breakage rate in the range of 0 to 100%.
[0138] By using nonlinear mapping transformation of the fully connected layer and normalization processing of the activation function, the high-dimensional abstract features of the previous step are transformed into initial cell wall breakage rate values with clear physical meaning, realizing accurate projection from the hidden feature space to the physical dimension space, and providing direct data support for subsequent filtering and control.
[0139] For example, let the feature vector of cell wall breakage rate be an 8-dimensional vector X, the weight matrix W1 of the first fully connected layer be 64×8, and the bias b1 be a 64-dimensional vector. Calculate Z1 = W1X + b1, and after ReLU activation, obtain A1 = max(0, Z1). The weight W2 of the second fully connected layer is 1×64, and the bias b2 is a scalar. Calculate Z2 = W2A1 + b2. Apply the Sigmoid function. The output p is obtained. If Z2 = 2.197, then p ≈ 0.9, and the initial estimated cell wall breakage rate is 90%. This process ensures that the model output always conforms to the physical boundary of the cell wall breakage rate, eliminates abnormal predictions of negative values or values exceeding 100%, and significantly improves the reliability of the estimation results.
[0140] like Figure 2 As shown, step S4: First-order low-pass filtering is applied to the online estimate of the current cell wall breakage rate to eliminate high-frequency fluctuation interference, and adaptive normalization correction is performed based on the statistical distribution characteristics of previous batch data to generate a stable cell wall breakage rate feedback quantity for closed-loop control decision-making. Specifically, this includes:
[0141] S4.1: Perform first-order low-pass filtering on the current online estimate of cell wall breakage rate, and use a digital filter with a cutoff frequency of 0.5Hz to filter out high-frequency fluctuation components caused by transient interference from mechanical vibration, so as to obtain smooth cell wall breakage rate sequence data with time continuity.
[0142] It should be noted that the stable cell wall breakage rate feedback quantity is the final cell wall breakage rate signal obtained by sequentially performing first-order low-pass filtering and adaptive normalization correction on the online estimated value output from the previous step, and is used for closed-loop control decision-making.
[0143] The system receives the online estimate of the initial current cell wall breakage rate, which is not filtered, as output from step S3.4. This value is input to the digital signal processing unit as a discrete-time series signal.
[0144] The cutoff frequency parameter of the first-order low-pass digital filter is set to 0.5Hz. This frequency threshold is determined based on the main spectral distribution characteristics of the transient interference of mechanical vibration during the Dendrobium pulverization process, aiming to retain the low-frequency components that reflect the slow change trend of the cell wall breakage rate.
[0145] Obtain the initial cell wall breakage rate estimate x(t) at the current sampling time t and the filtered output value y(t-1) at the previous sampling time t-1, and construct the basic data pair for recursive filtering calculation.
[0146] A first-order infinite impulse response filter algorithm is used to perform real-time smoothing calculations based on the following difference equation:
[0147]
[0148] in, This is the output value of the smooth cell wall breakage rate at the current moment. The original estimated input value at the current moment, The filter coefficient is calculated from the ratio of the cutoff frequency to the sampling frequency.
[0149] Based on the system's sampling frequency of 10kHz and cutoff frequency requirement of 0.5Hz, the calculated filter coefficient α is approximately 0.000314. This tiny coefficient ensures that the filter has extremely strong low-frequency pass-through characteristics and high-frequency suppression capabilities.
[0150] The original estimated value at the current moment is multiplied by the filtering coefficient α, and the filtered output value at the previous moment is multiplied by the complementary coefficient (1-α) to achieve weighted fusion of the old and new data.
[0151] The two weighted results are added together to generate the cell wall breakage rate value at the current moment after smoothing. This value effectively eliminates instantaneous jumps caused by local unevenness of materials or high-frequency noise from the sensor.
[0152] Update the internal status register and assign the currently calculated smooth cell wall breakage rate value y(t) to y(t-1) as the historical reference value for recursive calculation in the next sampling period.
[0153] By continuously executing the above recursive operation, a smooth cell wall breakage rate sequence data with time continuity is formed. This sequence removes high-frequency fluctuation interference and truly reflects the macroscopic evolution trend of the degree of cell wall rupture during the pulverization process.
[0154] By using a first-order low-pass filter, the initial estimate containing high-frequency noise from the previous step is transformed into a smooth cell wall breakage rate sequence with a high signal-to-noise ratio, thus achieving the expected technical effect of providing a stable data foundation for subsequent adaptive normalization correction.
[0155] For example, in the initial stage of the Dendrobium ultrafine grinding operation, due to the impact of material entering the gap between the grinding discs, the initial estimated cell wall breakage rate output by the CNN model fluctuated wildly between 82.5% and 87.3%. The system was set with a sampling period of 0.1 seconds and a cutoff frequency of 0.5 Hz, and the calculated filtering coefficient α was 0.000314. When the input value x(t) at time t was 87.3% and the filtered value y(t-1) at the previous time was 82.6%, substituting into the formula, we calculated: y(t) = 0.000314 × 87.3 + (1 - 0.000314) × 82.6 ≈ 82.6015%. After iterative filtering for 50 consecutive sampling periods, the output value stabilized and converged to around 82.8%, with the fluctuation range decreasing from 4.8% to less than 0.2%, significantly improving data stability and providing a reliable basis for subsequent control decisions.
[0156] S4.2: Based on the smooth cell wall breakage rate sequence data, extract the statistical characteristics of the first two minutes of the current batch, calculate the mean and variance parameters of the data in this period, and generate a real-time statistical distribution vector that characterizes the basic properties of the material in the current batch.
[0157] The system receives smoothed cell wall breakage rate sequence data after first-order low-pass filtering, sets the starting point of the time window as the feeding time of the current batch, and extracts the continuous estimated cell wall breakage rate values from that time to the next 120 seconds to form an initial statistical sample set.
[0158] Outlier removal preprocessing was performed on the initial statistical sample set. The three-standard-deviation criterion was used to identify and remove outlier data points caused by mechanical transient shocks, ensuring the purity and representativeness of the statistical basis.
[0159] Based on the cleaned sample set, the arithmetic mean algorithm is used to calculate the ratio of the sum of all data points to the total number of samples, thereby obtaining the mean parameter that characterizes the overall crushing difficulty of the current batch of Dendrobium materials.
[0160] Simultaneously calculate the sum of squares of the differences between each data point in the sample set and the mean parameter, divide by the degrees of freedom, and take the square root to obtain the variance parameter characterizing the dispersion of the cell wall breakage rate fluctuation.
[0161] The mean and variance parameters are encapsulated into a two-dimensional vector structure to generate a real-time statistical distribution vector, which serves as a digital fingerprint characterizing the basic properties of the current batch of materials.
[0162] By using the above statistical feature extraction and processing method, the time series data from the previous step is transformed into a statistical distribution vector reflecting the physical properties of the material, thereby realizing the quantitative characterization of the differences between different batches of raw materials and providing a benchmark for subsequent adaptive normalization correction.
[0163] For example, smoothed cell wall breakage rate data with a sampling frequency of 20Hz were extracted from the first 120 seconds of the current batch, totaling 2400 data points. After removing 15 outliers exceeding the mean ± 3 standard deviations, 2385 valid samples remained. The calculated mean parameter was 87.5%, and the variance parameter was 2.1%. A real-time statistical distribution vector [87.5, 2.1] was constructed. This vector accurately reflects the increased cell wall breakage resistance caused by the slightly higher moisture content of this batch of Dendrobium, enabling subsequent control steps to adjust the gap strategy accordingly, significantly improving the model's adaptability to material changes.
[0164] S4.3: Call the historical benchmark database pre-stored in the embedded industrial control computer, retrieve the standard mean and standard variance parameters of the previous normal production batches, and compare and analyze the real-time statistical distribution vector with the standard mean and standard variance parameters to construct an adaptive normalization correction coefficient for correcting batch differences.
[0165] The historical benchmark database is read from the non-volatile memory of the embedded industrial computer. This database records the statistical distribution data of vibration energy spectrum characteristics of at least twenty previous production batches that have passed offline microscopic counting.
[0166] The data in the historical benchmark database is filtered for validity, and abnormal batch records caused by sensor failure or material blockage are removed. Normal production batch data with a cell wall breakage rate that is stable within the target range of 90% to 95% are retained as the standard reference set.
[0167] The arithmetic mean of the energy spectrum characteristics of all batches in the standard reference set within the same time window is calculated and defined as the standard mean parameter, which characterizes the baseline level of vibration energy under ideal working conditions.
[0168] The standard variance parameter is obtained by calculating the dispersion of the characteristic values of the vibration energy spectrum of all batches in the standard reference set relative to the standard mean parameter, which characterizes the inherent fluctuation range of the vibration signal during normal crushing.
[0169] Obtain the real-time statistical distribution vector of the first two minutes of the current running batch generated in step S4.2. This vector contains the real-time mean and real-time variance of the vibration energy spectrum feature values of the current batch.
[0170] The difference between the real-time mean and the standard mean parameter is calculated to obtain the mean offset. This offset reflects the baseline drift of the overall vibration energy caused by the difference in moisture content or fiber hardness of the current batch of Dendrobium raw materials.
[0171] The variance scaling factor is obtained by calculating the ratio between the real-time variance and the standard variance parameter. This factor reflects the proportion of change in the amplitude of vibration signal fluctuation during the crushing process of the current batch of material.
[0172] Based on the mean offset and variance scaling factor, a linear transformation model is constructed, and an adaptive normalization correction coefficient is calculated. This coefficient includes translation and scaling components, which are used to map the vibration characteristic distribution of the current batch to the standard reference distribution space.
[0173] By using the above processing method, the real-time statistical distribution vector of the previous step is transformed into an adaptive normalized correction coefficient containing translation and scaling information, thereby eliminating the vibration energy baseline drift caused by the difference between raw material batches, ensuring the consistency of the soft measurement model input data distribution, and improving the generalization ability of cell wall breakage rate estimation across different material batches.
[0174] For example, the historical benchmark database stores standard data from 30 normal batches, of which 25 batches are retained after screening. The standard mean parameter of the 5-8kHz frequency band energy spectrum characteristics is calculated to be 120.5 mV² / Hz, and the standard variance parameter is 15.2 mV² / Hz. The real-time mean of the current batch in the first two minutes is 135.0 mV² / Hz, and the real-time variance is 18.0 mV² / Hz. The calculated mean offset is 14.5 mV² / Hz, and the variance scaling factor is 1.18. In the constructed adaptive normalization correction coefficient, the translation component is -14.5, and the scaling component is 0.85. This correction coefficient is used to correct the vibration energy characteristics of subsequent acquisitions in real time, ensuring that the feature distribution center and dispersion of the current batch are consistent with the standard reference set. This effectively offsets the overall increase in vibration energy caused by the high initial moisture content of the Dendrobium officinale raw material, ensuring the stability of the CNN model input.
[0175] S4.4: The adaptive normalization correction coefficient is used to perform linear transformation on the smooth cell wall breakage rate sequence data. By adjusting the center position and dispersion of the data distribution, the vibration energy baseline drift caused by the different initial moisture content of Dendrobium raw materials is offset, so as to output the standardized cell wall breakage rate value after batch adaptive correction.
[0176] It receives smooth cell wall breakage rate sequence data generated by the previous steps, real-time statistical distribution vector (including mean and variance) of the first two minutes of the current batch, and standard mean and standard variance parameters retrieved from the historical benchmark database.
[0177] A linear transformation mapping relationship based on Z-Score standardization is constructed, and the mean parameter in the real-time statistical distribution vector is used to center the smooth cell wall breakage rate sequence data to eliminate the overall offset of vibration energy baseline caused by the difference in initial moisture content of Dendrobium raw materials.
[0178] The standardized cell wall breakage rate after adaptive normalization correction is calculated using the following formula. :
[0179]
[0180] in, To smooth the instantaneous values in the cell wall breakage rate sequence data, and These are the mean and variance parameters of the data from the first two minutes of the current batch. and These are the standard deviation and standard mean parameters of the preceding normal production batches.
[0181] By introducing the standard mean and standard variance parameters from the historical benchmark database, inverse scaling reconstruction is performed on the standardized intermediate variables, and the data distribution is remapped to the physical dimension space under standard operating conditions.
[0182] Through the above linear transformation process, the smoothed cell wall breakage rate sequence data of the previous step is transformed into a standardized cell wall breakage rate value that offsets the differences in characteristics between batches of materials, so as to achieve the comparability and control stability of cell wall breakage rates of raw materials with different moisture contents under a unified benchmark.
[0183] For example, the average smooth cell wall disruption rate for the first two minutes of the current batch is set to 82.5%, with a variance of 1.2%; the standard mean in the historical benchmark database is 90.0%, with a standard deviation of 1.0%. When the instantaneous smooth cell wall disruption rate is collected as 83.7%, P is directly calculated by substituting it into the formula. std =((83.7-82.5) / 1.2)×1.0+90.0=91.0%. This treatment effectively eliminated the impact of the vibration energy baseline shift caused by the high moisture content of the raw materials, bringing the estimated value back to 91.0% corresponding to the standard operating condition, and significantly improving the generalization accuracy of the closed-loop control command across different batches of materials.
[0184] S4.5: Perform validity boundary determination and final encapsulation processing on the standardized cell wall breakage rate value, eliminate abnormal jump points that exceed the physical possibility range, and confirm that the value is in the effective range of zero to 100%, so as to generate a stable cell wall breakage rate feedback quantity for closed-loop control decision.
[0185] Step S5: Compare the stable cell wall breakage rate feedback with the first preset threshold (lower limit) and the second preset threshold (upper limit). If the value is lower than the first threshold, a gap reduction command is generated; if it is higher than the second threshold, a gap increase command is generated; if it is between the two, a zero-position hold command is generated. Based on the comparison result, the dynamic adjustment target value of the axial position of the small end of the conical grinding disc is output. Specifically, this includes:
[0186] S5.1: Obtain the first-order low-pass filter cell wall breakage rate estimate and the adaptive normalized cell wall breakage rate benchmark value output from the previous steps, and input the first-order low-pass filter cell wall breakage rate estimate into the segmented threshold comparator to complete the preliminary determination of the cell wall breakage rate status under the current crushing condition.
[0187] The system receives the estimated cell wall breakage rate after processing by the first-order low-pass filter in step S4. This value has eliminated high-frequency mechanical vibration noise and undergone batch adaptive normalization correction, resulting in a high signal-to-noise ratio and consistency with operating conditions.
[0188] The estimated cell wall breaking rate of the first-order low-pass filter is loaded into the input port of the segmented threshold comparator of the embedded controller as a real-time status monitoring variable under the current crushing condition.
[0189] Both the first preset threshold and the second preset threshold are percentage thresholds of the cell wall breakage rate, corresponding to the lower limit and upper limit of the target cell wall breakage rate control range, respectively.
[0190] The first cell wall breakage rate lower limit threshold parameter, which is pre-stored in the controller's non-volatile memory, is invoked. This parameter is set to 85% and is used to define the critical boundary of the under-pulverized state.
[0191] The second upper limit threshold parameter of cell wall breaking rate, which is pre-stored in the controller's non-volatile memory, is invoked. This parameter is set to 95% and is used to define the critical boundary of the over-pulverized state.
[0192] Perform numerical comparison logic operations to determine whether the current input first-order low-pass filter cell wall breaking rate estimate is less than the first cell wall breaking rate lower limit threshold.
[0193] If the judgment result is valid, i.e. the estimated value is less than 85%, an under-grinding status indicator signal is generated, indicating that the current grinding gap is too large, resulting in insufficient shear stress.
[0194] Perform numerical comparison logic operations to determine whether the current input first-order low-pass filter cell wall breaking rate estimate is greater than the second cell wall breaking rate upper limit threshold.
[0195] If the judgment result is valid, i.e. the estimated value is higher than 95%, an over-grinding status indicator signal is generated, indicating that the current grinding gap is too small, which increases the risk of frictional temperature rise.
[0196] If neither of the above two judgment conditions is met, that is, the estimated value is within the range of 85% to 95%, then a normal maintenance state indicator signal is generated, indicating that the current working condition meets the target requirements.
[0197] By using a segmented threshold comparison processing method, the continuous cell wall breakage rate value output in the previous step is transformed into a discrete working condition status indicator signal, thereby achieving a preliminary qualitative judgment on the degree of Dendrobium pulverization and providing a logical basis for subsequent differentiated gap adjustment strategies.
[0198] For example, a first lower threshold for cell wall disruption rate is set to 85%, and a second upper threshold for cell wall disruption rate is set to 95%. When the input first-order low-pass filtered cell wall disruption rate estimate is 82.4%, the comparator determines that 82.4 < 85 and outputs an under-disruption status indicator. When the input value is 96.1%, the comparator determines that 96.1 > 95 and outputs an over-disruption status indicator. When the input value is 91.5%, the comparator determines that 85 ≤ 91.5 ≤ 95 and outputs a normal maintenance status indicator. This process is completed within 1 millisecond, ensuring real-time control response, effectively distinguishing different degrees of disruption, and avoiding malfunctions.
[0199] S5.2: Based on the preliminary judgment result of the cell wall breakage rate status, the preset first cell wall breakage rate lower limit threshold and the second cell wall breakage rate upper limit threshold are called to perform dual boundary verification. If the cell wall breakage rate estimate of the first-order low-pass filter is lower than the first cell wall breakage rate lower limit threshold, it is marked as under-pulverized. If it is higher than the second cell wall breakage rate upper limit threshold, it is marked as over-pulverized, so as to generate a cell wall breakage rate deviation discrimination signal containing the status identifier.
[0200] The system receives the stable cell wall breakage rate feedback output from step S4.5. This value is a standardized estimate of the cell wall breakage rate after low-pass filtering and adaptive normalization correction, serving as the core state parameter under the current pulverization condition. It then calls the first lower limit threshold parameter for cell wall breakage rate, pre-stored in the embedded controller's memory. This threshold is set to 85% and is used to define the critical boundary of the under-pulverized state. A numerical comparison is performed between the stable cell wall breakage rate feedback and the first lower limit threshold. If the feedback is strictly less than 85%, it is determined that the current degree of dendrobium cell wall breakage is insufficient, generating an under-pulverized state indicator signal. Next, the system calls the second upper limit threshold parameter for cell wall breakage rate, pre-stored in the embedded controller's memory. This threshold is set to 95% and is used to define the critical boundary of the over-pulverized state. A numerical comparison is performed between the stable cell wall breakage rate feedback and the second upper limit threshold. If the feedback is strictly greater than 95%, it is determined that the current pulverization intensity is too high, posing a risk of degradation of heat-sensitive components, and generating an over-pulverized state indicator signal. If the stable cell wall breakage rate feedback is between the first lower threshold and the second upper threshold, i.e., within the 85% to 95% range, then the current operating condition is determined to be within the target control domain, and a normal maintenance state indicator signal is generated. Based on the above comparison results, a cell wall breakage rate deviation discrimination signal containing state type codes is constructed, where under-grinding state is encoded as -1, over-grinding state is encoded as 1, and normal maintenance state is encoded as 0. Through dual boundary verification logic processing, continuous cell wall breakage rate values are converted into discrete state control command indicators, realizing the quantitative grading judgment of grinding accuracy and providing a clear directional basis for subsequent gap adjustment.
[0201] For example, a first lower threshold for cell wall breakage rate is set to 85%, and a second upper threshold for cell wall breakage rate is set to 95%. When the stable cell wall breakage rate feedback after real-time acquisition and processing is 82.4%, the system executes the comparison logic: 82.4 < 85 is true, indicating an under-crushing state, outputting a status flag of -1, and triggering a subsequent gap reduction command to enhance shear force. When the stable cell wall breakage rate feedback is 96.1%, the system executes the comparison logic: 96.1 > 95 is true, indicating an over-crushing state, outputting a status flag of 1, and triggering a subsequent gap increase command to reduce temperature rise. When the stable cell wall breakage rate feedback is 92.5%, the system executes the comparison logic: 85 ≤ 92.5 ≤ 95 is true, indicating a normal maintenance state, outputting a status flag of 0, and maintaining the current gap unchanged. This process ensures the determinism and real-time performance of control decisions and effectively avoids control oscillations caused by measurement noise.
[0202] S5.3: For the under-grinding state, a decreasing operation is performed according to the preset gap reduction step size parameter. The current real-time gap value is subtracted from the gap reduction step size parameter to calculate the gap reduction control target value used to enhance shear stress.
[0203] Receives the under-grinding status indicator and the smoothed real-time gap value output from S5.2. Calls the gap reduction step size parameter pre-stored in the controller memory; this parameter is set to 0.01mm to avoid transient vibration interference caused by mechanical impact through fine-tuning. Performs a subtraction operation, using the current real-time gap value as the minuend and the preset gap reduction step size parameter as the subtrahend.
[0204] The formula for calculating the target value of the gap adjustment is unified as follows:
[0205]
[0206] in, To reduce the control target value for the gap, This is the current real-time gap value. This is the preset gap reduction step size parameter. In the under-grinding state, use "-" to perform the reduction operation; in the over-grinding state, use "+" to perform the increase operation.
[0207] The calculated clearance reduction control target value is subjected to physical boundary verification to determine whether it is less than the minimum safe clearance threshold allowed by the conical grinding disc. If the verification passes, the target value is encapsulated into a control command data packet containing an axial displacement direction identifier and the specific displacement amount. If the verification fails, the clearance reduction control target value is forcibly clamped to the minimum safe clearance threshold to prevent rigid collision between the tool and the grinding disc.
[0208] Through the above-mentioned decreasing calculation and boundary verification processing, the under-crushing state judgment result of the previous step is transformed into a specific mechanical displacement control target value, so as to achieve the expected technical effect of enhancing shear stress to promote cell wall rupture.
[0209] For example, the current real-time gap value G current The preset gap reduction step size parameter ΔG is 0.01mm, with a value of 0.35mm. Substituting these values into the formula, the target value G for gap reduction control is calculated. target The minimum safe gap threshold is 0.34mm. The system retrieves the minimum safe gap threshold as 0.05mm, and determines that 0.34mm is greater than 0.05mm, thus the verification passes. The controller generates a displacement command pointing towards the small end of the grinding disc, driving the servo motor to perform a 0.01mm feed motion. This operation significantly enhances the shear stress field intensity in the grinding area, causing incompletely broken Dendrobium fibers to break in subsequent rotation cycles, effectively improving cell wall disruption efficiency, while avoiding severe equipment vibration caused by excessive step size.
[0210] S5.4: For the over-grinding state, perform incremental calculations based on the preset gap increase step size parameter, add the current real-time gap value to the gap increase step size parameter, and calculate the gap increase control target value to reduce friction temperature rise.
[0211] Receive the over-grinding status flag and the current real-time gap value generated in step S5.2, and call the pre-stored gap increase step size parameter as the control increment reference.
[0212] Based on the thermosensitive properties of Dendrobium polysaccharides and the mechanism of frictional heat generation, the gap increase step size was set to 0.01 mm. This value has been experimentally verified to effectively reduce shear heat while avoiding drastic fluctuations in cell wall breakage rate.
[0213] Read the current smooth real-time gap value output by the micro-gap capacitive sensor and use it as the base addend in the increment operation.
[0214] The addition arithmetic logic unit operation is executed to accumulate the current real-time gap value with the preset gap increase step size of 0.01mm.
[0215] The calculated clearance increase control target value is subjected to physical boundary constraint checks to ensure that it does not exceed the maximum safe clearance limit of 2.0 mm allowed by the conical grinding disc mechanical structure.
[0216] If the calculation result exceeds the maximum safe clearance limit, the clearance will be increased and the control target value will be forcibly clamped to 2.0mm to prevent the tool from leaving the effective grinding area, which could cause the equipment to run idle or vibrate abnormally.
[0217] If the calculation result is within a safe range, the value is retained as the final control command parameter.
[0218] Through the above incremental calculations and boundary constraint processing, the over-grinding state judgment in the previous step is transformed into a specific target value for increasing the gap, thereby achieving the expected technical effect of reducing the frictional temperature rise in the grinding zone and protecting the heat-sensitive polysaccharide components of Dendrobium from degradation.
[0219] For example, when the system detects a stable cell wall breakage rate feedback of 96.5%, which is higher than the second cell wall breakage rate upper limit threshold of 95%, it determines that the grinding is in an over-grinding state. At this time, the current real-time gap value is read as 0.15mm. The preset gap increase step size parameter of 0.01mm is called. An addition operation is performed: 0.15mm + 0.01mm = 0.16mm. It is checked whether 0.16mm is less than the maximum safe gap of 2.0mm, and the result is yes. Therefore, the gap increase control target value is determined to be 0.16mm. This target value is encapsulated in a dynamic adjustment target value data packet and sent to the servo motor drive module. After execution, the grinding disc gap widens, the shear stress decreases, the heat generated by friction is significantly reduced, the temperature of Dendrobium powder is controlled below 45℃, effectively avoiding polysaccharide degradation, and the cell wall breakage rate gradually falls back to the target range of 90%-95%.
[0220] S5.5: Combines the target value for reducing backlash, the target value for increasing backlash, and the zero-position holding command to maintain the current state, constructs a dynamic adjustment target value data package containing the axial displacement direction and displacement amount, and outputs it to the servo motor drive module to perform the final tool position adjustment.
[0221] It receives the gap reduction control target value, the gap increase control target value, and the zero-position holding command to maintain the current state generated by the previous steps, as the basic input data for dynamic adjustment decision.
[0222] Construct a data structure containing a displacement direction identifier and an absolute displacement parameter. The displacement direction identifier is used to distinguish the positive and negative logic of axial feed, and the absolute displacement parameter represents the specific distance the tool moves relative to the small end of the conical grinding disc.
[0223] Based on the preset control priority logic, the three status commands are mutually exclusive to ensure that only one adjustment command is effective in the same control cycle, thus preventing the servo motor from oscillating or malfunctioning due to signal conflicts.
[0224] The verified displacement direction identifier is mapped to the pulse direction level signal of the servo driver. A high level represents the direction of gap reduction, a low level represents the direction of gap increase, and a zero level represents the state of remaining stationary.
[0225] The absolute displacement parameter is converted into the corresponding pulse equivalent number. Combined with the electronic gear ratio setting of the servo motor, the total number of precise stepping pulses is calculated to quantify the linear displacement accuracy of the tool.
[0226] The encapsulated displacement direction level signal and the total number of step pulses form a standard dynamic adjustment target value data packet. This data packet conforms to the communication protocol format of the embedded motion control card and includes a start bit, data payload, and checksum.
[0227] Through the above data packet construction and encapsulation process, the logical judgment result of the previous step is transformed into a digital control command stream that can be recognized by the servo motor, thereby achieving the expected technical effect of precise command transmission from cell wall breaking rate estimation to mechanical actuator.
[0228] For example, when the stable cell wall breakage rate feedback is 82%, it is determined to be in an under-crushing state, and the system generates a gap reduction command. The gap reduction step size is set to 0.01mm, the servo motor electronic gear ratio is 10000 pulses / revolution, and the ball screw lead is 5mm. The calculated number of pulses corresponding to a single step displacement is 20 pulses. The displacement direction identifier is set to 1 (high level). The dynamic adjustment target value data packet contains direction bit 1 and pulse count 20. After this data packet is sent to the motion control card, the servo motor precisely rotates the corresponding angle, pushing the cutter axially by 0.01mm, significantly increasing the shear stress, and causing the cell wall breakage rate to rise back to the target range.
[0229] like Figure 3 As shown, step S6: Based on the dynamically adjusted target value, the servo motor is driven to perform displacement, thereby changing the axial position of the cutter at the small end of the conical grinding disc and adjusting the grinding gap. This, in turn, achieves real-time closed-loop control of the Dendrobium cell wall breakage rate by changing the shear stress distribution. Specifically, this includes:
[0230] S6.1: Based on the dynamically adjusted target value, the displacement direction indicator and the number of step pulses are parsed out. The motion control card is used to generate a pulse sequence signal containing acceleration and deceleration curves to obtain a digital control command stream for driving the servo motor.
[0231] Receive the dynamic adjustment target value data packet generated by the previous steps, and parse the displacement direction identifier and target displacement value contained therein.
[0232] If the direction indicator is positive, it is determined that the grinding gap needs to be increased; if it is negative, it is determined that the grinding gap needs to be decreased. Based on the preset servo motor electronic gear ratio and ball screw lead parameters, the target displacement is converted into the corresponding total number of pulse equivalents. A trapezoidal acceleration / deceleration programming algorithm is used to set the maximum operating speed, acceleration, and deceleration parameters, constructing a smooth speed-time curve.
[0233] Based on the difference between the current real-time position of the motor and the target position, the required pulse quantity distribution for each of the acceleration, constant speed, and deceleration phases is calculated. The motion control card's internal timer generates high-frequency pulse sequences with microsecond-level precision, gradually increasing the pulse frequency during acceleration, maintaining a constant frequency during constant speed, and gradually decreasing the pulse frequency during deceleration. Orthogonal pulse signals and directional level signals are output through a differential signal driver to ensure signal anti-interference capability.
[0234] By using the aforementioned digital instruction stream generation method, the macroscopic clearance adjustment requirements are transformed into a microscopic pulse sequence that can be executed by the servo system, thereby achieving shock-free, high-precision step-by-step adjustment of the tool position.
[0235] For example, setting the servo motor resolution to 10,000 pulses / revolution and the ball screw lead to 5mm, a single pulse corresponds to a displacement of 0.0005mm. When a gap reduction of 0.01mm is required, it is calculated that 20 pulses need to be sent. The maximum speed is set to 1000 pulses / second, and the acceleration to 500 pulses / second². The motion control card generates a trapezoidal pulse train that first accelerates and then decelerates, with a total time of approximately 0.14 seconds. This process avoids mechanical resonance caused by step signals, ensuring smooth and precise gap adjustment.
[0236] S6.2: Perform power amplification and impedance matching processing on the digital control command stream, and use the servo driver to convert the low-level logic signal into a high-torque three-phase AC signal to generate an electromagnetic driving torque that can drive the rotor of the servo motor to rotate.
[0237] The system receives a digital control command stream from the motion control card, which contains low-level logic signals representing the displacement direction and the number of step pulses. Level conversion is performed on the low-level logic signals, using an optocoupler isolation circuit to convert the TTL level to a differential signal to eliminate common-mode interference during long-distance transmission and achieve electrical isolation.
[0238] The isolated differential signal is input to the pulse receiving port of the servo driver. The control unit inside the driver analyzes the pulse frequency and direction mark to generate the corresponding three-phase sinusoidal modulation signal reference value.
[0239] Space vector pulse width modulation (SVM) technology is used to decompose the reference voltage vector and calculate the conduction time and duty cycle of each bridge arm switch to synthesize the required phase voltage waveform.
[0240] Using an insulated-gate bipolar transistor (IGBT) power module, the DC current is chopped at high frequency according to the calculated duty cycle to generate an equivalent three-phase AC signal. Current closed-loop feedback regulation is then applied to the generated three-phase AC signal, and the motor stator current is acquired in real time and compared with the command current to correct the PWM waveform and compensate for nonlinear load characteristics.
[0241] Through power amplification and impedance matching, low-level logic instructions are converted into high-torque three-phase AC signals, generating electromagnetic drive torque capable of driving the rotor of a servo motor.
[0242] For example, the servo driver receives a pulse sequence with a frequency of 50kHz, and the DC bus voltage is set to 310V. When the modulation ratio is set to 0.8, the phase voltage vector amplitude is 248V. The effective value of the driver's output three-phase current is stable at 5A, and the phase error is controlled within 2 degrees. The motor rotor generates a holding torque of 1.5N·m, with a response delay of less than 1ms, ensuring the immediacy and accuracy of the tool axial displacement control and effectively eliminating the hysteresis effect caused by mechanical transmission backlash.
[0243] S6.3: Based on the electromagnetic drive torque to drive the servo motor to perform precise angular displacement operation, the built-in high-resolution encoder provides real-time feedback of the rotor's current position information to form a position closed loop, so as to obtain accurate motor shaft rotation data that eliminates accumulated errors.
[0244] It receives the three-phase AC signal output from the servo driver, establishes a rotating magnetic field in the stator winding, and generates an electromagnetic driving torque to drive the rotor to rotate.
[0245] Based on the electromagnetic driving torque to overcome the load resistance torque and friction torque, the servo motor rotor is driven to perform precise angular displacement motion, changing the mechanical angular position of the rotor relative to the stator.
[0246] By using a built-in high-resolution photoelectric encoder installed at the tail of the motor, the absolute position or relative incremental change of the rotor magnetic poles is detected in real time, and the corresponding high-frequency pulse signal or digital coded data is generated.
[0247] The position pulse signal fed back by the encoder is transmitted to the position loop control module of the servo driver, and is compared and calculated at high speed with the target position command pulse issued by the motion control card.
[0248] The position deviation is calculated using a proportional-integral-derivative (PID) control algorithm. The frequency and amplitude of the inverter output current are dynamically adjusted according to the speed correction, forming a closed position negative feedback control loop to compensate for the angle error caused by load fluctuations in real time.
[0249] The steady-state error of the position loop is continuously monitored. When the error converges to within the preset accuracy threshold, the current rotor position is locked, and the accurate rotation angle data of the motor shaft after eliminating the accumulated error is output.
[0250] By using a closed-loop position feedback control method, the electromagnetic drive torque from the previous step is converted into high-precision motor shaft rotation angle data, thereby achieving the expected technical effect of micron-level positioning accuracy for tool gap adjustment.
[0251] For example, an AC servo motor with a rated torque of 5 Nm is selected, paired with an absolute encoder with a resolution of 23 bits (i.e., 8,388,608 pulses per revolution). When a control command to reduce the gap by 0.01 mm is received, the motion control card calculates the corresponding number of pulses to be approximately 16,777. The servo driver receives the pulses and drives the motor to rotate, while the encoder provides real-time feedback on the current position. If the load suddenly increases due to uneven hardness of the Dendrobium officinale material, and the position loop detects a hysteresis error of 5 pulses, the PID controller immediately increases the q-axis current output to enhance the electromagnetic torque and eliminate the error. Ultimately, the motor shaft stops at the target position, with the actual rotation angle error controlled within ±0.001 degrees, ensuring that the linear displacement error converted from the ball screw is less than 0.5 μm, significantly improving the repeatability of the gap adjustment.
[0252] S6.4: The precise rotation angle data of the motor shaft is transmitted to the precision ball screw pair through a high-rigidity coupling. The rotational motion is converted into linear feed motion by using the helical transmission principle to obtain the linear displacement that drives the tool assembly to move axially.
[0253] It receives precise rotation angle data of the motor shaft from the servo motor encoder, which represents the amount of rotational displacement at the drive source.
[0254] The precise rotation angle data of the motor shaft is transmitted to the input end of the precision ball screw pair through a high-rigidity coupling. The zero backlash characteristic of the coupling is used to eliminate transmission backlash and ensure that the rotational torque is transmitted to the screw spindle without loss.
[0255] The nut assembly inside the precision ball screw pair, driven by the rotation of the screw, converts the rotational motion into linear feed motion along the axial direction based on the principle of helical transmission, thereby calculating the linear displacement of the tool assembly.
[0256] Based on the calculated linear displacement, the tool assembly, which is fixed to the nut, is driven to produce a precise physical displacement along the axis of the conical grinding disc.
[0257] The actual position of the tool assembly is monitored in real time to confirm that its displacement is consistent with the dynamic adjustment target value, thus completing the precise conversion from rotary motion to linear feed motion.
[0258] By using the principle of helical transmission, the rotational angular displacement of the motor shaft is converted into the axial linear displacement of the tool assembly, thereby obtaining a precise linear displacement that drives the tool assembly to move axially, achieving the expected technical effect of micron-level precise control of the grinding gap.
[0259] For example, the ball screw lead P is set to 5mm, and the reduction ratio i is 1. When the controller issues a command to reduce the backlash, requiring a displacement of 0.01mm, the motor shaft needs to rotate by an angle θ of 0.72 degrees. A high-rigidity coupling transmits this 0.72-degree rotation angle to the screw without backlash, and the rotation of the screw drives the nut and cutter assembly to move axially by 0.01mm. This process eliminates backlash errors in mechanical transmission, ensuring that the gap adjustment accuracy for Dendrobium officinale crushing reaches the micron level, and significantly improving the stability of cell wall breakage rate control.
[0260] S6.5: Based on the change of the physical distance between the cutter end face and the small end of the conical grinding disc by linear displacement, the shear stress field distribution in the material crushing area is reconstructed by the mechanical contact mechanics principle, so as to achieve real-time closed-loop control effect for the cell wall breakage rate of Dendrobium officinale.
[0261] It receives the axial linear displacement of the tool assembly generated by the conversion of the precision ball screw pair, which directly affects the relative positional relationship between the small end of the conical grinding disc and the fixed tool.
[0262] The physical values of the grinding gap are updated in real time based on the linear displacement, and the normal contact pressure distribution of the material under variable gap conditions is calculated using Hertzian contact theory.
[0263] Based on the updated gap value and grinding disc speed, the shear stress field in the pulverizing area is reconstructed. The change in shear stress directly regulates the mechanical load intensity acting on the Dendrobium fibers and cell walls. When the gap decreases, the denominator decreases, resulting in a significant increase in shear stress, which promotes cell wall rupture.
[0264] When the gap increases, the shear stress decreases, reducing the heat accumulation caused by excessive friction and protecting the heat-sensitive polysaccharide components from degradation.
[0265] Real-time monitoring of the impact of shear stress field changes on cell wall disruption efficiency forms a closed-loop regulatory link from mechanical displacement to biophysical effects.
[0266] By altering the physical gap through linear displacement and reconstructing the shear stress field distribution, mechanical control commands are transformed into specific cell wall disruption dynamics conditions, enabling real-time, precise, closed-loop regulation of the Dendrobium cell wall disruption rate.
[0267] For example, the grinding disc rotation speed ω is set to 3000 rad / min, the radial position r to 0.05 m, and the friction coefficient μ to 0.3. The initial gap h is 0.5 mm. When the cell wall breakage rate is detected to be below 85%, a gap reduction operation is performed, with a linear displacement Δh of -0.01 mm. The calculated new gap is 0.49 mm, and the shear stress τ increases from its original value to approximately 1.02 times. The increased shear stress accelerates the breakage of Dendrobium cell walls, causing the cell wall breakage rate to recover to the 90%–95% range in the next control cycle. If the cell wall breakage rate exceeds 95%, a gap increase operation is performed, with Δh of +0.01 mm, resulting in a new gap of 0.51 mm. This reduces the shear stress, prevents the temperature rise from exceeding 45°C, and ensures a significant increase in the retention rate of Dendrobium polysaccharide activity.
[0268] Step S7: Record the vibration spectrum, real-time gap value, and stable cell breakage rate feedback of the current operating cycle to form an operating log; use the actual cell breakage rate measured by offline microscopic counting method after the current batch ends as the calibration value, calculate the deviation between the online estimated value and the calibration value of each sampling cycle within the batch, and take the arithmetic mean of the deviations of all sampling cycles within the batch as the average deviation; when the average deviation of multiple consecutive batches accumulates to exceed a preset threshold, trigger incremental model updates to adjust the weight parameters of the convolutional neural network soft measurement model to optimize its cross-batch generalization ability. Specifically, this includes:
[0269] S7.1: Perform timestamp alignment and formatted encapsulation on the original vibration time series data, real-time gap values, and stable cell wall breakage rate feedback after first-order low-pass filtering within the current operating cycle to generate structured operating log entries containing multimodal sensing information and control decision results.
[0270] S7.2: Based on the stable cell wall breakage rate feedback extracted from the generated structured operation log entries, and combined with the corresponding offline microscopic counting method real cell wall breakage rate calibration value obtained from the preset sampling batch node, perform difference operation and absolute value calculation to generate a single-point deviation index that characterizes the instantaneous estimation accuracy of the deep convolutional neural network soft measurement model.
[0271] Extract the stable cell wall breakage rate feedback sequence of the current batch from the structured operation log entries, and use it as the benchmark data for evaluating the instantaneous estimation accuracy of the model.
[0272] Based on the preset sampling batch node identifier, the offline microscopic counting method true cell wall breakage calibration value corresponding to the batch is retrieved and matched to establish the physical true value reference point for error calculation.
[0273] Perform point-by-point interpolation to subtract the actual cell breakage rate calibration value from the stable cell breakage rate feedback value, and calculate the original deviation data that characterizes the degree of deviation between the estimated value and the true value.
[0274] Calculate the absolute value of the single-point deviation index, verify the validity of the calculated absolute value, eliminate non-physical extreme values caused by sensor transient failures, and ensure that the deviation index reflects the true model estimation performance.
[0275] Through the above difference operation and absolute value conversion, the multi-source heterogeneous online estimation data and offline calibration data are transformed into a single-point deviation index with unified dimensions, realizing the quantitative characterization of the instantaneous estimation accuracy of the deep convolutional neural network soft measurement model.
[0276] For example, the stable cell wall breakage rate feedback during the pulverization process of the 5th batch of Dendrobium officinale was extracted, with an average value of 92.5%. The actual cell wall breakage rate calibration value of this batch, measured by microscopic counting method after the machine was stopped, was obtained as 91.8%. The deviation was calculated as: E = |92.5 - 91.8| = 0.7%. This single-point deviation index of 0.7% was recorded in the deviation accumulation queue for subsequent sliding window statistical analysis. If a batch experiences a fluctuation in the vibration baseline due to a sudden change in the moisture content of the raw materials, with an estimated value of 96.0% and an actual value of 92.0%, the calculated deviation index is 4.0%, significantly higher than the normal fluctuation range, thus triggering the subsequent model generalization performance degradation early warning mechanism.
[0277] S7.3: Perform sliding window statistical analysis and cumulative error summation on the single-point deviation index generated from multiple consecutive production batches to generate a cumulative quantitative value of deviation that reflects the trend of the model’s generalization performance degradation across different material batches.
[0278] Obtain the sequence of single-point deviation indicators for multiple consecutive production batches generated by the previous steps, and store them in a sliding window buffer. The length of the sliding window is set to cover the data volume of the most recent 10 complete production batches.
[0279] The single-point deviation index within the sliding window is subjected to time-weighted processing, and the weight coefficients are assigned according to the order in which the batches were generated. The weight coefficients of recent batches are higher than those of distant batches, so as to enhance the model's sensitivity to changes in current material properties.
[0280] The weighted cumulative deviation value is calculated using the exponentially weighted moving average algorithm and used as the quantified value of cumulative deviation.
[0281] Trend slope analysis was performed on the calculated cumulative deviation quantization value. The deviation change curves of the most recent 5 batches were fitted using the least squares method, and the slope parameter was extracted to characterize the decay rate of the model's generalization performance.
[0282] The cumulative deviation quantization value is combined with the degradation rate parameter to generate a multidimensional performance degradation feature vector, which comprehensively reflects the degree of deviation and deterioration trend of the current soft measurement model in the estimation accuracy under cross-batch operating conditions.
[0283] Through the above sliding window statistical analysis and weighted cumulative calculation, the discrete single-point deviation index is transformed into a quantitative value of cumulative deviation that reflects the decline trend of model generalization performance. This enables a quantitative assessment of the model's health status and provides an accurate criterion for triggering subsequent incremental updates.
[0284] For example, the sliding window length N_w is set to 10, and the weight coefficients w_i are distributed according to an exponential decay law, with the weight of the most recent batch being 0.15 and the weight of the furthest batch being 0.05. If the single-point deviation indices of the most recent 10 batches are 1.2%, 1.5%, 1.1%, 1.8%, 2.1%, 2.5%, 2.8%, 3.0%, 3.2%, and 3.5%, respectively, the weighted cumulative deviation quantification value E is 2.45%. Simultaneously, the fitted deviation change slope is 0.25% / batch, indicating that the model error shows a linear growth trend. This cumulative deviation quantification value is significantly higher than the preset static threshold, accurately reflecting the decrease in model generalization ability caused by fluctuations in the moisture content of Dendrobium officinale raw materials, ensuring the timeliness of incremental update triggering.
[0285] S7.4: Logically compare the generated cumulative deviation quantization value with the preset allowable error threshold range. If the cumulative deviation quantization value is determined to exceed the allowable error threshold range, generate a model incremental update trigger command to start the online optimization process for the deep convolutional neural network soft measurement model.
[0286] Obtain the cumulative deviation quantization value generated by the preceding step S7.3, which characterizes the degree of estimation accuracy degradation of the deep convolutional neural network soft measurement model in continuous production batches.
[0287] The preset allowable error threshold range is read from the non-volatile memory area of the embedded industrial control computer. This range includes an upper threshold and a lower threshold, which are used to define the acceptable boundaries of the model's generalization performance.
[0288] The cumulative quantized value of the deviation is compared with the upper limit of the allowable error threshold range to determine whether the cumulative error of the current model exceeds the system tolerance limit.
[0289] If the cumulative quantized value of the judgment bias exceeds the upper limit of the allowable error threshold range, it is confirmed that the model's generalization ability has significantly deteriorated, and an online optimization process needs to be executed.
[0290] Generate a high-priority model incremental update trigger instruction, which includes the trigger timestamp, the current deviation value, and the identifier of the model to be updated.
[0291] By using a logical comparison and judgment processing method, the cumulative quantized value of the deviation is transformed into a model state judgment signal, thereby achieving the expected technical effect of starting the adaptive online optimization process for the soft measurement model of the deep convolutional neural network.
[0292] For example, the upper limit of the preset allowable error threshold range is 0.05. When the cumulative quantized value of the deviation calculated by S7.3 is 0.062, the system performs a comparison operation and determines that 0.062 is greater than 0.05. The system then generates a model incremental update trigger instruction, and the instruction data frame contains the trigger type code 0x01 and the current batch ID. This instruction is sent to the model management module to start the fine-tuning program for the weights of the fully connected layers, ensuring that the model restores its high-precision estimation capability in the next batch of production and effectively suppresses estimation drift caused by fluctuations in the moisture content of raw materials.
[0293] S7.5: In response to the generated model incremental update trigger command, the vibration energy spectrum feature vector and real-time gap value in the latest batch of collected structured operation log entries are used as fine-tuning training samples to perform backpropagation iterative update operation on the weight parameters of the fully connected layer of the deep convolutional neural network soft measurement model, so as to output an updated deep convolutional neural network soft measurement model with enhanced generalization ability.
[0294] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.
[0295] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the element or object preceding “comprising” or “including” encompasses the element or object listed following “comprising” or “including” and its equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.
[0296] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A method for intelligent control of cell disruption based on real-time estimation of Dendrobium cell wall disruption rate, characterized in that, include: Wideband vibration acceleration signal and micro-gap capacitor voltage are synchronously acquired in the conical grinding disc bearing housing. The acceleration is sampled to obtain the original vibration time sequence data, and the micro-gap capacitor voltage is calibrated and converted to obtain the real-time gap value. The original vibration time series data is bandpass filtered to retain the 5-8kHz cell wall breaking characteristic frequency band and suppress the interference frequency band. The filtered signal is then subjected to short-time Fourier transform to extract the frequency band energy spectrum feature vector. The frequency band energy spectrum feature vector and the real-time gap value are combined into a continuous time window feature sequence, which is then input into a pre-trained convolutional neural network soft measurement model to output an online estimate of the current cell wall breakage rate. The online estimated value is subjected to first-order low-pass filtering and adaptive normalization correction is performed in combination with the statistical distribution characteristics of the previous batch data to generate a stable cell wall breakage rate feedback. The stable cell wall breakage rate feedback is compared with the first preset threshold for the lower limit of the cell wall breakage rate and the second preset threshold for the upper limit of the cell wall breakage rate. If it is lower than the first threshold, a gap reduction command is generated; if it is higher than the second threshold, a gap increase command is generated; if it is between the two, a zero position hold command is generated. Based on the comparison result, the dynamic adjustment target value of the axial position of the small end of the conical grinding disc is output. Based on the dynamically adjusted target value, the servo motor is driven to perform displacement, thereby changing the axial position of the cutter at the small end of the conical grinding disc and adjusting the grinding gap to achieve real-time closed-loop control of the cell wall breaking rate.
2. The intelligent control method for cell disruption based on real-time estimation of Dendrobium cell wall disruption rate according to claim 1, characterized in that, After adjusting the grinding gap, the method further includes: Record the vibration spectrum, real-time gap value, and stable cell breakage rate feedback of the current operating cycle to form an operating log; The actual cell wall breakage rate measured by offline microscopic counting method after the current batch ends is used as the calibration value. The deviation between the online estimated value and the calibration value of each sampling period in the batch is calculated, and the arithmetic mean of the deviations of all sampling periods in the batch is taken as the average deviation. When the average deviation of multiple consecutive batches accumulates to exceed a preset threshold, an incremental update of the model is triggered to adjust the weight parameters of the convolutional neural network soft measurement model in order to optimize its cross-batch generalization ability.
3. The intelligent control method for cell disruption based on real-time estimation of Dendrobium cell wall disruption rate according to claim 1, characterized in that, The process of synchronously acquiring broadband vibration acceleration signals and micro-gap capacitor voltages in a conical grinding disc bearing housing, sampling the acceleration to obtain raw vibration time-series data, and calibrating and converting the micro-gap capacitor voltages to obtain real-time gap values includes: Based on the radial and axial intersection of the conical grinding disc bearing seat, a broadband vibration sensor and a micro-gap capacitive sensor are fixed using a rigid mounting base to eliminate the phase shift caused by the difference in mechanical vibration transmission paths, thereby obtaining a sensor physical mounting structure with a fixed relative positional relationship. Based on the physical installation structure of the sensor, the triaxial acceleration signal and the gap voltage signal are synchronously acquired using the multi-channel synchronous sampling mode of the data acquisition card to generate the original vibration acceleration sequence and the original gap voltage sequence containing strictly aligned timestamp information. Based on the original gap voltage sequence, a nonlinear mapping conversion process from voltage to distance is performed using a preset sensor calibration curve to obtain the real-time dynamic gap original value that characterizes the actual physical distance between the small end of the grinding disc and the cutting tool. Based on the original value of the real-time dynamic gap, a high-frequency noise suppression process is performed using a moving average filtering algorithm to obtain a smooth real-time gap value after eliminating random interference. Based on the original vibration acceleration sequence and the smoothed real-time gap value, a data fusion encapsulation protocol is used to perform time-series data pairing and formatted encapsulation processing to obtain a dataset of original vibration time-series data and real-time gap value combination containing timestamp alignment information for subsequent feature extraction.
4. The intelligent control method for cell disruption based on real-time estimation of Dendrobium cell wall disruption rate according to claim 1, characterized in that, The step of combining the frequency band energy spectrum feature vector with the real-time gap value into a continuous time window feature sequence, inputting it into a pre-trained convolutional neural network soft measurement model, and outputting an online estimate of the current cell wall breakage rate includes: Based on the frequency band energy spectrum feature vector and the real-time gap value, a multi-period sliding window splicing process is performed to generate a continuous time window feature sequence containing temporal evolution information, which serves as the standardized input tensor of the deep convolutional neural network soft measurement model. The continuous time window feature sequence is input into a pre-constructed deep convolutional neural network soft measurement model, and local feature extraction operations are performed using multi-layer one-dimensional convolutional kernels to obtain a high-dimensional abstract feature map characterizing the fiber fracture mode of Dendrobium officinale. Based on the high-dimensional abstract feature map, batch normalization and nonlinear activation processing are performed, and spatial dimension compression is performed through a global average pooling layer to generate a cell wall breakage rate state feature vector with strong generalization ability. The cell wall breakage rate state feature vector is transformed by a fully connected layer using a nonlinear mapping transformation, which projects the high-dimensional hidden space features to the physical dimension space to output an online estimate of the initial current cell wall breakage rate without filtering.
5. The intelligent control method for cell disruption based on real-time estimation of Dendrobium cell wall disruption rate according to claim 1, characterized in that, The short-time Fourier transform includes: The filtered signal is windowed and truncated by applying a window function. It is then divided into discrete frames by sliding according to a preset overlap rate. Fourier transform is performed on each frame and the power spectral density is obtained by calculating the square of the modulus. The frequency band energy spectrum feature value is obtained by integrating and accumulating the feature values within the 5-8kHz frequency band. A multidimensional feature vector is constructed using the feature values of the current and a preset number of historical periods, which serves as the 5-8kHz frequency band energy spectrum feature vector.
6. The intelligent control method for cell disruption based on real-time estimation of Dendrobium cell wall disruption rate according to claim 1, characterized in that, The cutoff frequency of the first-order low-pass filter is configured to be lower than the lowest frequency of the high-frequency fluctuation component in the online estimation value.
7. The intelligent control method for cell disruption based on real-time estimation of Dendrobium cell wall disruption rate according to claim 1, characterized in that, The adaptive normalization correction includes: The statistical distribution parameters of the smooth cell wall breakage rate sequence data of the previous period of the current batch are extracted, compared with the standard statistical distribution parameters in the pre-stored historical benchmark database, and a correction coefficient is constructed. The smooth cell wall breakage rate sequence data is linearly transformed, and after validity boundary judgment and removal of abnormal jump points, the stable cell wall breakage rate feedback quantity is generated.
8. The intelligent control method for cell disruption based on real-time estimation of Dendrobium cell wall disruption rate according to claim 1, characterized in that, The method for generating the reduced gap command is as follows: The current real-time gap value is subtracted from the preset gap reduction step size parameter to obtain the gap reduction control target value; the gap increase command is generated by adding the current real-time gap value to the preset gap increase step size parameter to obtain the gap increase control target value; the gap reduction control target value, the gap increase control target value, or the zero position holding command are combined to construct the dynamic adjustment target value data packet.
9. The intelligent control method for cell disruption based on real-time estimation of Dendrobium cell wall disruption rate according to claim 1, characterized in that, The deep convolutional neural network soft measurement model comprises a multi-layer one-dimensional convolutional structure. After convolution, it is processed by batch normalization and activation function, and then by global pooling and fully connected layers to output the online estimated value.
10. The intelligent control method for cell disruption based on real-time estimation of Dendrobium cell wall disruption rate according to claim 1, characterized in that, The bandpass filter includes: The original vibration time series data is processed to remove the DC component to eliminate the sensor's zero bias error. Then, a Butterworth digital filter is used to filter the DC-removed vibration signal to retain the cell wall breaking characteristic frequency band and suppress the interference frequency band.
