Intelligent prediction method and system for wet granulation end point based on torque dynamic characteristics

Abstract

The application relates to the technical field of automatic processing control, and discloses a wet granulation endpoint intelligent prediction method and system based on torque dynamic characteristics, which comprises the following steps: collecting spindle motor output torque data and running time stamps, constructing a one-dimensional original torque time sequence, and applying a sliding time window to extract a multi-dimensional rheological characteristic vector sequence; inputting the multi-dimensional rheological characteristic vector sequence into an endpoint prediction algorithm model, outputting a predicted remaining time, performing serial verification, and obtaining a determined predicted remaining time; based on this, a countdown trigger is generated in a hardware timer, and a hardware interrupt signal is output when the countdown trigger is zeroed and a brake instruction is issued to a frequency converter. The application realizes high-fidelity multi-dimensional decoupling of mixed signals, eliminates system response time delay, realizes closed-loop brake regulation with hard real-time feedforward prediction, and effectively avoids excessive material stirring to damage the particle properties.

Description

Technical Field

[0001] This invention relates to the field of automated processing control technology, specifically to a method and system for intelligent prediction of the endpoint of wet granulation based on dynamic torque characteristics. Background Technology

[0002] With the development of automated processing technology, wet granulation equipment typically uses a spindle motor to drive an agitator to mix and granulate materials. In existing processes, the main control unit often employs pre-set static control logic, relying on an independent torque sensor installed at the rear of the equipment to intermittently acquire readings at a low sampling rate, attempting to determine whether granulation has reached its endpoint by comparing the single-point amplitude with a static threshold.

[0003] However, under dynamic and continuous stirring and shearing, the rheological evolution of materials from loose powder to high-viscosity agglomerates instantaneously changes the load state of the spindle motor, resulting in an overlap of global phase transition trends and local transient disturbances in the one-dimensional torque signal. Since static threshold parameters cannot perceive this multi-dimensional rheological evolution, low-frequency single-point sampling is prone to misjudgment at the critical material forming stage, leading to unavoidable physical delays in traditional hysteresis feedback control, causing excessive stirring and damaging particle properties. Furthermore, blindly introducing purely data-driven neural network predictions for breakthroughs can easily result in model outputs being affected by residual random noise, causing non-physical abnormal jumps; and traditional software logic is prone to operating system task scheduling delays when executing feedback shutdown, and the timing misalignment of instruction issuance between hardware and software also prevents the optimal granulation shutdown endpoint from being accurately reproduced by the underlying frequency converter. Therefore, overcoming the effects of torque signal aliasing and system response delays, and achieving closed-loop braking control with high-fidelity decoupling and hard real-time feedforward prediction, is a pressing technical problem that needs to be solved in this field. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the present invention aims to provide an intelligent prediction method and system for the endpoint of wet granulation based on dynamic torque characteristics. This method constructs a one-dimensional original torque time series by collecting the output torque data and running timestamps of the main shaft motor, and extracts a multi-dimensional rheological feature vector sequence by applying a sliding time window. This achieves high-fidelity decoupling of the global phase transition trend and local transient disturbances in the torque signal. The endpoint prediction algorithm model evolves the features into a predicted remaining time. Combining non-negativity verification, upper bound constraint verification, and temporal monotonicity verification, the model output is transformed into a deterministic predicted remaining time with hard real-time execution conditions, accurately eliminating abnormal jumps. A local system hardware timer, combined with a countdown trigger and frequency converter linkage, achieves precise timing alignment, ensuring braking reliability, eliminating physical delays, avoiding excessive material mixing, and protecting particle properties.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A smart prediction method for the endpoint of wet granulation based on dynamic torque characteristics includes: The output torque data and running timestamp of the spindle motor are obtained, and a one-dimensional original torque time series is constructed in chronological order. A sliding time window is applied to the one-dimensional original torque time series to perform decoupling extraction, resulting in a multi-dimensional rheological feature vector sequence. The multidimensional rheological feature vector sequence is input into the pre-trained endpoint prediction algorithm model, the predicted remaining time is output, and a concatenated verification is performed on the predicted remaining time to obtain the deterministic predicted remaining time. The main controller receives the deterministic prediction of the remaining time, generates and dynamically refreshes the countdown trigger in the local system hardware timer, and outputs a hardware interrupt signal when the countdown trigger returns to zero. The main controller responds to the hardware interrupt signal, sends a braking command to the frequency converter through the hardware control interface, and generates a granulation batch endpoint control record.

[0006] Furthermore, the method for constructing the one-dimensional original torque time series includes: The output torque data and corresponding running timestamp of the spindle motor are read in real time through the communication bus of the spindle motor driver at a preset sampling frequency. The output torque data read at each preset sampling time and its corresponding running timestamp are taken as a data pair, and the data pairs are cached sequentially in the circular data buffer inside the main controller according to the time order of the running timestamps. The ordered set of all output torque data arranged in chronological order of runtime timestamps in the circular data buffer is defined as a one-dimensional original torque time series.

[0007] Furthermore, the method for obtaining the multidimensional rheological feature vector sequence includes: A sliding time window with a preset window length and sliding step size is applied to the one-dimensional original torque time series, and a corresponding number of data points are sequentially extracted to form a subset of window data. Within each sliding time window, the arithmetic mean of the output torque data values ​​is calculated to obtain the mean feature, the mean of the absolute difference between adjacent data points is calculated to obtain the first-order difference feature, and the variance feature around the mean feature is calculated. The mean feature, first-order difference feature, and variance feature are concatenated to form a three-dimensional rheological feature vector; Arrange the three-dimensional rheological feature vectors corresponding to all sliding time windows in chronological order to form a multidimensional rheological feature vector sequence.

[0008] Furthermore, the endpoint prediction algorithm model includes: The three-dimensional rheological feature vectors corresponding to the last consecutive sliding time window of a preset number of input time steps are extracted from the multidimensional rheological feature vector sequence to form the historical feature matrix. The arithmetic mean and standard deviation of each feature column of the historical feature matrix are calculated independently. Zero mean and unit variance standardization is performed on each feature element in the column to obtain the standardized input matrix. The standardized input matrix is ​​fed into the input layer of the pre-trained endpoint prediction algorithm model row by row in chronological order.

[0009] Furthermore, the method for obtaining the predicted remaining time includes: The endpoint prediction algorithm model includes a hidden layer and an output layer that are cascaded sequentially with the input layer; Using the row-by-row input process of the standardized input matrix as a time step, at each time step, the hidden layer receives the three-dimensional rheological feature vector corresponding to the current row and the hidden state vector passed from the previous time step, calls the internal door control logic to generate and update the current hidden state vector, and performs recursive operations. After completing the recursive operation of all time steps corresponding to the number of input time steps, the final hidden state vector output by the hidden layer at the last time step is extracted and sent to the output layer. The output layer performs a fully connected linear mapping on the final hidden state vector, and the output is a scalar value predicting the remaining time.

[0010] Furthermore, the method for obtaining the deterministic prediction of the remaining time includes: Obtain the output prediction remaining time and perform non-negativity check: determine if the prediction remaining time is less than zero. If the prediction remaining time is less than zero, correct the prediction remaining time to zero. If the prediction remaining time is greater than or equal to zero, leave it unchanged. Perform upper bound constraint verification: Determine whether the predicted remaining time after non-negativity verification is greater than the preset upper limit of process time. If the predicted remaining time after non-negativity verification is greater than the upper limit of process time, then correct it to the value of the upper limit of process time. If it is less than or equal to the upper limit of process time, then leave it unchanged.

[0011] Furthermore, the method for obtaining the deterministic prediction of the remaining time also includes: Set the inference cycle, and retrieve the prediction remaining time output from the previous inference cycle as the preceding determined value; Perform time series monotonicity check, calculate the algebraic difference between the predicted remaining time after upper bound constraint check and the previously determined value, and determine whether the algebraic difference is greater than the preset monotonicity tolerance threshold. If the algebraic difference is greater than the monotonicity tolerance threshold, the prediction result will be corrected to the difference between the preceding determined value and the inference period. If the algebraic difference is less than or equal to the monotonicity tolerance threshold, the value of the remaining prediction time after the upper bound constraint verification remains unchanged. The final value obtained after sequentially passing through nonnegativity check, upper bound constraint check, and temporal monotonicity check is defined as the deterministic prediction remaining time.

[0012] Furthermore, the method for obtaining the hardware interrupt signal includes: Obtain the clock frequency of the local system hardware master clock; The initial countdown count is obtained by multiplying the deterministic prediction remaining time obtained for the first time by the clock frequency. At the arrival of each new inference cycle, the main controller reads the current remaining count value of the decrementing count register in the local system hardware timer and multiplies the new deterministic prediction remaining time by the clock frequency to obtain the countdown target count value. The main controller directly writes the countdown target value into the decrementing count register, replacing the current remaining count value and writing it over to refresh the register; When the decrementing counter register reaches zero, the local system hardware timer outputs a hardware interrupt signal.

[0013] Furthermore, the execution process of the braking command includes: The main controller responds to the hardware interrupt signal and enters the interrupt service routine, switches the granulation operation flag in the internal status register to the stop state, and triggers the braking command issuance sequence. Obtain the instruction frame header, as well as the control data segment containing the zero-hertz target frequency value and braking deceleration time parameters; Based on the instruction frame header and the control data segment, a cyclic redundancy check code is calculated and generated, and the cyclic redundancy check code is encapsulated into the check tail segment. The command frame header, the control data segment, and the check tail segment are concatenated in sequence to construct and generate a braking command. The main controller sends the braking command to the frequency converter through the hardware control interface.

[0014] Furthermore, the method for calculating the cyclic redundancy check code includes: The instruction frame header and control data segment are concatenated and spliced ​​to form a binary bit sequence to be verified. Instantiate a 16-bit cyclic redundancy check register with all bits initialized to the value 1; Extract the current byte from the binary bit sequence to be checked, perform a bitwise XOR operation with the lower eight bits of the 16-bit cyclic redundancy check register, and update the lower eight bits with the result of the XOR operation. The entire 16-bit cyclic redundancy check register is shifted one bit to the right and zeros are added to the high bits. If the least significant bit of the overflow before the shift is one, the shifted 16-bit cyclic redundancy check register is XORed with the preset generator polynomial hexadecimal characteristic constant and stored back. If it is zero, the value remains unchanged. Repeat the shift and conditional XOR operation eight times to complete the current byte verification, until all bytes in the binary bit sequence to be verified have been traversed, and the final value stored in the sixteen-bit cyclic redundancy check register is established as the cyclic redundancy check code.

[0015] Furthermore, the method for obtaining the granulation batch endpoint control record includes: The frequency converter receives the braking command and parses the control data segment, controls the output frequency of the inverter inside the frequency converter, and causes the speed of the main shaft motor to decrease synchronously with the rotational angular velocity of the agitator. When the output frequency drops to zero hertz, the frequency converter cuts off the power supply to the main shaft motor, triggering the agitator to stop rotating completely. After the agitator has completely stopped, the main controller records the timestamp of the shutdown completion. The main controller combines the timestamp of the shutdown completion, the deterministic prediction of the remaining time that triggered this shutdown, the timestamp of the countdown trigger returning to zero, and the three-dimensional rheological feature vector corresponding to the countdown trigger returning to zero as a set of granulation batch endpoint control records and stores them in the non-volatile memory of the main controller.

[0016] A smart prediction system for the endpoint of wet granulation based on torque dynamic characteristics is used to implement the aforementioned smart prediction method for the endpoint of wet granulation based on torque dynamic characteristics. The system includes: Torque decoupling module: used to acquire the output torque data and running timestamp of the spindle motor, and construct a one-dimensional original torque time series in chronological order. A sliding time window is applied to the one-dimensional original torque time series to perform decoupling extraction, and a multi-dimensional rheological feature vector sequence is obtained. Time verification module: It is used to input the multidimensional rheological feature vector sequence into the pre-trained endpoint prediction algorithm model, output the predicted remaining time, and perform concatenated verification on the predicted remaining time to obtain the deterministic predicted remaining time. Interruption braking module: The main controller receives the deterministic prediction of the remaining time, generates and dynamically refreshes the countdown trigger in the local system hardware timer, outputs a hardware interrupt signal when the countdown trigger returns to zero, the main controller responds to the hardware interrupt signal, sends a braking command to the frequency converter through the hardware control interface, and generates a granulation batch end-point control record.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention constructs a one-dimensional original torque time series by synchronously acquiring output torque data and running timestamps, and maps it to a multi-dimensional rheological feature vector sequence. This effectively decouples the global rheological phase change trend from high-frequency transient disturbances, solving the problem of misjudgment at critical stages caused by traditional low-frequency single-point sampling and static threshold comparison. The endpoint prediction algorithm model transforms the traditional hysteresis feedback into feedforward inference prediction based on mean features, first-order difference features, and variance features. It also uses the series verification of non-negativity, upper bound constraints, and temporal monotonicity to generate deterministic prediction of the remaining time, accurately eliminating non-physical abnormal jumps caused by noise interference and improving system reliability. The local system hardware timer combined with the countdown trigger realizes the conversion of software prediction to underlying hardware interrupt signals, eliminating task scheduling delays, ensuring that braking commands are accurately issued to the frequency converter, avoiding excessive stirring of materials and damaging particle properties, and comprehensively improving granulation quality. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart of the intelligent prediction method for wet granulation endpoint based on torque dynamic characteristics provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the spatial motion monitoring coordinate system and the target vehicle's reversing into the parking space, provided in an embodiment of the present invention. Figure 3 This is a schematic diagram of the partition structure of the settlement determination plane provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the partition structure of the settlement determination plane provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the partition structure of the settlement determination plane provided in an embodiment of the present invention; Figure 6 This is a functional block diagram of the intelligent prediction system for the endpoint of wet granulation based on torque dynamic characteristics provided in an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1 Please see Figure 1 As shown, this embodiment provides an intelligent prediction and control method for the endpoint of wet granulation based on torque dynamic characteristics, including: Step S10: Obtain the output torque data and running timestamp of the spindle motor, and construct a one-dimensional original torque time series in chronological order. Apply a sliding time window to the one-dimensional original torque time series to perform decoupling extraction and obtain a multi-dimensional rheological feature vector sequence.

[0022] Further, step S10 includes: Step S11: Read the output torque data of the spindle motor and the corresponding running timestamp in real time through the communication bus of the spindle motor driver at a preset sampling frequency to construct a one-dimensional original torque time series.

[0023] In the automated production process of wet granulation equipment, the spindle motor drives the agitator to continuously mix and granulate the material. As the material transforms from loose powder to wet granules, its rheological properties continuously evolve, causing the load torque on the spindle motor to change constantly. In existing technologies, a separate torque sensor installed at the rear end of the spindle motor is typically used to intermittently acquire torque readings at a low sampling rate. These readings are then compared with a fixed static torque threshold at a single point to determine whether granulation has reached its endpoint. However, on the one hand, the separate torque sensor, as an additional invasive measuring element, requires modification of the spindle motor's mechanical structure, increasing the equipment's complexity and failure risk. On the other hand, the low sampling rate prevents the system from capturing the details of high-frequency transient torque fluctuations caused by local agglomeration of high-viscosity materials and non-Newtonian fluid shear thickening during the critical forming stage of wet granulation, resulting in a lack of sufficient temporal information for subsequent endpoint determination. Therefore, a non-invasive high-frequency torque timing signal synchronous acquisition mechanism based on the spindle motor driver communication bus is established. The aim is to record the entire dynamic evolution process of the spindle motor's output torque with sufficient temporal resolution without altering the equipment's mechanical structure.

[0024] Specifically, the main controller reads the output torque data and corresponding running timestamp of the spindle motor in real time through the communication bus of the spindle motor driver. The main controller refers to a programmable logic controller or industrial computer deployed in the control cabinet of the wet granulation equipment, possessing real-time data acquisition and processing capabilities, used to execute all data processing and control command issuance operations in this method. The spindle motor driver refers to a variable frequency drive that provides power to the spindle motor and has a built-in current loop feedback loop. While driving the spindle motor, the variable frequency drive continuously calculates and updates the instantaneous output torque of the spindle motor based on the real-time feedback current of its internal current loop. The communication bus refers to an industrial fieldbus communication link connecting the spindle motor driver and the main controller, used to enable the main controller to read the internal operating parameters of the spindle motor driver in real time. The output torque data refers to a scalar value calculated in real time by the spindle motor driver based on the internal current loop feedback current, reflecting the rotational torque experienced by the spindle motor when driving the agitator to cut material at the current physical instant. The running timestamp refers to a digital time stamp generated by the internal clock register of the spindle motor driver after each completion of the output torque data calculation, used to accurately mark the absolute physical instant of the torque sampling.

[0025] To ensure the acquired torque signal has sufficient time resolution to capture high-frequency transient torque fluctuations, the main controller periodically reads output torque data from the communication bus at a preset sampling frequency. The sampling frequency refers to the number of times the main controller reads output torque data from the spindle motor driver per unit time. The sampling frequency is set based on the typical duration of transient torque spikes caused by localized agglomeration of high-viscosity materials during the critical forming stage of wet granulation. This typical duration is obtained through pre-experimental calibration. The sampling frequency must meet the requirements of the Nyquist sampling theorem, i.e., the sampling frequency must be at least twice the highest frequency component contained in the transient torque spike, to ensure that no detailed information about high-frequency torque fluctuations is lost. For example, the sampling frequency is set to 1000 Hz, meaning output torque data is read once every 1 millisecond. Based on the sampling frequency, the main controller reads a set of output torque data and its corresponding running timestamp from the communication bus at each continuously advancing sampling moment. The sampling time is determined as follows: The physical instant the main controller receives the wet granulation process start command is taken as the zero point of time. The sequential index value of the currently executing reading action is obtained. Each sequential index value is arithmetically multiplied with the reciprocal of the sampling frequency to obtain a series of discrete time nodes that increase in an arithmetic sequence, which are then used as sampling times. The wet granulation process start command refers to a logic trigger level signal issued by the host computer's human-machine interface and transmitted to the main controller via the industrial network. It is used to instruct the main controller to activate a preset material mixing logic sequence and serves as the reference starting point for the system's global time axis. The main controller treats the output torque data read at each sampling time and its corresponding running timestamp as a data pair, and caches them sequentially in the circular data buffer within the main controller according to the chronological order of the running timestamps. The circular data buffer is a contiguous memory space with a fixed storage capacity allocated within the main controller. When the buffer is full, newly written data pairs will overwrite the oldest written data pairs, thus always retaining the torque sampling records within the most recent period. The storage capacity of the annular data buffer is set based on the product of the longest process duration of the wet granulation process from the addition of the binder to the end of granulation and the sampling frequency. The purpose is to ensure that the buffer can accommodate all torque sampling data throughout the entire granulation process without data loss. For example, the storage capacity of the annular data buffer is set to accommodate 600,000 data pairs. The ordered set of all output torque data arranged in chronological order of runtime timestamps in the annular data buffer is defined as a one-dimensional raw torque time series. The one-dimensional raw torque time series is expressed as follows: , where i represents the index of the sampling time, ranging from 1 to K, and K represents the total number of sampling times that have been completed up to the current sampling time, which is a dynamic quantized value that changes in real time with the system's operating status. This represents the output torque data value corresponding to the i-th sampling moment in the one-dimensional original torque time series; This represents the instantaneous torque scalar value calculated and output by the spindle motor driver at the i-th sampling moment; the one-dimensional original torque time series is a one-dimensional time-domain signal that continuously increases as the granulation process progresses, and fully records the entire dynamic evolution of the spindle motor torque from the start of the process to the current sampling moment.

[0026] Step S12: Apply a sliding time window to the one-dimensional original torque time series and map a multi-dimensional rheological feature vector sequence within each sliding time window.

[0027] After obtaining the one-dimensional raw torque time series, in order to simultaneously extract low-frequency baseline information characterizing the global viscosity increase trend of the material and high-frequency disturbance information characterizing the local agglomeration state of the material from this one-dimensional time domain signal, and to structurally stitch the two types of information in the time dimension to achieve multi-dimensional decoupling of the torque signal, a sliding time window is applied to the one-dimensional raw torque time series, and multi-dimensional statistical feature calculations are performed within the window. The aim is to transform the high-frequency torque spikes, which are filtered out as "system noise" in traditional methods, into physically meaningful predictive feature parameters, thereby solving the technical problem of the inseparability of global rheological phase transition trends and local transient disturbances in one-dimensional torque signals.

[0028] Specifically, the main controller applies a sliding time window of a preset length to the one-dimensional raw torque time series. The sliding time window refers to a continuous data extraction frame defined on the one-dimensional raw torque time series, having a fixed number of data points and gradually shifting forward along the time axis according to a preset step size. The preset length of the sliding time window is the number of sampled data points contained within the window, defined as the window length. The window length is set based on the product of the typical rheological transition time experienced by the material during wet granulation from local agglomeration to the formation of stable particle cores and the sampling frequency. The aim is to ensure that the data within the window covers at least one complete cycle to capture high-frequency disturbance characteristics, without obscuring the time-varying nature of the global trend due to an excessively long window. For example, the window length is set to 500 data points, corresponding to a time span of 0.5 seconds at a sampling frequency of 1000 Hz. The preset step size of the sliding time window is defined as the sliding step size. The sliding step size refers to the number of sampled data points skipped each time the sliding time window moves forward along the time axis. Its setting is based on balancing the temporal resolution of the feature sequence with the real-time computing power of the main controller. The aim is to ensure sufficient data overlap between adjacent sliding time windows to avoid feature breakage. For example, setting the sliding step size to 50 data points means that there is an overlap of 450 data points between adjacent windows.

[0029] Based on the window length and sliding step size, the main controller starts from the beginning of the one-dimensional original torque time series and sequentially extracts data points corresponding to the window length, forming the data subset within the first sliding time window. Then, it shifts the starting point backward by the corresponding sliding step size of data points and extracts data points again, forming the data subset within the second sliding time window, and so on, until the end of the window reaches the latest sampled data point of the one-dimensional original torque time series. The data subsets contained within all sliding time windows are defined as window data subsets, expressed in the following form: Where n represents the sequential number of the sliding time window, ranging from 1 to A, and A represents the total number of sliding time windows generated up to the current moment, which is a dynamically quantified value that changes in real time with the system's operating state. The output torque data value at the corresponding index position in the one-dimensional original torque time series is represented by S, where S represents the sliding step size and L represents the window length. Within each sliding time window, the main controller calculates three statistical features: mean feature, first-order difference feature, and variance feature. The mean feature is a scalar value obtained by performing an arithmetic mean operation on all output torque data values ​​within a single sliding time window. It is used to construct a low-frequency trend baseline characterizing the gradual increase in the global viscosity of the material as the granulation process progresses. The low-frequency trend baseline is a smooth curve formed by sequentially arranging the mean features of multiple consecutive sliding time windows in the time dimension. This curve reflects the macroscopic change trend of the average load level borne by the spindle motor during the transition of the material from a low-viscosity loose state to a high-viscosity agglomerate state. The first-order difference feature is a scalar value obtained by performing a pairwise algebraic difference operation on the output torque data values ​​of adjacent sampling data points within a single sliding time window, taking the absolute value of all difference results, and calculating the arithmetic mean. It is used to quantify the degree of instantaneous change in the torque signal between adjacent sampling times within the window. The variance feature refers to a scalar value obtained by summing the squares of the deviations of all output torque data values ​​from the mean feature within a single sliding time window, using the mean feature as a benchmark, and then dividing by the number of data points. This value is used to quantify the discrete fluctuation energy of the torque signal around the mean feature within the window. The first-order difference feature and the variance feature together constitute high-frequency perturbation parameters characterizing the local agglomeration state of the material. The high-frequency perturbation parameters refer to a set of statistics extracted from the one-dimensional original torque time series that reflect the transient high-frequency fluctuation characteristics of the torque signal within the window due to the local non-uniformity of the material.

[0030] After calculating the three statistical features, the main controller concatenates the mean, first-order difference, and variance features of the nth sliding time window along the feature dimension to form the three-dimensional rheological feature vector corresponding to the nth sliding time window. All three-dimensional rheological feature vectors corresponding to the sliding time windows are then arranged sequentially along the time dimension according to the sliding time window numbering, forming a multidimensional rheological feature vector sequence. (See also...) Figure 2 This is a schematic diagram illustrating the mapping and decoupling principle from a one-dimensional original torque time series to a multi-dimensional rheological feature vector sequence, provided by an embodiment of the present invention. Figure 2 The upper half of the diagram shows the waveform of the original one-dimensional torque time series, with the horizontal axis representing the sampling time and the vertical axis representing the output torque data. The rectangular dashed box in the diagram represents a sliding time window applied to the data sequence, which is shifted from left to right along the sampling time according to a preset sliding step size. The vertical dashed arrow in the middle of the diagram represents the computational transformation process of multi-dimensional statistical feature extraction performed by the system within the sliding time window. The lower half of the diagram shows the reconstructed multi-dimensional rheological feature vector sequence, with the horizontal axis representing the time dimension and the vertical axis representing statistical features. The thick black solid line represents the evolution curve of the mean feature, the short dashed line represents the evolution curve of the first-order difference feature, and the gray solid line represents the evolution curve of the variance feature. This visually demonstrates the transformation logic from a single aliased signal to multi-dimensional statistical feature extraction.

[0031] Step S10 solves the technical problems of traditional invasive measurement, which easily increases the risk of equipment failure and loses high-frequency fluctuation details due to low sampling rate, and the inability to separate the global rheological phase transition trend and local transient disturbance in the one-dimensional torque signal by using one-dimensional original torque time series, mean feature, first-order difference feature, variance feature and multi-dimensional rheological feature vector sequence. It realizes high-fidelity non-invasive acquisition of the dynamic evolution process of spindle motor torque, and decouples the mixed one-dimensional time domain torque signal to the multi-dimensional rheological feature space. Among them, the one-dimensional original torque time series fully recorded the dynamic evolution of the main shaft motor torque throughout the granulation process through high-frequency synchronous sampling, providing a sufficient temporal information basis for subsequent multi-dimensional feature extraction; the mean feature eliminated high-frequency fluctuations by performing an arithmetic average operation within a sliding time window, constructing a low-frequency trend baseline characterizing the increase in global viscosity of the material; the first-order difference feature captured the frequency and amplitude information of instantaneous torque jumps caused by local agglomeration of the material by quantifying the mean of the absolute change in torque values ​​at adjacent sampling times; the variance feature quantified the fluctuation energy of high-frequency torque peaks around the mean by calculating the dispersion of the torque signal within the window; the multi-dimensional rheological feature vector sequence structurally spliced ​​the mean feature, the first-order difference feature, and the variance feature in the time dimension, successfully transforming the system noise filtered out in traditional methods into predictive feature parameters with physical meaning.

[0032] Step S20: Input the multidimensional rheological feature vector sequence into the pre-trained endpoint prediction algorithm model, output the predicted remaining time, and perform a series verification on the predicted remaining time to obtain the deterministic predicted remaining time.

[0033] Further, step S20 includes: Step S21: Input the multidimensional rheological feature vector sequence into the pre-trained endpoint prediction algorithm model and output the predicted remaining time.

[0034] After obtaining the multidimensional rheological feature vector sequence, in order to utilize the dynamic evolution law of the high-frequency perturbation parameters contained in the multidimensional rheological feature vector sequence on the low-frequency trend baseline, and to achieve early prediction of the optimal granulation shutdown endpoint, an endpoint prediction algorithm model capable of processing time-series data and outputting the predicted remaining time is constructed. The aim is to upgrade the traditional "post-event feedback control," that is, waiting for the torque to reach the threshold to trigger shutdown, to "pre-event feedforward control," that is, predicting the remaining time before the material reaches the optimal granulation state, thereby eliminating the impact of system response delay on the granulation endpoint accuracy.

[0035] Specifically, the main controller extracts the three-dimensional rheological feature vectors corresponding to the last M consecutive sliding time windows from the multidimensional rheological feature vector sequence to form a historical feature matrix. M is the input time step, and its value satisfies... The input time step is set based on the ratio of the typical physical time period from the appearance of local agglomeration points to complete granulation in the wet granulation process to the sliding step. The purpose is to ensure that the historical feature matrix can cover the complete evolution trajectory of the material's rheological state from a "quasi-steady state" to a "mutational state." For example, the input time step M is set to 60. The historical feature matrix is ​​a two-dimensional structured data matrix with M rows and 3 columns. In the historical feature matrix, the row dimension represents the time evolution sequence, that is, from the first row to the Mth row, they correspond to M consecutive sliding time windows from the past to the latest moment; the column dimension represents the physical feature classification, that is, all elements in the first column correspond to the mean feature, all elements in the second column correspond to the first-order difference feature, and all elements in the third column correspond to the variance feature. To eliminate the absolute numerical difference between the mean feature representing global viscosity and the variance and difference features representing local agglomeration, the main controller independently performs zero-mean, unit-variance standardization on each column of the historical feature matrix. Specifically, for the first, second, and third columns of the historical feature matrix, the arithmetic mean and standard deviation of all M feature elements in each column are calculated independently. Each feature element in a column is then subtracted from the arithmetic mean of its column and divided by the standard deviation of that column to obtain the standardized element value. The historical feature matrix after standardization is defined as the standardized input matrix.

[0036] The endpoint prediction algorithm model employs a Long Short-Term Memory (LSTM) recurrent neural network (RNN) structure. LSM is a deep learning network architecture specifically designed for processing time-series data. It achieves selective memorization and forgetting of historical information over long time spans through three gating mechanisms: a forget gate, an input gate, and an output gate, overcoming the inherent gradient vanishing defect of ordinary recurrent neural networks during long-sequence training. The main controller instantiates this network structure in its internal memory. Its complete structure consists of an input layer, a hidden layer, and an output layer cascaded sequentially. The input layer contains three input neurons, corresponding one-to-one with the three feature columns of the standardized input matrix, used to synchronously receive a three-dimensional feature vector containing the mean, difference, and variance at each time step. The hidden layer is configured as a single-layer LSM RNN unit. This unit contains a cell state vector and forget gates, input gates, and output gates for controlling the ratio of information inflow to outflow. The number of neurons in the hidden layer is defined as the hidden dimension. For example, the hidden dimension is set to 32 to provide sufficient internal parameter capacity to fit the nonlinear phase transition trajectory of complex material rheology; the output layer is configured as a fully connected linear mapping layer, whose input is connected to the output of the hidden layer, and the output contains one computing node, used to finally reduce the dimensionality of the rheological time series features extracted by the hidden layer in high-dimensional space to a specific physical time scalar value. The training process is as follows: Obtain the original torque data of historical operating batches of the wet granulation equipment within the health calibration period and with granulation quality meeting standards, as the training basis data; convert the training basis data into the corresponding historical multidimensional rheological feature vector sequence; define the absolute physical moment when each historical batch reaches the optimal particle density, as the historical calibration endpoint, through review by process experts. Using a sliding window overlapping truncation method, extract multiple segments of standardized historical input matrices of length M from the historical multidimensional rheological feature vector sequence as input samples; simultaneously calculate the physical time difference between the truncation time of each input sample and the corresponding historical calibration endpoint, as the true remaining time label; initialize all network weight parameters within the endpoint prediction algorithm model in the external training server. The network weight parameters include the gated weight matrix and recurrent connection weight matrix within the hidden layers, as well as the weight vector and bias scalar of the output layer. A truncated normal distribution initialization method is used to initialize all network weight parameters to random values ​​with a mean of 0 and a standard deviation of 0.01 to prevent gradient vanishing within the gated units during the early stages of training. Input samples are fed into the input layer step by step, and state update operations are performed through the hidden layers. The output state of the hidden layer at the last time step is then fed into the fully connected linear mapping layer of the output layer to complete the forward propagation process, obtaining the model prediction value. The model prediction value refers to the ideal estimated remaining time output by the endpoint prediction algorithm model based on the current input sample.Calculate the square of the difference between each model prediction and the corresponding true remaining time label, and average the squared differences of all samples. Define this average as the mean squared error loss function. Utilize the adaptive moment estimation optimization algorithm, perform backpropagation based on the gradient of the partial derivatives of the mean squared error loss function with respect to each network weight parameter, and update all network weight parameters. Set the learning rate to ensure smooth convergence of the network weight parameter update step size in the complex rheological feature space, preventing numerical oscillations near the minimum value of the loss function due to excessively large step sizes. For example, set it to 0.001. Set the convergence threshold to ensure that the model training accuracy meets the stringent tolerance requirements of industrial-grade wet granulation processes for downtime physical time errors, preventing downtime lag from causing particle structure damage. For example, set the mean squared error convergence threshold to 0.25. Set the maximum number of iterations to prevent the model from getting trapped in local minima and endlessly consuming computational resources, based on the computational resource constraints of the training platform. For example, set it to 1000 iterations. When the mean squared error loss function value is less than the convergence threshold or the maximum number of iterations is reached, parameter updates are stopped, and all network weight parameters, including the weight vector and bias scalar, that have finally converged are solidified and downloaded and deployed to the non-volatile memory of the field main controller, thus completing the training and construction of the endpoint prediction algorithm model.

[0037] The main controller feeds the standardized input matrix generated in real time into the input layer row by row in chronological order. At each time step, the hidden layer units receive the 3D feature vector of the current row and the hidden state vector passed from the previous time step, and update the current hidden state vector through internal control logic. The hidden state vector refers to the high-dimensional digital representation vector generated by the neurons inside the model, which integrates all previous historical rheological information. After completing the recursive operation of all M time steps, the main controller extracts the final hidden state vector output by the hidden layer at the last time step and sends it to the output layer. The output layer outputs the predicted remaining time. The predicted remaining time refers to the remaining physical time required for the current material physical state to reach the optimal particle density, i.e., the theoretical optimal shutdown endpoint.

[0038] Step S22: Perform a cascaded verification on the predicted remaining time to obtain the deterministic predicted remaining time.

[0039] After obtaining the predicted remaining time from the endpoint prediction algorithm model, the neural network model may be affected by residual random noise in the input features or abnormal conditions such as localized large-scale collapses of materials during the inference process. This can lead to the predicted remaining time output by the model exhibiting negative values ​​that do not conform to physical logic, excessively long values ​​that exceed process norms, or drastic jumps between adjacent time points. If such unstable original prediction values ​​are directly used to trigger the underlying braking action, it will cause malfunctions in the control system or even vibration damage to the mechanical structure. Therefore, a logic filtering mechanism based on engineering constraints is established. The purpose is to ensure the engineering reliability of the output results by performing closed-loop verification of the physical boundaries and timing logic on the original prediction values.

[0040] Specifically, the main controller acquires the predicted remaining time output by the endpoint prediction algorithm model at the current sampling time. The inference period is defined as the time interval between the start times of two adjacent sliding time windows, and the value of the inference period is equal to the quotient obtained by dividing the sliding step size by the sampling frequency. The main controller performs a series verification on the predicted remaining time according to three rules. The first verification rule is non-negativity verification. The main controller determines whether the predicted remaining time is less than zero. If the predicted remaining time is less than zero, the main controller determines that the material rheological state has exceeded the theoretical granulation endpoint, and forcibly corrects the predicted remaining time to zero; conversely, if the predicted remaining time is greater than or equal to zero, the main controller determines that the current predicted value conforms to the physical property of positive time flow, and keeps the value of the predicted remaining time unchanged. The second verification rule is upper bound constraint verification. The main controller determines whether the predicted remaining time after non-negativity verification is greater than the preset upper limit of the process time. The upper limit of the process time is set based on the longest safe duration of a single batch stirring operation specified in the wet granulation formula, with the aim of filtering out abnormal predicted values ​​caused by abnormal model input. For example, the upper limit of the process time is set to 300 seconds. If the predicted remaining time after non-negativity constraint verification is greater than the upper limit of the process time, the main controller will correct the predicted remaining time after non-negativity constraint verification to the upper limit of the process time; if the predicted remaining time after non-negativity constraint verification is less than or equal to the upper limit of the process time, the main controller will keep the predicted remaining time unchanged. The third verification rule is timing monotonicity verification. The main controller retrieves the predicted remaining time output from the previous inference cycle from its internal register, which is defined as the preceding determined value. The main controller calculates the algebraic difference between the predicted remaining time after upper limit constraint verification and the preceding determined value. It determines whether the algebraic difference is greater than a preset monotonicity tolerance threshold. The monotonicity tolerance threshold is set based on 1.5 to 2 times the inference cycle value, which allows the prediction result to have a small positive fluctuation when the material rheology fluctuates drastically, while blocking non-physical jumps. For example, the monotonicity tolerance threshold is set to 0.1 seconds. If the algebraic difference is greater than the monotonicity tolerance threshold, the prediction result is corrected to the difference between the preceding determined value and the inference cycle, to force the maintenance of the logical trend that the prediction time decreases as the process progresses; if the algebraic difference is less than or equal to the monotonicity tolerance threshold, the main controller keeps the current value unchanged. The main controller defines the value obtained after processing through the above three verification rules as the deterministic prediction remaining time. The deterministic prediction remaining time is the final time parameter that meets the conditions for hard real-time control execution after physical boundary alignment and timing logic smoothing.

[0041] See Figure 3 This is a schematic diagram illustrating the principle of multidimensional rheological feature evolution and endpoint prediction path provided in an embodiment of the present invention. Figure 3As shown, the upper part of the figure is displayed here. Figure 2 The figure describes the synchronous evolution trajectory of the mean characteristic curve, first-order difference characteristic curve, and variance characteristic curve over time. The difference lies in the vertical dashed line, which marks the transient jump characteristics of the high-frequency perturbation parameters captured by the endpoint prediction algorithm model on the low-frequency trend baseline, defined as the dynamic mutation point. The lower half of the figure shows the evolution trajectory of the model's output prediction remaining time over time, starting from this dynamic mutation point. In this trajectory, the prediction starting point corresponds to the physical instant when the dynamic mutation point is first captured; subsequently, the predicted value continuously converges along a smooth path, and the zeroing point at the end of the trajectory corresponds to the optimal granulation shutdown endpoint for the material. This figure intuitively reveals the evolutionary logic of achieving feedforward prediction of the granulation endpoint by identifying dynamic mutation points of rheological characteristics and combining them with deep learning deduction. See also... Figure 4 This is a flowchart illustrating the logical verification process for generating a deterministic prediction of remaining time, provided by an embodiment of the present invention. Figure 4 As shown, the predicted remaining time output by the endpoint prediction algorithm model enters the verification system and undergoes a cascaded processing path involving nonnegativity verification, upper bound constraint verification, and temporal monotonicity verification. At each verification node, the system identifies the physical rationality of the predicted value through preset logical judgment conditions and performs corresponding correction and compensation operations for values ​​that do not conform to material rheology logic or process constraints. Through the cascaded verification of the above three rules, a deterministic predicted remaining time that meets the conditions for hard real-time control execution is finally output.

[0042] Step S20 solves the technical problems of traditional hysteresis feedback control causing granulation endpoint accuracy deviation due to physical time delay, and the output of pure data-driven neural network models being susceptible to noise interference, resulting in non-physical abnormal jumps, by using the endpoint prediction algorithm model, predicting the remaining time, and deterministically predicting the remaining time. It realizes the intelligent control upgrade from passive threshold triggering to pre-prediction and ensures the absolute safety and engineering reliability of the predicted time in the underlying braking execution link of the industrial site. Among them, the endpoint prediction algorithm model learns and fits the dynamic evolution law of high-frequency disturbance parameters on the low-frequency trend baseline through the gating mechanism of long short-term memory recurrent neural network, and maps the complex multidimensional rheological time series characteristics into a time scalar by reducing the dimensionality. The predicted remaining time provides a forward-looking physical countdown of the current material state from the theoretical optimal shutdown endpoint, which completely compensates for the system delay of signal acquisition and mechanical shutdown. The deterministic prediction of the remaining time is verified by the series closed-loop verification of three engineering rules: non-negativity, upper bound constraint and time series monotonicity. This eliminates negative value overflow, abnormal anomalies and non-monotonic jumps caused by model inference fluctuations, and ensures that the final output time parameter has stable and continuous hard real-time control execution conditions.

[0043] In step S30, the main controller receives the deterministic prediction of the remaining time, generates and dynamically refreshes the countdown trigger in the local system hardware timer, outputs a hardware interrupt signal when the countdown trigger returns to zero, the main controller responds to the hardware interrupt signal, sends a braking command to the frequency converter through the hardware control interface, and generates a granulation batch endpoint control record.

[0044] Further, step S30 includes: Step S31: The main controller receives the deterministic prediction of the remaining time, generates and dynamically refreshes the countdown trigger in the local system hardware timer, and outputs a hardware interrupt signal when the countdown trigger returns to zero.

[0045] After obtaining the deterministic predicted remaining time, in order to transform the purely software-level prediction and logic verification results into a hardware execution mechanism capable of precisely controlling the shutdown time of physical equipment, the main controller needs to establish a countdown triggering mechanism based on the local system hardware timer. The aim is to ensure the precise triggering of the shutdown command at the predicted time through hard real-time timer interrupts, eliminating the impact of operating system task scheduling delays on shutdown accuracy.

[0046] Specifically, after receiving the deterministic prediction of the remaining time in each inference cycle, the main controller performs a dynamic update operation on the countdown trigger. The countdown trigger refers to a decrementing counter register maintained in the local system hardware timer within the main controller. This decrementing counter register is a hardware physical register unit within the hardware timer specifically used to temporarily store digitized count values ​​and decrement its internal value by one when triggered by an external clock pulse. The controller also obtains the clock frequency of the local system hardware master clock. The local system hardware master clock refers to a high-precision hardware time base driven by a crystal oscillator within the main controller, used to provide a stable decrementing count for the countdown trigger. The clock frequency refers to the number of clock pulses generated per second by the local system hardware master clock. The countdown trigger performs decrementing counting at the clock frequency of the main controller's local system hardware master clock, and automatically triggers a hardware interrupt signal when the value in the decrementing counter register decreases to zero.

[0047] The initialization and dynamic update process of the countdown trigger is as follows: The main controller obtains the clock frequency of the local system hardware master clock, performs an arithmetic product operation on the initially obtained deterministic predicted remaining time and the clock frequency to obtain the initial count value of the countdown. The initial count value of the countdown represents the total number of clock pulses that the countdown trigger needs to experience from the current moment to the predicted optimal granulation stop endpoint.

[0048] At the arrival of each new inference cycle, i.e., after generating a new deterministic predicted remaining time, the main controller performs a dynamic refresh operation on the countdown trigger. Specifically, the main controller reads the current remaining count value from the decrementing counter register in the local system hardware timer, re-multiplies the new deterministic predicted remaining time with the clock frequency to obtain a new initial count value, defined as the countdown target count value. The main controller directly overwrites the countdown target count value into the decrementing counter register, replacing the current remaining count value, thereby achieving real-time calibration of the countdown trigger. When the remaining count value in the decrementing counter register of the local system hardware timer decreases to zero, the local system hardware timer sends a hardware interrupt signal to the central processing unit of the main controller. The hardware interrupt signal is a level transition signal automatically generated by the comparator circuit of the local system hardware timer when it detects that the value of the decrementing counter register is equal to zero. This signal has the highest interrupt priority and can immediately interrupt any software task currently being executed by the main controller, ensuring that the shutdown response has absolute determinism.

[0049] In step S32, the main controller responds to the hardware interrupt signal, sends a braking command to the frequency converter through the hardware control interface, and generates a granulation batch endpoint control record.

[0050] After the local system hardware timer outputs a hardware interrupt signal, in order to convert the underlying hardware timing trigger into a physical braking action of the spindle motor, the main controller uses the instruction transmission link between the hardware control interface and the frequency converter to realize the physical process of driving the agitator to decelerate and stop. The purpose is to ensure that the transmission and execution of braking commands have industrial-grade determinism and reliability, so that the agitator stops precisely at the predicted optimal pelleting time.

[0051] Specifically, upon receiving a hardware interrupt signal, the central processing unit of the main controller immediately enters the interrupt service routine. The interrupt service routine is a sequence of instructions pre-programmed into the main controller's non-volatile memory, specifically designed to respond to countdown trigger reset events. In the interrupt service routine, the main controller first switches the granulation operation flag in its internal status register from "running state" to "stopped state." The granulation operation flag is a binary logic variable maintained internally by the main controller, taking the value of either "running state" or "stopped state," used to mark the current execution status of the wet granulation process at the system level. The "running state" refers to the sequential working state in which the various software control tasks inside the main controller are in a high-frequency polling scheduling, the communication bus periodically reads the output torque data, and the endpoint prediction algorithm model is continuously activated and performs online inference. At this time, the underlying frequency converter maintains normal power output to drive the agitator to continuously shear and mix the material. The "stop state" refers to the control lock state in which the main controller suspends the online inference task of the endpoint prediction algorithm model, locks the data reading action of the communication bus, and triggers the braking command issuance sequence. At this time, the scheduling resources of the main controller are forcibly focused on driving the spindle motor to perform safe braking and deceleration.

[0052] The main controller sends braking commands to the frequency converter of the spindle motor via a hardware control interface. The hardware control interface refers to the physical communication connection established between the main controller and the frequency converter, which uses an industrial fieldbus protocol to transmit digital commands. The frequency converter is a power conversion device that provides adjustable frequency AC power to the spindle motor, controlling the spindle motor's speed by receiving frequency commands from the main controller. The braking command is a structured control message containing a target frequency of zero Hz and a braking deceleration time parameter. The braking deceleration time parameter refers to the time required for the frequency converter to linearly decrease from its current output frequency to zero Hz. Its setting is based on the rotational inertia of the spindle motor and the agitator, as well as the maximum braking torque capacity of the frequency converter's built-in braking unit. The purpose is to ensure that the agitator stops smoothly under a controllable deceleration curve, avoiding mechanical impact damage to the equipment and materials caused by sudden stops. For example, the braking deceleration time parameter is set to 2 seconds. The braking command structure consists of three sequentially concatenated parts: the first part is the command frame header, which contains the device address of the main controller and the slave address of the frequency converter, used to establish precise point-to-point addressing routing on the communication bus; the second part is the control data segment, which contains the target frequency value and braking deceleration time parameters, encapsulated in binary encoding; the third part is the check tail segment, which contains a cyclic redundancy check (CRC) code calculated based on the command frame header and the control data segment. The CRC code is a high-order CRC code that uses binary polynomial division and XOR logic operations to detect errors in serially transmitted data. Specifically: The main controller concatenates the instruction frame header and the control data segment in byte order to form a one-dimensional binary bit sequence to be verified; the main controller instantiates a 16-bit cyclic redundancy check (CRC) register in its internal register and initializes all bits of the 16-bit CRC register to the value 1; the main controller extracts the first byte from the binary bit sequence to be verified, performs a bitwise XOR operation between this byte and the lower eight bits of the 16-bit CRC register, and replaces the lower eight bits of the original 16-bit CRC register with the resulting eight bits; subsequently... The main controller shifts all bits of the 16-bit Cyclic Redundancy Check (CRC) register one bit to the right, padding the high-order bits with zeros, and checks whether the least significant bit that overflowed before the shift is a value of 1. If the least significant bit that overflowed is a value of 1, the main controller calls a preset generator polynomial hexadecimal characteristic constant. The generator polynomial hexadecimal characteristic constant refers to a specific characteristic value used in the CRC algorithm to perform a conditional XOR operation with the data register to generate an error detection code. This value is obtained by the main controller reading the industrial fieldbus communication protocol compatible with the spindle motor inverter, such as the Modbus-RTU protocol standard specification manual.The shifted 16-bit Cyclic Redundancy Check (CRC) register is XORed with the generator polynomial hexadecimal characteristic constant, and the result is stored back in the CRC register. If the least significant bit of the overflow is detected to be zero, the main controller blocks the XOR operation, keeping the value of the shifted 16-bit CRC register unchanged. The main controller repeats the above shift and conditional XOR operation eight times to complete the check cycle for the first byte. Then, the main controller sequentially introduces the next byte in the binary bit sequence to be checked, and repeats the above bitwise XOR, shift, and conditional XOR operations until all bytes in the binary bit sequence to be checked have been traversed. At this point, the main controller establishes the final 16-bit binary value remaining in the CRC register as the CRC code. The CRC code is sent to the frequency converter along with the control data segment, so that the frequency converter can perform communication error detection at the receiving end according to the same polynomial division principle to ensure the data integrity of the braking command during transmission.

[0053] After receiving the braking command, the frequency converter analyzes the target frequency value and braking deceleration time parameter in the control data segment, and controls the output frequency of its internal inverter to decrease from the current operating frequency to zero Hz according to a linear decrease curve within the time length specified by the braking deceleration time parameter. During the output frequency decrease process, the speed of the main shaft motor decreases synchronously, and the rotational angular velocity of the agitator continuously decreases accordingly. When the output frequency of the frequency converter drops to zero Hz, the frequency converter cuts off the power supply to the main shaft motor, and the agitator completely stops rotating. After the agitator completely stops, the main controller records the timestamp of the shutdown completion, and stores this timestamp, along with the deterministic predicted remaining time that triggered this shutdown, the timestamp of the countdown trigger resetting to zero, and the three-dimensional rheological feature vector corresponding to the countdown trigger resetting to zero, as a set of granulation batch endpoint control records in the non-volatile memory of the main controller. The granulation batch endpoint control record is a structured data file containing time information and rheological feature snapshots. See also Figure 5 This is a timing diagram for the execution of a variable frequency linkage closed-loop control provided in an embodiment of the present invention. Figure 5As shown in the figure, the horizontal axis represents physical time. The upper part of the figure shows the overall trend of the countdown register in the local system hardware timer gradually decreasing from the initial count value to zero. The lower part of the figure shows the physical process of the inverter controlling its internal inverter output frequency to decrease from the current operating frequency to zero Hz according to a linear decreasing curve after receiving the braking command, and the spindle motor speed decreasing synchronously, and the rotational angular velocity of the agitator continuously decreasing until it stops rotating completely. The upper and lower parts are connected by a vertical dashed line starting from the moment the countdown register value decreases to zero. This vertical dashed line represents the hardware timer interrupt trigger point in terms of timing, intuitively showing the causal linkage between the hardware interrupt signal generated by the local system hardware timer and the underlying inverter and agitator physical braking action.

[0054] Step S30, through countdown triggers, braking commands, and batch end-of-pipe control records, solves the technical problem in traditional wet granulation control where the system response delay between the threshold determination and actual shutdown leads to excessive material agitation and damage to particle properties. It achieves precise time-domain conversion of deterministic prediction of remaining time to underlying physical braking actions, accurate shutdown of the agitator at the optimal rheological state point, and full traceability recording of the granulation end-of-pipe control process. Specifically, the countdown trigger, through the decrementing count of the local system hardware timer and hardware interrupt mechanism, transforms the software prediction and verification results into a hard real-time trigger signal with a deterministic response, eliminating the impact of software task scheduling delays on shutdown accuracy. The braking command, through structured control messages, accurately transmits the target frequency value and braking deceleration time parameters to the frequency converter, ensuring the agitator stops smoothly under a controllable linear deceleration curve. The batch end-of-pipe control record structurally archives the time information and rheological characteristic snapshots at the shutdown moment, providing high-fidelity data evidence for process quality traceability and model iterative optimization.

[0055] Example 2 This embodiment, based on Embodiment 1, provides a closed-loop automatic assembly and adjustment system for optical systems based on real-time wavefront feedback, such as... Figure 6 As shown, it includes: Torque decoupling module: used to acquire the output torque data and running timestamp of the spindle motor, and construct a one-dimensional original torque time series in chronological order. A sliding time window is applied to the one-dimensional original torque time series to perform decoupling extraction, and a multi-dimensional rheological feature vector sequence is obtained. Time verification module: It is used to input the multidimensional rheological feature vector sequence into the pre-trained endpoint prediction algorithm model, output the predicted remaining time, and perform concatenated verification on the predicted remaining time to obtain the deterministic predicted remaining time. Interruption braking module: The main controller receives the deterministic prediction of the remaining time, generates and dynamically refreshes the countdown trigger in the local system hardware timer, outputs a hardware interrupt signal when the countdown trigger returns to zero, the main controller responds to the hardware interrupt signal, sends a braking command to the frequency converter through the hardware control interface, and generates a granulation batch end-point control record.

[0056] In the torque decoupling module, the output torque data and running timestamp of the spindle motor are acquired, and a one-dimensional original torque time series is constructed in chronological order. A sliding time window is applied to the one-dimensional original torque time series to perform decoupling extraction, resulting in a multi-dimensional rheological feature vector sequence, including: Step S11: Read the output torque data of the spindle motor and the corresponding running timestamp in real time through the communication bus of the spindle motor driver at a preset sampling frequency to construct a one-dimensional original torque time series. Step S12: Apply a sliding time window to the one-dimensional original torque time series and map a multi-dimensional rheological feature vector sequence within each sliding time window.

[0057] In the time verification module, the step of inputting the multidimensional rheological feature vector sequence into the pre-trained endpoint prediction algorithm model, outputting the predicted remaining time, and performing concatenated verification on the predicted remaining time to obtain the deterministic predicted remaining time includes: Step S21: Input the multidimensional rheological feature vector sequence into the pre-trained endpoint prediction algorithm model and output the predicted remaining time. Step S22: Perform a cascaded verification on the predicted remaining time to obtain the deterministic predicted remaining time.

[0058] In the interrupt braking module, the main controller receives a deterministic prediction of the remaining time, generates and dynamically refreshes a countdown trigger within the local system hardware timer, and outputs a hardware interrupt signal when the countdown trigger reaches zero. The main controller responds to the hardware interrupt signal, sends a braking command to the frequency converter through the hardware control interface, and generates a granulation batch endpoint control record, including: Step S31: The main controller receives the deterministic prediction of the remaining time, generates and dynamically refreshes the countdown trigger in the local system hardware timer, and outputs a hardware interrupt signal when the countdown trigger returns to zero. In step S32, the main controller responds to the hardware interrupt signal, sends a braking command to the frequency converter through the hardware control interface, and generates a granulation batch endpoint control record.

[0059] In addition, the parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of the corresponding technical solutions in the prior art have not been described in detail, so as to avoid excessive elaboration.

[0060] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely preferred embodiments of the present invention, and the scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for intelligent prediction of the endpoint of wet granulation based on dynamic torque characteristics, characterized in that, The method includes: The output torque data and running timestamp of the spindle motor are obtained, and a one-dimensional original torque time series is constructed in chronological order. A sliding time window is applied to the one-dimensional original torque time series to perform decoupling extraction, resulting in a multi-dimensional rheological feature vector sequence. The multidimensional rheological feature vector sequence is input into the pre-trained endpoint prediction algorithm model, the predicted remaining time is output, and a concatenated verification is performed on the predicted remaining time to obtain the deterministic predicted remaining time. The main controller receives the deterministic prediction of the remaining time, generates and dynamically refreshes the countdown trigger in the local system hardware timer, and outputs a hardware interrupt signal when the countdown trigger returns to zero. The main controller responds to the hardware interrupt signal, sends a braking command to the frequency converter through the hardware control interface, and generates a granulation batch endpoint control record.

2. The intelligent prediction method for the endpoint of wet granulation based on dynamic torque characteristics according to claim 1, characterized in that, The method for constructing the one-dimensional original torque time series includes: The output torque data and corresponding running timestamp of the spindle motor are read in real time through the communication bus of the spindle motor driver at a preset sampling frequency. The output torque data read at each preset sampling time and its corresponding running timestamp are taken as a data pair, and the data pairs are cached sequentially in the circular data buffer inside the main controller according to the time order of the running timestamps. The ordered set of all output torque data arranged in chronological order of runtime timestamps in the circular data buffer is defined as a one-dimensional original torque time series.

3. The intelligent prediction method for the endpoint of wet granulation based on dynamic torque characteristics according to claim 1, characterized in that, The method for obtaining the multidimensional rheological feature vector sequence includes: A sliding time window with a preset window length and sliding step size is applied to the one-dimensional original torque time series, and a corresponding number of data points are sequentially extracted to form a subset of window data. Within each sliding time window, the arithmetic mean of the output torque data values ​​is calculated to obtain the mean feature, the mean of the absolute difference between adjacent data points is calculated to obtain the first-order difference feature, and the variance feature around the mean feature is calculated. The mean feature, first-order difference feature, and variance feature are concatenated to form a three-dimensional rheological feature vector; Arrange the three-dimensional rheological feature vectors corresponding to all sliding time windows in chronological order to form a multidimensional rheological feature vector sequence.

4. The intelligent prediction method for the endpoint of wet granulation based on dynamic torque characteristics according to claim 3, characterized in that, The endpoint prediction algorithm model includes: The three-dimensional rheological feature vectors corresponding to the last consecutive sliding time window of a preset number of input time steps are extracted from the multidimensional rheological feature vector sequence to form the historical feature matrix. The arithmetic mean and standard deviation of each feature column of the historical feature matrix are calculated independently. Zero mean and unit variance standardization is performed on each feature element in the column to obtain the standardized input matrix. The standardized input matrix is ​​fed into the input layer of the pre-trained endpoint prediction algorithm model row by row in chronological order.

5. The intelligent prediction method for the endpoint of wet granulation based on dynamic torque characteristics according to claim 4, characterized in that, The method for obtaining the predicted remaining time includes: The endpoint prediction algorithm model includes a hidden layer and an output layer that are cascaded sequentially with the input layer; Using the row-by-row input process of the standardized input matrix as a time step, at each time step, the hidden layer receives the three-dimensional rheological feature vector corresponding to the current row and the hidden state vector passed from the previous time step, calls the internal door control logic to generate and update the current hidden state vector, and performs recursive operations. After completing the recursive operation of all time steps corresponding to the number of input time steps, the final hidden state vector output by the hidden layer at the last time step is extracted and sent to the output layer. The output layer performs a fully connected linear mapping on the final hidden state vector, and the output is a scalar value predicting the remaining time.

6. The intelligent prediction method for the endpoint of wet granulation based on dynamic torque characteristics according to claim 1, characterized in that, The method for obtaining the deterministic prediction of the remaining time includes: Obtain the output prediction remaining time and perform non-negativity check: determine if the prediction remaining time is less than zero. If the prediction remaining time is less than zero, correct the prediction remaining time to zero. If the prediction remaining time is greater than or equal to zero, leave it unchanged. Perform upper bound constraint verification: Determine whether the predicted remaining time after non-negativity verification is greater than the preset upper limit of process time. If the predicted remaining time after non-negativity verification is greater than the upper limit of process time, then correct it to the value of the upper limit of process time. If it is less than or equal to the upper limit of process time, then leave it unchanged.

7. The intelligent prediction method for the endpoint of wet granulation based on dynamic torque characteristics according to claim 6, characterized in that, The method for obtaining the deterministic prediction of the remaining time also includes: Set the inference cycle, and retrieve the prediction remaining time output from the previous inference cycle as the preceding determined value; Perform time series monotonicity check, calculate the algebraic difference between the predicted remaining time after upper bound constraint check and the previously determined value, and determine whether the algebraic difference is greater than the preset monotonicity tolerance threshold. If the algebraic difference is greater than the monotonicity tolerance threshold, the prediction result will be corrected to the difference between the preceding determined value and the inference period. If the algebraic difference is less than or equal to the monotonicity tolerance threshold, the value of the remaining prediction time after the upper bound constraint verification remains unchanged. The final value obtained after sequentially passing through nonnegativity check, upper bound constraint check, and temporal monotonicity check is defined as the deterministic prediction remaining time.

8. The intelligent prediction method for the endpoint of wet granulation based on dynamic torque characteristics according to claim 1, characterized in that, The method for obtaining the hardware interrupt signal includes: Obtain the clock frequency of the local system hardware master clock; The initial countdown count is obtained by multiplying the deterministic prediction remaining time obtained for the first time by the clock frequency. At the arrival of each new inference cycle, the main controller reads the current remaining count value of the decrementing count register in the local system hardware timer and multiplies the new deterministic prediction remaining time by the clock frequency to obtain the countdown target count value. The main controller directly writes the countdown target value into the decrementing count register, replacing the current remaining count value and writing it over to refresh the register; When the decrementing counter register reaches zero, the local system hardware timer outputs a hardware interrupt signal.

9. The intelligent prediction method for the endpoint of wet granulation based on dynamic torque characteristics according to claim 1, characterized in that, The execution process of the braking command includes: The main controller responds to the hardware interrupt signal and enters the interrupt service routine, switches the granulation operation flag in the internal status register to the stop state, and triggers the braking command issuance sequence. Obtain the instruction frame header, as well as the control data segment containing the zero-hertz target frequency value and braking deceleration time parameters; Based on the instruction frame header and the control data segment, a cyclic redundancy check code is calculated and generated, and the cyclic redundancy check code is encapsulated into the check tail segment. The command frame header, the control data segment, and the check tail segment are concatenated in sequence to construct and generate a braking command. The main controller sends the braking command to the frequency converter through the hardware control interface.

10. The intelligent prediction method for the endpoint of wet granulation based on dynamic torque characteristics according to claim 9, characterized in that, The method for calculating the cyclic redundancy check code includes: The instruction frame header and control data segment are concatenated and spliced ​​to form a binary bit sequence to be verified. Instantiate a 16-bit cyclic redundancy check register with all bits initialized to the value 1; Extract the current byte from the binary bit sequence to be checked, perform a bitwise XOR operation with the lower eight bits of the 16-bit cyclic redundancy check register, and update the lower eight bits with the result of the XOR operation. The entire 16-bit cyclic redundancy check register is shifted one bit to the right and zeros are added to the high bits. If the least significant bit of the overflow before the shift is one, the shifted 16-bit cyclic redundancy check register is XORed with the preset generator polynomial hexadecimal characteristic constant and stored back. If it is zero, the value remains unchanged. Repeat the shift and conditional XOR operation eight times to complete the current byte verification, until all bytes in the binary bit sequence to be verified have been traversed, and the final value stored in the sixteen-bit cyclic redundancy check register is established as the cyclic redundancy check code.

11. The intelligent prediction method for the endpoint of wet granulation based on dynamic torque characteristics according to claim 9, characterized in that, The method for obtaining the granulation batch endpoint control record includes: The frequency converter receives the braking command and parses the control data segment, controls the output frequency of the inverter inside the frequency converter, and causes the speed of the main shaft motor to decrease synchronously with the rotational angular velocity of the agitator. When the output frequency drops to zero hertz, the frequency converter cuts off the power supply to the main shaft motor, triggering the agitator to stop rotating completely. After the agitator has completely stopped, the main controller records the timestamp of the shutdown completion. The main controller combines the timestamp of the shutdown completion, the deterministic prediction of the remaining time that triggered this shutdown, the timestamp of the countdown trigger returning to zero, and the three-dimensional rheological feature vector corresponding to the countdown trigger returning to zero as a set of granulation batch endpoint control records and stores them in the non-volatile memory of the main controller.

12. A wet granulation endpoint intelligent prediction system based on torque dynamic characteristics, used to implement the wet granulation endpoint intelligent prediction method based on torque dynamic characteristics as described in any one of claims 1-11, characterized in that, The system includes: Torque decoupling module: used to acquire the output torque data and running timestamp of the spindle motor, and construct a one-dimensional original torque time series in chronological order. A sliding time window is applied to the one-dimensional original torque time series to perform decoupling extraction, and a multi-dimensional rheological feature vector sequence is obtained. Time verification module: It is used to input the multidimensional rheological feature vector sequence into the pre-trained endpoint prediction algorithm model, output the predicted remaining time, and perform concatenated verification on the predicted remaining time to obtain the deterministic predicted remaining time. Interruption braking module: The main controller receives the deterministic prediction of the remaining time, generates and dynamically refreshes the countdown trigger in the local system hardware timer, outputs a hardware interrupt signal when the countdown trigger returns to zero, the main controller responds to the hardware interrupt signal, sends a braking command to the frequency converter through the hardware control interface, and generates a granulation batch end-point control record.

Images