A deep learning-based continuous deodorization adaptive control method and system
The continuous deodorization adaptive control method built through deep learning solves the problem that fixed parameters are difficult to adapt to dynamic changes, achieves stable improvement in deodorization efficiency and production capacity, reduces the risk of clumping, and enhances the system's adaptability and operational stability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUANGZHOU KELISHI TECH CO LTD
- Filing Date
- 2026-02-28
- Publication Date
- 2026-06-09
AI Technical Summary
Existing continuous deodorization control methods are based on fixed process parameters or human experience, which makes it difficult to adapt to dynamic changes in the type of incoming materials, the proportion of volatile components, and environmental conditions. This results in insufficient deodorization or fluctuations in production capacity, and is also prone to problems such as material adhesion and clumping.
A deep learning-based continuous deodorization adaptive control method is adopted. By collecting sensor data from the silo and heating chamber, a time-series feature sequence is constructed. A recurrent neural network model with dual output branches is used for time series modeling to predict the deodorization completion rate and production capacity range. The residence time and feeding speed are dynamically adjusted. Combined with real-time weighing data, a fusion calculation is performed to achieve quantitative assessment and gradient adjustment of the risk of agglomeration.
It ensures deodorization efficiency and production stability under complex working conditions, reduces the probability of clumping, improves the level of automation, and avoids passive processing and equipment load fluctuations in traditional methods.
Smart Images
Figure CN122172559A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of material deodorization control technology, and more particularly to a continuous deodorization adaptive control method and system based on deep learning. Background Technology
[0002] In the fields of chemical materials, fine chemicals, polymer materials, and related continuous production, continuous deodorization processes are commonly used to remove volatile odor components from raw materials or intermediate products. These processes typically employ structures such as silos, heating chambers, and screw conveyors to allow materials to undergo desorption and removal of odor components under specific temperature, vacuum, and residence time conditions, thereby reducing the content of volatile organic compounds (VOCs).
[0003] Existing continuous deodorization control methods are mostly based on fixed process parameters or manual experience, such as pre-setting the feed rate, heating temperature, vacuum level, and material residence time in the heating chamber. Manual intervention is then performed during production based on experience or feedback from a single sensor, resulting in low automation levels. Furthermore, due to significant differences in incoming material type, volatile component ratio, material flowability, and environmental conditions, fixed-parameter control methods struggle to adapt to dynamic changes in actual production, easily leading to insufficient deodorization or capacity fluctuations. In addition, continuous deodorization often involves material adhesion and agglomeration within the heating chamber, especially under conditions of high filling rates, high viscosity, or large temperature fluctuations. Current technologies often passively address this by reducing the feed rate, increasing stirring intensity, or increasing the number of recirculation cycles, lacking quantitative assessment methods for agglomeration risk. Adjustments are made only after agglomeration has occurred, impacting deodorization efficiency and potentially causing equipment load fluctuations or even shutdowns. Summary of the Invention
[0004] To address the aforementioned problems, this invention provides a continuous odor removal adaptive control method and system based on deep learning.
[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A continuous odor removal adaptive control method based on deep learning includes the following steps: S1. Collect sensor data from the silo and heating chamber during the continuous deodorization production process, and construct a time-series feature sequence according to a preset sliding time window. The sensor data includes weighing data, temperature data, screw current, screw speed and exhaust gas VOC concentration. S2. Based on a recurrent neural network model with dual output branches, perform time series modeling and regression calculation on the time series feature sequence to obtain the predicted value of deodorization completion at the current moment and the predicted value of production capacity range within a preset time period in the future. S3. Calculate the residence time correction coefficient corresponding to the target deodorization completion based on the predicted value of deodorization completion, and determine the maximum feasible feed rate based on the predicted value of the production capacity range. Integrate the residence time correction coefficient and the maximum feasible feed rate as dynamic control parameters. S4. Based on the dynamic control parameters and combined with real-time weighing data, the instantaneous production capacity is fused and calculated to obtain the feeding frequency parameters, heating chamber transfer timing parameters and transfer weight parameters, and the feeding frequency converter and heating chamber are controlled and adjusted. S5. During the control and regulation process, the residence time distribution, filling rate, temperature, vacuum degree and torque fluctuation of the heating chamber are vectorized to obtain the state feature vector. S6. Perform regression calculation on the agglomeration probability of the state feature vector to obtain the agglomeration probability value, and adjust the feeding frequency parameter, transfer weight parameter and reflux and stirring control parameter according to the risk level corresponding to the agglomeration probability value.
[0006] Further, S1 includes the following steps: Based on the weighing sensors, temperature sensors, screw current sensors, screw speed sensors, and exhaust gas VOC concentration sensors installed in the hopper and heating chamber respectively, the sensor data in the continuous deodorization production process is periodically sampled to obtain the corresponding sensor data. Based on the sampling timestamps of various sensor data, the sensor data is time-synchronized and aligned to obtain multi-source synchronized data arranged along a unified time axis; Based on the preset sliding time window length and step size parameters, the multi-source synchronized data is windowed to obtain windowed data segments arranged in chronological order. Normalization and scaling processing are performed on the sensor data within each windowed data segment, and the data are combined according to a preset feature arrangement order to obtain a time-series feature sequence.
[0007] Further, S2 includes the following steps: Based on the time-series feature sequence, the time-series feature sequence is input into the shared recurrent layer of the recurrent neural network in chronological order, and the state dependency relationship between adjacent time steps is recursively calculated to obtain the time-series hidden state sequence corresponding to each time step. Based on the time-series hidden state sequence, select the hidden state corresponding to the current time step and input it into the odor removal completion prediction output branch. Perform linear mapping and regression calculation on the hidden state to obtain the odor removal completion prediction value at the current time. Based on the time-series hidden state sequence, a hidden state within a preset prediction time range is selected and input into the capacity interval prediction output branch. Interval regression calculation is performed on the hidden state to obtain the capacity interval prediction value within a future preset time period. The predicted values for odor removal completion and production capacity range are subjected to output format constraints and numerical validity checks, and then output.
[0008] Furthermore, the recurrent neural network model containing dual output branches is trained through the following steps: A training sample set is constructed based on sensor data from the historical continuous deodorization production process. Each training sample includes a corresponding time-series feature sequence and the true value of deodorization completion and the true value of production capacity range that are time-aligned with the time-series feature sequence. The time-series feature sequences in the training sample set are input into the recurrent neural network model, and forward recursive calculation is performed on the time-series feature sequences to obtain the predicted value of deodorization completion and the predicted value of production capacity range, respectively. The first prediction error is calculated based on the predicted value of odor removal completion and the corresponding actual value of odor removal completion; the second prediction error is calculated based on the predicted value of production capacity range and the corresponding actual value of production capacity range; and a joint loss function is constructed based on the first prediction error and the second prediction error. Backpropagation and parameter updates are performed on the network parameters of the recurrent neural network model based on the joint loss function until the joint loss function meets the preset convergence condition.
[0009] Further, the step of calculating the dwell time correction coefficient corresponding to the target odor removal completion rate based on the predicted odor removal completion rate includes: Based on the predicted value of odor removal completion at the current moment and the preset target odor removal completion, the deviation between the two is calculated to obtain the odor removal completion deviation. Based on the deviation in odor removal completion, and in conjunction with a preset correction mapping relationship, the baseline dwell time is proportionally amplified or reduced to obtain the corresponding quasi-correction coefficient. The dwell time correction coefficient is obtained by applying upper and lower limit constraints and change rate limits to the quasi-correction coefficient.
[0010] Furthermore, the process of determining the maximum feasible feed rate based on the predicted capacity range to obtain control parameters for adjusting the continuous deodorization process includes: Based on the upper limit of the production capacity, which represents the upper limit of the processing capacity within a future preset time period in the predicted production capacity range, and combined with the current effective volume of the heating chamber and the preset safe filling rate, the maximum allowable material processing volume is calculated to obtain the corresponding theoretical maximum feed volume. Based on the theoretical maximum feed rate and the length of the preset time period, the feed rate is converted to obtain the upper limit of the feed rate. The upper limit of the feeding speed is compared and adjusted with the equipment's rated feeding capacity constraint to obtain the maximum feasible feeding speed that meets the equipment's operating constraints.
[0011] Further, S4 includes the following steps: Based on real-time weighing data, the change in material weight within a preset time interval is calculated to obtain the actual instantaneous production capacity at the current moment. Based on the maximum feasible feed rate, the actual instantaneous capacity is limited, and based on the residence time correction coefficient, the limited instantaneous capacity is proportionally corrected to obtain the feed adjustment reference value. Based on the aforementioned feed adjustment reference value, and combined with the mapping relationship between the feed speed of the feed device and the frequency of the inverter, the feed adjustment reference value is converted to obtain the feed frequency parameter used to control the feed inverter. Based on the dwell time correction coefficient, the reference dwell time of the heating chamber is corrected and calculated to obtain the target dwell time, and the timing parameters for transferring the heating chamber are determined according to the target dwell time. Based on the transfer timing parameters and the real-time weighing data, the allowable material weight for a single transfer is calculated to obtain the transfer weight parameters. The feed frequency inverter is controlled and adjusted according to the feed frequency parameters, and the heating chamber is controlled and adjusted according to the transfer timing parameters and transfer weight parameters.
[0012] Further, S5 includes the following steps: Based on the operating status of the heating chamber during the control and regulation process, the residence time distribution, filling rate, temperature, vacuum degree and torque fluctuation of the material in the heating chamber are calculated and collected to obtain status data; The state data is normalized and then vectorized and concatenated according to a preset feature arrangement order to obtain a state feature vector.
[0013] Further, the process of performing regression calculations on the state feature vector to obtain the block probability value includes the following steps: Based on the statistical distribution characteristics of each dimension of the state feature vector in the historical continuous deodorization production process, a weighted linear combination calculation is performed on the state feature vector to obtain an intermediate feature value that reflects the intensity of the clumping trend. A nonlinear mapping calculation is performed on the intermediate feature value to map the intermediate feature value to a preset probability interval, thereby obtaining a block probability value that represents the current operating state.
[0014] A deep learning-based continuous odor removal adaptive control system, applied to any of the aforementioned deep learning-based continuous odor removal adaptive control methods, includes: The data acquisition module is used to collect sensor data from the silo and heating chamber during the continuous deodorization production process, and construct a time-series feature sequence according to a preset sliding time window. The sensor data includes weighing data, temperature data, screw current, screw speed and exhaust gas VOC concentration. The odor removal completion and production capacity prediction module is used to perform time series modeling and regression calculation on the time series feature sequence based on a recurrent neural network model containing dual output branches, so as to obtain the odor removal completion prediction value at the current moment and the production capacity range prediction value within a preset time period in the future. The control parameter generation module is used to calculate the residence time correction coefficient corresponding to the target deodorization completion based on the predicted value of deodorization completion, and to determine the maximum feasible feed rate based on the predicted value of the production capacity range, and to integrate the residence time correction coefficient and the maximum feasible feed rate as dynamic control parameters. The feeding and transfer control module is used to perform fusion calculation and processing on instantaneous production capacity based on the dynamic control parameters and combined with the real-time weighing data collected by the weighing sensor, to obtain feeding frequency parameters, heating chamber transfer timing parameters and transfer weight parameters, and to perform control and adjustment on the feeding frequency converter and heating chamber. The state feature construction module is used to vectorize the residence time distribution, filling rate, temperature, vacuum degree and torque fluctuation of the heating chamber during the control and regulation execution process to obtain the state feature vector; The agglomeration assessment and adjustment module is used to perform regression calculation on the agglomeration probability of the state feature vector to obtain the agglomeration probability value, and to perform gradient adjustment on the feed frequency parameter, transfer weight parameter and reflux and stirring control parameter according to the risk level corresponding to the agglomeration probability value.
[0015] The beneficial effects of this invention are as follows: By simultaneously collecting multi-source operational data such as weighing, temperature, screw current, screw speed, and VOC concentration in the exhaust gas from the silo and heating chamber during continuous deodorization production, and constructing a multi-dimensional time-series feature sequence based on a sliding time window, this invention provides a unified model of the dynamic operating conditions during production. This overcomes the problem in existing technologies that rely solely on fixed parameters or single sensor feedback, making it difficult to reflect the temporal changes in operating conditions. Furthermore, by performing time-series modeling and regression calculations on the time-series feature sequence, and simultaneously predicting the current deodorization completion rate of the material and the future short-term production capacity range, the control system can anticipate changes in deodorization effect and production capacity before actual adjustments, thereby avoiding insufficient deodorization or unstable production capacity due to fluctuations in operating conditions. Next, the residence time correction coefficient is dynamically calculated based on the predicted value of deodorization completion, and the maximum feasible feed rate is determined by combining the predicted results of the production capacity range. These two are integrated into dynamic control parameters, and the instantaneous production capacity is calculated by fusing with real-time weighing data. This achieves coordinated adjustment of feeding frequency, transfer timing, and transfer weight, thereby replacing the traditional static setting method that relies on experience. This allows the continuous deodorization process to adapt to the operating requirements of different material types and different load conditions. During the control and adjustment process, the residence time distribution, filling rate, temperature, vacuum degree, and stirring torque fluctuation of the material in the heating chamber are vectorized to form a state feature vector that reflects the material flow state and stress state. Based on this state feature vector, regression calculation is performed on the material agglomeration risk to achieve a quantitative assessment of the agglomeration trend. On this basis, this scheme adjusts the feeding, transfer, reflux, and stirring control parameters in a gradient according to the agglomeration risk level, transforming the anti-agglomeration adjustment from a passive post-event intervention to a pre-event prediction and graded suppression. This reduces the probability of agglomeration while ensuring deodorization efficiency, and improves the stability and automation level of the continuous deodorization process under complex operating conditions. Attached Figure Description
[0016] Figure 1 This is a flowchart of the steps of a continuous odor removal adaptive control method based on deep learning in this invention.
[0017] Figure 2 This is a flowchart of step S4 in this invention. Detailed Implementation
[0018] Please see Figures 1-2 As shown, this invention relates to a continuous odor removal adaptive control method based on deep learning, comprising the following steps: S1. Collect sensor data from the silo and heating chamber during the continuous deodorization production process, and construct a time-series feature sequence according to a preset sliding time window. The sensor data includes weighing data, temperature data, screw current, screw speed and exhaust gas VOC concentration. S2. Based on a recurrent neural network model with dual output branches, perform time series modeling and regression calculation on the time series feature sequence to obtain the predicted value of deodorization completion at the current moment and the predicted value of production capacity range within a preset time period in the future. S3. Calculate the residence time correction coefficient corresponding to the target deodorization completion based on the predicted value of deodorization completion, and determine the maximum feasible feed rate based on the predicted value of the production capacity range. Integrate the residence time correction coefficient and the maximum feasible feed rate as dynamic control parameters. S4. Based on the dynamic control parameters and combined with real-time weighing data, the instantaneous production capacity is fused and calculated to obtain the feeding frequency parameters, heating chamber transfer timing parameters and transfer weight parameters, and the feeding frequency converter and heating chamber are controlled and adjusted. S5. During the control and regulation process, the residence time distribution, filling rate, temperature, vacuum degree and torque fluctuation of the heating chamber are vectorized to obtain the state feature vector. S6. Perform regression calculation on the agglomeration probability of the state feature vector to obtain the agglomeration probability value, and adjust the feeding frequency parameter, transfer weight parameter and reflux and stirring control parameter according to the risk level corresponding to the agglomeration probability value.
[0019] In some embodiments, the multi-dimensional operating status of the material conveying, heating, and deodorization processes is synchronously collected first by weighing sensors, temperature sensors, screw current and speed sensors, and an online VOC detection device installed in the silo and heating chamber. Unlike the existing technology that judges based on a single process parameter or instantaneous threshold, this embodiment does not directly analyze the raw sensor data independently. Instead, it performs time alignment and windowing processing on multi-source data based on a preset sliding time window to construct a multi-dimensional time-series feature sequence reflecting the evolution of operating conditions. This explicitly preserves the temporal correlation between material load changes, thermal state evolution, and the deodorization process at the data level. The time-series feature sequence is then input into a recurrent neural network model with dual output branches for modeling. The model recursively encodes the operating condition evolution of historical time steps through a shared loop structure and performs differential regression calculations in two output branches: one output branch characterizes the deodorization completion rate of the current material under given operating conditions, and the other output branch predicts the stable capacity range that the system can bear within a preset future time period. This dual-output modeling approach differs from existing technologies that only predict a single indicator. It allows the system to simultaneously obtain joint predictions of both "odor removal effect status" and "capacity carrying capacity" within the same temporal semantic space, providing a forward-looking basis for subsequent control decisions rather than retrospective corrections. Furthermore, based on the deviation between the predicted odor removal completion rate and the target odor removal completion rate, a residence time correction coefficient is dynamically calculated. Combined with the upper limit of processing capacity reflected in the capacity range prediction results, the maximum feasible feed rate under the current operating conditions is determined. Unlike traditional methods that use fixed residence times or empirically set feed limits, this embodiment integrates the residence time correction coefficient and the maximum feasible feed rate to form dynamic control parameters that update in real time with changing operating conditions. This ensures that control variables are no longer adjusted in isolation but evolve collaboratively under the joint constraints of the odor removal target and capacity constraints. Real-time weighing data is further introduced to calculate the material weight change per unit time, obtaining an instantaneous capacity indicator reflecting the actual load level. This instantaneous capacity is then fused with the dynamic control parameters to determine the feed frequency parameter, heating chamber transfer timing parameter, and single transfer weight parameter. Through the aforementioned fusion mechanism, feeding and transfer control no longer rely on preset cycles or manual intervention, but instead adjust in conjunction with prediction results and real-time load conditions. This avoids problems such as excessively rapid feeding, insufficient residence time, or localized accumulation during sudden changes in operating conditions. Subsequently, during the continuous execution of control and adjustment, the residence time distribution, filling rate, temperature field, vacuum changes, and torque fluctuations during stirring of the material in the heating chamber are continuously collected and calculated. These state variables are then normalized and vectorized, and a state feature vector is constructed that comprehensively reflects the material flow state, heating state, and mechanical stress state.This state feature vector is not a simple set of process monitoring indicators, but rather a targeted construction based on the multi-physical quantity coupling characteristics involved in the agglomeration mechanism. This differs from existing technologies that indirectly determine agglomeration risk solely through the occurrence of stall or current surges. Finally, agglomeration probability regression calculation is performed on the state feature vector to obtain the probability value representing the material agglomeration under the current operating state. Based on the risk level corresponding to this probability value, graded and gradual adjustments are made to the feed frequency parameter, transfer weight parameter, and reflux and stirring control parameters. Through this method, agglomeration prevention and control shifts from the traditional "passive handling after agglomeration occurs" to "early identification and graded suppression of agglomeration trends." This allows the continuous deodorization process to effectively reduce the probability of agglomeration while ensuring deodorization efficiency and stable production capacity, significantly improving the system's adaptability and operational stability under complex operating conditions.
[0020] Further, S1 includes the following steps: Based on the weighing sensors, temperature sensors, screw current sensors, screw speed sensors, and exhaust gas VOC concentration sensors installed in the hopper and heating chamber respectively, the sensor data in the continuous deodorization production process is periodically sampled to obtain the corresponding sensor data. Based on the sampling timestamps of various sensor data, the sensor data is time-synchronized and aligned to obtain multi-source synchronized data arranged along a unified time axis; Based on the preset sliding time window length and step size parameters, the multi-source synchronized data is windowed to obtain windowed data segments arranged in chronological order. Normalization and scaling processing are performed on the sensor data within each windowed data segment, and the data are combined according to a preset feature arrangement order to obtain a time-series feature sequence.
[0021] Specifically, the continuous deodorization production line is equipped with weighing sensors, temperature sensors, screw current sensors, screw speed sensors, and online VOC concentration sensors in the exhaust gas at the material hopper and heating chamber locations, respectively, for monitoring the entire process of material conveying, heating status, and deodorization. The system synchronously collects the outputs of various sensors at a preset sampling period to obtain raw sensor data reflecting changes in different physical quantities. Considering the differences in sampling frequency, communication delay, and triggering mechanisms among different sensors, this embodiment does not directly use the raw data for analysis. Instead, it first performs time synchronization and alignment processing on the multi-source sensor data based on the timestamp information corresponding to each sensor data, mapping the data from different sensors to a unified time axis, thereby eliminating the time sequence offset problem caused by asynchronous sampling and forming multi-source synchronous data arranged in a unified time order. Furthermore, based on preset sliding time window length and step size parameters, the multi-source synchronous data is windowed, dividing the continuous time series into several overlapping or adjacent windowed data segments. Each windowed data segment covers a complete operational process, characterizing the dynamic features of material load changes, thermal state evolution, and deodorization processes within that time period. Compared to existing technologies that directly rely on single-moment data or simple averages for control and judgment, this windowing process explicitly preserves the trend information of the operating conditions over time, providing a stable data foundation for subsequent time-series modeling. Normalization and scaling are then performed on various sensor data within each windowed data segment, mapping physical quantities with different dimensions and numerical ranges to a unified numerical interval, preventing any single feature from having a dominant effect during model training or inference due to excessively large numerical amplitudes. After normalization, the multi-source sensor data within the window are combined and stitched together according to a preset feature arrangement order to construct a time-series feature sequence with consistent structure and fixed dimensions.
[0022] Further, S2 includes the following steps: Based on the time-series feature sequence, the time-series feature sequence is input into the shared recurrent layer of the recurrent neural network in chronological order, and the state dependency relationship between adjacent time steps is recursively calculated to obtain the time-series hidden state sequence corresponding to each time step. Based on the time-series hidden state sequence, select the hidden state corresponding to the current time step and input it into the odor removal completion prediction output branch. Perform linear mapping and regression calculation on the hidden state to obtain the odor removal completion prediction value at the current time. Based on the time-series hidden state sequence, a hidden state within a preset prediction time range is selected and input into the capacity interval prediction output branch. Interval regression calculation is performed on the hidden state to obtain the capacity interval prediction value within a future preset time period. The predicted values for odor removal completion and production capacity range are subjected to output format constraints and numerical validity checks, and then output.
[0023] It should be noted that a recurrent neural network model containing a shared recurrent layer and dual output branches is used for joint prediction modeling of the continuous deodorization process. The time-series feature sequence consists of multi-dimensional feature vectors arranged in chronological order across multiple time steps, where each time step feature vector synchronously represents the material load state, thermal state, and odor release state at the corresponding moment. The model first inputs this time-series feature sequence into the shared recurrent layer step by step, and models the feature correlations in the time dimension through recursion. The shared recurrent layer adopts a gated recurrent structure. In each time step, the current input feature vector and the hidden state of the previous time step jointly participate in the state update calculation. By introducing state retention gates and update gates, historical information and new input information are adaptively weighted and fused, thereby suppressing the impact of short-term noise fluctuations on the model output and enhancing the ability to represent the slow changing trends during the deodorization process. Through this recursive mechanism, the model can accumulate and reflect the heating history, residence time evolution, and time-dependent characteristics of volatile component release of the material in the heating chamber in the hidden state, making the hidden state no longer a simple mapping of the working conditions at a single time point, but a compressed expression of the dynamic evolution path of the continuous deodorization process. After the above recursive calculation, a temporal hidden state sequence covering the entire sliding time window is obtained. After obtaining the temporal hidden state sequence, differentiated decoding processing is performed on the hidden state sequence based on different control objectives. Specifically, for the task of predicting odor removal completion, the model selects the hidden state corresponding to the current time step from the temporal hidden state sequence as the instantaneous state representation and inputs it into the odor removal completion prediction output branch. This output branch maps the hidden state to a predicted odor removal completion value through linear mapping and continuous regression calculations, characterizing the removal progress level of volatile odor components in the material under the current operating conditions. Since this hidden state has integrated the operating condition information of multiple previous time steps, the predicted odor removal completion value can simultaneously reflect the current instantaneous state and the cumulative effect of historical operating conditions, avoiding local judgment bias caused by relying solely on a single VOC concentration or temperature signal. Meanwhile, for the task of predicting production capacity, the model does not directly perform single-point prediction based on the hidden state of the current time step, but instead selects multiple hidden states covering a preset prediction time range from the temporal hidden state sequence and inputs them as a whole into the production capacity interval prediction output branch. This output branch performs interval regression calculations on the multiple hidden states. By simultaneously learning the upper and lower bound constraints of capacity changes within a future time period, it outputs a predicted capacity range representing the system's stable operating capacity within a preset future time period. Compared to traditional capacity forecasting methods that only output a single predicted value, this interval regression result can explicitly characterize the uncertainties of future operating conditions, providing robust boundary conditions for subsequent feed rate limiting and safe filling control.Further format constraints and numerical validity checks are performed on the predicted values for odor removal completion and production capacity range. This includes imposing physical feasible range restrictions on the prediction results, suppressing abnormal jumps, and smoothing the output results to ensure the stability and usability of the prediction results in the control loop. The verified prediction results are used as core input parameters for subsequent residence time correction coefficient calculation and maximum feasible feed rate determination.
[0024] Furthermore, the recurrent neural network model containing dual output branches is trained through the following steps: A training sample set is constructed based on sensor data from the historical continuous deodorization production process. Each training sample includes a corresponding time-series feature sequence and the true value of deodorization completion and the true value of production capacity range that are time-aligned with the time-series feature sequence. The time-series feature sequences in the training sample set are input into the recurrent neural network model, and forward recursive calculation is performed on the time-series feature sequences to obtain the predicted value of deodorization completion and the predicted value of production capacity range, respectively. The first prediction error is calculated based on the predicted value of odor removal completion and the corresponding actual value of odor removal completion; the second prediction error is calculated based on the predicted value of production capacity range and the corresponding actual value of production capacity range; and a joint loss function is constructed based on the first prediction error and the second prediction error. Backpropagation and parameter updates are performed on the network parameters of the recurrent neural network model based on the joint loss function until the joint loss function meets the preset convergence condition.
[0025] In some embodiments, firstly, a training sample set is constructed based on multi-source sensor data accumulated over a long period during the historical continuous deodorization production process. For each training sample, the corresponding temporal feature sequence is extracted from the historical operating data according to the same sliding time window and time alignment rules as in step S1, ensuring consistency in the input data distribution between the model training stage and the online prediction stage. The true value of deodorization completion is not directly determined by a single instantaneous VOC concentration value, but is comprehensively calculated based on the actual residence time of the material in the heating chamber, the rate of change of VOC concentration in and out of the chamber, and the exhaust gas emission characteristics during the stable operation period. This comprehensive calculation characterizes the actual removal degree of odor components within the time window, thus avoiding noise interference caused by relying solely on instantaneous detection values as monitoring signals. The true value of the production capacity range is statistically estimated based on the actual feed volume, transfer frequency, and stable operation constraints of the equipment within a future preset time period from historical operating records, forming range labels with upper and lower bounds. The true value of the production capacity range constructed in this way can realistically reflect the feasible operating boundaries of the process under different operating conditions, rather than an idealized point-value production capacity indicator. During model training, the temporal feature sequences from the training sample set are input into the recurrent neural network model in chronological order, and forward recursive calculations are performed on these sequences. A shared recurrent layer extracts the implicit dynamic representation of the temporal features through a time-step recursive approach, and feeds the resulting temporal hidden states into the deodorization completion prediction output branch and the production capacity range prediction output branch, respectively. This allows for the simultaneous generation of deodorization completion prediction values and production capacity range prediction values within the same temporal representation space. By sharing the underlying recurrent structure, the model can automatically learn common temporal patterns affecting deodorization efficiency and production capacity changes during training, avoiding data redundancy and prediction inconsistencies introduced by building independent models for different prediction targets. In the loss function construction stage, the first prediction error and the second prediction error are calculated separately to address the differences in the physical meaning of the dual-output prediction targets. The system employs a first prediction error to measure the deviation between the predicted odor removal completion rate and the corresponding actual odor removal completion rate, thereby constraining the model's fitting accuracy to the odor removal process. The second prediction error measures the deviation between the predicted capacity range and the corresponding actual capacity range, and by simultaneously constraining the upper and lower bounds of the prediction range, the model learns to reasonably characterize the uncertainty of capacity fluctuations during training. Subsequently, the system constructs a joint loss function based on the first and second prediction errors, and weights and fuses the two types of errors by introducing preset weight coefficients. This balances the prediction accuracy of odor removal effect with the prediction stability of capacity range during model training, preventing one type of task from dominating the other during gradient updates. Finally, backpropagation is performed on the network parameters of the recurrent neural network model based on the joint loss function, and the model parameters are iteratively updated according to a preset learning rate strategy.During training, the trend of the joint loss function is continuously monitored. The training process is terminated when the function meets the preset convergence condition or reaches a stable decreasing range, resulting in the final recurrent neural network model used for online prediction. The model obtained through this training method not only maintains prediction consistency on historical data but also demonstrates good generalization ability to changes in operating conditions during actual continuous deodorization production.
[0026] Further, the step of calculating the dwell time correction coefficient corresponding to the target odor removal completion rate based on the predicted odor removal completion rate includes: Based on the predicted value of odor removal completion at the current moment and the preset target odor removal completion, the deviation between the two is calculated to obtain the odor removal completion deviation. Based on the deviation in odor removal completion, and in conjunction with a preset correction mapping relationship, the baseline dwell time is proportionally amplified or reduced to obtain the corresponding quasi-correction coefficient. The dwell time correction coefficient is obtained by applying upper and lower limit constraints and change rate limits to the quasi-correction coefficient.
[0027] In some embodiments, at the current moment, the predicted deodorization completion rate output by the recurrent neural network model is first read and compared with the preset target deodorization completion rate to calculate the deviation between the two. This deviation rate characterizes the degree of inadequacy or redundancy of the material deodorization status relative to the target status under the current operating conditions. When the deviation rate is positive, it indicates that the current predicted deodorization completion rate is lower than the target requirement, and the effective resolution of the material in the heating chamber is insufficient; when the deviation rate is negative, it indicates that there is redundancy in the deodorization completion rate, and continuing to extend the residence time will lead to a decrease in production capacity without generating significant deodorization benefits. Based on this, according to the deodorization completion rate deviation rate and combined with a preset correction mapping relationship, a proportional correction calculation is performed on the benchmark residence time. The correction mapping relationship is not a one-to-one correspondence, but is pre-established based on the response characteristics between changes in deodorization completion rate and residence time adjustments in historical production data. This ensures that only a small correction is triggered when the deviation rate is small, while the correction magnitude is gradually amplified when the deviation rate gradually increases, thereby avoiding frequent and large changes in residence time due to prediction errors or short-term fluctuations. The quasi-correction coefficient obtained through this mapping relationship reflects the proportional trend of whether the residence time should be extended or shortened relative to the baseline value under the current operating conditions. To further ensure the operational stability of the continuous deodorization process, upper and lower limit constraints and change rate limits are applied to the quasi-correction coefficient after its generation. On the one hand, by setting maximum and minimum values for the correction coefficient, it is prevented that the residence time will be adjusted beyond the safe operating range of the equipment under abnormal predictions or extreme operating conditions. On the other hand, by limiting the change amplitude of the correction coefficient within adjacent control cycles, it is possible to suppress drastic oscillations of the residence time on the time axis and avoid triggering sudden changes in the heating chamber filling rate, material flow instability, or chain disturbances in the downstream control chain.
[0028] Furthermore, the process of determining the maximum feasible feed rate based on the predicted capacity range to obtain control parameters for adjusting the continuous deodorization process includes: Based on the upper limit of the production capacity, which represents the upper limit of the processing capacity within a future preset time period in the predicted production capacity range, and combined with the current effective volume of the heating chamber and the preset safe filling rate, the maximum allowable material processing volume is calculated to obtain the corresponding theoretical maximum feed volume. Based on the theoretical maximum feed rate and the length of the preset time period, the feed rate is converted to obtain the upper limit of the feed rate. The upper limit of the feeding speed is compared and adjusted with the equipment's rated feeding capacity constraint to obtain the maximum feasible feeding speed that meets the equipment's operating constraints.
[0029] Specifically, firstly, an upper limit value representing the upper limit of processing capacity within a preset future time period is extracted from the predicted capacity range. This upper limit value reflects the maximum material processing capacity that the heating chamber can stably bear and complete the deodorization process within a future period under the combined effects of current material characteristics, thermal conditions, and operating status. Unlike traditional methods that rely solely on historical average capacity or instantaneous load, this upper limit value includes predictive information about upcoming changes in operating conditions, thereby preventing the feeding strategy from lagging behind actual changes in processing capacity. After obtaining the upper limit value, the maximum material processing volume allowed to enter the heating chamber and participate in the deodorization process is calculated by combining the current effective volume of the heating chamber and the preset safe filling rate. The safe filling rate is used to constrain the material filling state inside the heating chamber, ensuring that it can maintain stable heat transfer efficiency, gas escape channels, and agitated flow patterns even under high load operating conditions, thereby avoiding the risk of localized overheating, material stagnation, or agglomeration caused by overfilling. The theoretical maximum feed rate calculated using this constraint essentially defines the matching relationship between the total material input and the deodorization processing capacity within the predicted time window. Then, based on the theoretical maximum feed rate and the corresponding preset time period length, a conversion process is performed on the feed rate, transforming the "maximum total material that can be processed within the time period" into an upper limit value for the feed rate under continuous control. This conversion process allows the predicted capacity constraint to directly apply to the real-time feed control, avoiding its existence as a cumulative quantity that cannot participate in instantaneous control decisions. Furthermore, the upper limit value for the feed rate is compared with the equipment's rated feed capacity constraint, and a trimming process is performed. This trimming process ensures that the feed rate does not exceed the maximum operating capacity allowed by the equipment structure, drive system, and safety regulations, while also preventing blindly increasing the feed rate when the predicted capacity is high, which could lead to equipment overload or mechanical shock.
[0030] Further, S4 includes the following steps: Based on real-time weighing data, the change in material weight within a preset time interval is calculated to obtain the actual instantaneous production capacity at the current moment. Based on the maximum feasible feed rate, the actual instantaneous capacity is limited, and based on the residence time correction coefficient, the limited instantaneous capacity is proportionally corrected to obtain the feed adjustment reference value. Based on the aforementioned feed adjustment reference value, and combined with the mapping relationship between the feed speed of the feed device and the frequency of the inverter, the feed adjustment reference value is converted to obtain the feed frequency parameter used to control the feed inverter. Based on the dwell time correction coefficient, the reference dwell time of the heating chamber is corrected and calculated to obtain the target dwell time, and the timing parameters for transferring the heating chamber are determined according to the target dwell time. Based on the transfer timing parameters and the real-time weighing data, the allowable material weight for a single transfer is calculated to obtain the transfer weight parameters. The feed frequency inverter is controlled and adjusted according to the feed frequency parameters, and the heating chamber is controlled and adjusted according to the transfer timing parameters and transfer weight parameters.
[0031] In some embodiments, the instantaneous capacity is first limited based on the maximum feasible feed rate predicted from the aforementioned capacity range. This limiting is not a simple upper limit cutoff, but rather constrains the current instantaneous capacity within the safe processing capacity allowed by the prediction model, thereby preventing the feed strategy from further amplifying the load during sudden increases in instantaneous flow, which could lead to excessively high heating chamber filling rates or insufficient deodorization. Subsequently, the system introduces a residence time correction coefficient calculated from the deodorization completion prediction result to perform a proportional correction on the limited instantaneous capacity. Through this proportional correction, the feed adjustment is not only constrained by the upper limit of processing capacity, but can also actively compensate based on the degree to which the current deodorization completion deviates from the target value: when the predicted deodorization completion is low, the correction coefficient prompts the system to lower the equivalent feed reference value to extend the residence time of the material in the heating chamber; when the predicted deodorization completion is close to or reaches the target value, the feed reference level is allowed to be increased within a safe range, thereby improving the overall capacity utilization rate. The feed adjustment reference value obtained after the above dual constraints and corrections essentially forms an intermediate control quantity that simultaneously considers deodorization effect and capacity stability. Further, based on the structural characteristics of the feeding device and the mapping relationship between its feeding speed and the inverter output frequency, the feeding adjustment reference value is converted. This mapping relationship can be established through calibration experiments or historical operating data, transforming the mass flow control problem into an inverter frequency control problem, thereby obtaining the feeding frequency parameters used to directly drive the feeding inverter. In this way, the capacity adjustment results at the algorithm level can be seamlessly mapped to the underlying actuators, avoiding the involvement of human experience. Based on the same residence time correction coefficient, the baseline residence time of the heating chamber is dynamically corrected and calculated to obtain the target residence time. This target residence time reflects the optimal material residence scale required to meet the target deodorization completion under the current predicted operating conditions. Further, based on this target residence time, combined with the volume characteristics of the heating chamber and the material conveying cycle time, the timing parameters for transferring the heating chamber are determined, ensuring that the material is transferred in an orderly manner after reaching the target residence conditions, avoiding insufficient deodorization due to premature transfer or the risk of local stagnation and agglomeration due to excessively late transfer. After determining the timing of the transfer, the system further utilizes real-time weighing data to calculate the allowable weight of material to be transferred during a single transfer, obtaining the transfer weight parameter. This parameter constrains the scale of material transferred each time, ensuring that the transfer operation is consistent with the current fill rate, instantaneous capacity, and residence time requirements, thereby preventing disruption of the thermal balance and flow state inside the heating chamber due to excessive transfer volume in a single operation. Finally, the system performs real-time control and adjustment of the feed inverter based on the feed frequency parameter, and performs coordinated control of the heating chamber based on the transfer timing parameter and transfer weight parameter.
[0032] Further, S5 includes the following steps: Based on the operating status of the heating chamber during the control and regulation process, the residence time distribution, filling rate, temperature, vacuum degree and torque fluctuation of the material in the heating chamber are calculated and collected to obtain status data; The state data is normalized and then vectorized and concatenated according to a preset feature arrangement order to obtain a state feature vector.
[0033] Specifically, during the process of issuing and executing parameters such as feeding frequency, transfer timing, and transfer weight, the system continuously observes the actual operating status of the material in the heating chamber. This includes the distribution characteristics of the material's residence time in the chamber at different time periods, the current material filling rate, changes in the heating temperature field, vacuum fluctuations, and the torque changes corresponding to the screw or stirring mechanism during the stirring process. Further, based on a preset time window, statistical calculations are performed on each state quantity. For example, residence time is characterized using distributional features, filling rate, temperature, and vacuum are characterized using interval averages and fluctuation amplitudes, and the torque signal's fluctuation intensity and trend are extracted, thereby reducing the interference of instantaneous noise on subsequent model calculations. The state data generated after the above processing exhibits significant differences in numerical scale, physical dimensions, and range of variation. Therefore, the system further performs normalization processing on the state data, mapping state features from different sources and with different dimensions to a unified numerical space to avoid feature imbalance caused by inconsistent dimensions in subsequent regression calculations.
[0034] Further, the process of performing regression calculations on the state feature vector to obtain the block probability value includes the following steps: Based on the statistical distribution characteristics of each dimension of the state feature vector in the historical continuous deodorization production process, a weighted linear combination calculation is performed on the state feature vector to obtain an intermediate feature value that reflects the intensity of the clumping trend. A nonlinear mapping calculation is performed on the intermediate feature value to map the intermediate feature value to a preset probability interval, thereby obtaining a block probability value that represents the current operating state.
[0035] In some embodiments, firstly, based on the recorded state feature vector samples and corresponding clumping occurrences from the historical continuous deodorization production process, the statistical distribution characteristics of each dimension of the state feature vector are analyzed before clumping, during clumping, and under normal operating conditions, thereby obtaining the sensitivity of each dimension of the feature to clumping formation. Based on the statistical analysis results, corresponding weight coefficients are assigned to each dimension of the state feature vector, so that different features can participate in the calculation according to their influence on clumping behavior in subsequent calculations. During real-time operation, the system inputs the state feature vector obtained at the current moment into the clumping trend regression calculation module, performs a weighted linear combination calculation on each dimension of the state feature vector and its corresponding weight coefficient, and obtains an intermediate feature value used to characterize the intensity of the clumping development trend under the current operating condition. Then, a nonlinear mapping calculation is performed on the intermediate feature value to map it to a preset probability interval, thereby obtaining a clumping probability value characterizing the possibility of clumping occurring under the current operating condition. Through this nonlinear mapping process, the clumping risk under different operating conditions can be expressed in a unified probability form, making the clumping risk assessment results have good numerical stability and comparability. The above method transforms the original process of judging agglomeration, which relied on manual experience or post-hoc observation, into a continuous probability regression calculation process based on state feature vectors, enabling early quantitative assessment of agglomeration trends. It not only identifies rising risk trends before agglomeration becomes apparent but also provides clear quantitative basis for gradient adjustments of feeding, transfer, reflux, and mixing parameters.
[0036] The present invention also includes a deep learning-based continuous odor removal adaptive control system, comprising: The data acquisition module is used to collect sensor data from the silo and heating chamber during the continuous deodorization production process, and construct a time-series feature sequence according to a preset sliding time window. The sensor data includes weighing data, temperature data, screw current, screw speed and exhaust gas VOC concentration. The odor removal completion and production capacity prediction module is used to perform time series modeling and regression calculation on the time series feature sequence based on a recurrent neural network model containing dual output branches, so as to obtain the odor removal completion prediction value at the current moment and the production capacity range prediction value within a preset time period in the future. The control parameter generation module is used to calculate the residence time correction coefficient corresponding to the target deodorization completion based on the predicted value of deodorization completion, and to determine the maximum feasible feed rate based on the predicted value of the production capacity range, and to integrate the residence time correction coefficient and the maximum feasible feed rate as dynamic control parameters. The feeding and transfer control module is used to perform fusion calculation and processing on instantaneous production capacity based on the dynamic control parameters and combined with the real-time weighing data collected by the weighing sensor, to obtain feeding frequency parameters, heating chamber transfer timing parameters and transfer weight parameters, and to perform control and adjustment on the feeding frequency converter and heating chamber. The state feature construction module is used to vectorize the residence time distribution, filling rate, temperature, vacuum degree and torque fluctuation of the heating chamber during the control and regulation execution process to obtain the state feature vector; The agglomeration assessment and adjustment module is used to perform regression calculation on the agglomeration probability of the state feature vector to obtain the agglomeration probability value, and to perform gradient adjustment on the feed frequency parameter, transfer weight parameter and reflux and stirring control parameter according to the risk level corresponding to the agglomeration probability value.
[0037] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.
Claims
1. A continuous odor removal adaptive control method based on deep learning, characterized in that, Includes the following steps: S1. Collect sensor data from the silo and heating chamber during the continuous deodorization production process, and construct a time-series feature sequence according to a preset sliding time window. The sensor data includes weighing data, temperature data, screw current, screw speed and exhaust gas VOC concentration. S2. Based on a recurrent neural network model with dual output branches, perform time series modeling and regression calculation on the time series feature sequence to obtain the predicted value of deodorization completion at the current moment and the predicted value of production capacity range within a preset time period in the future. S3. Calculate the residence time correction coefficient corresponding to the target deodorization completion based on the predicted value of deodorization completion, and determine the maximum feasible feed rate based on the predicted value of the production capacity range. Integrate the residence time correction coefficient and the maximum feasible feed rate as dynamic control parameters. S4. Based on the dynamic control parameters and combined with real-time weighing data, the instantaneous production capacity is fused and calculated to obtain the feeding frequency parameters, heating chamber transfer timing parameters and transfer weight parameters, and the feeding frequency converter and heating chamber are controlled and adjusted. S5. During the control and regulation process, the residence time distribution, filling rate, temperature, vacuum degree and torque fluctuation of the heating chamber are vectorized to obtain the state feature vector. S6. Perform regression calculation on the agglomeration probability of the state feature vector to obtain the agglomeration probability value, and adjust the feeding frequency parameter, transfer weight parameter and reflux and stirring control parameter according to the risk level corresponding to the agglomeration probability value.
2. The continuous odor removal adaptive control method based on deep learning according to claim 1, characterized in that, S1 includes the following steps: Based on the weighing sensors, temperature sensors, screw current sensors, screw speed sensors, and exhaust gas VOC concentration sensors installed in the hopper and heating chamber respectively, the sensor data in the continuous deodorization production process is periodically sampled to obtain the corresponding sensor data. Based on the sampling timestamps of various sensor data, the sensor data is time-synchronized and aligned to obtain multi-source synchronized data arranged along a unified time axis; Based on the preset sliding time window length and step size parameters, the multi-source synchronized data is windowed to obtain windowed data segments arranged in chronological order. Normalization and scaling processing are performed on the sensor data within each windowed data segment, and the data are combined according to a preset feature arrangement order to obtain a time-series feature sequence.
3. The continuous odor removal adaptive control method based on deep learning according to claim 1, characterized in that, S2 includes the following steps: Based on the time-series feature sequence, the time-series feature sequence is input into the shared recurrent layer of the recurrent neural network in chronological order, and the state dependency relationship between adjacent time steps is recursively calculated to obtain the time-series hidden state sequence corresponding to each time step. Based on the time-series hidden state sequence, select the hidden state corresponding to the current time step and input it into the odor removal completion prediction output branch. Perform linear mapping and regression calculation on the hidden state to obtain the odor removal completion prediction value at the current time. Based on the time-series hidden state sequence, a hidden state within a preset prediction time range is selected and input into the capacity interval prediction output branch. Interval regression calculation is performed on the hidden state to obtain the capacity interval prediction value within a future preset time period. The predicted values for odor removal completion and production capacity range are subjected to output format constraints and numerical validity checks, and then output.
4. The continuous odor removal adaptive control method based on deep learning according to claim 1, characterized in that, The recurrent neural network model containing dual output branches is trained through the following steps: A training sample set is constructed based on sensor data from the historical continuous deodorization production process. Each training sample includes a corresponding time-series feature sequence and the true value of deodorization completion and the true value of production capacity range that are time-aligned with the time-series feature sequence. The time-series feature sequences in the training sample set are input into the recurrent neural network model, and forward recursive calculation is performed on the time-series feature sequences to obtain the predicted value of deodorization completion and the predicted value of production capacity range, respectively. The first prediction error is calculated based on the predicted value of odor removal completion and the corresponding actual value of odor removal completion; the second prediction error is calculated based on the predicted value of production capacity range and the corresponding actual value of production capacity range; and a joint loss function is constructed based on the first prediction error and the second prediction error. Backpropagation and parameter updates are performed on the network parameters of the recurrent neural network model based on the joint loss function until the joint loss function meets the preset convergence condition.
5. The continuous odor removal adaptive control method based on deep learning according to claim 1, characterized in that, The step of calculating the dwell time correction coefficient corresponding to the target odor removal completion rate based on the predicted odor removal completion rate includes: Based on the predicted value of odor removal completion at the current moment and the preset target odor removal completion, the deviation between the two is calculated to obtain the odor removal completion deviation. Based on the deviation in odor removal completion, and in conjunction with a preset correction mapping relationship, the baseline dwell time is proportionally amplified or reduced to obtain the corresponding quasi-correction coefficient. The dwell time correction coefficient is obtained by applying upper and lower limit constraints and change rate limits to the quasi-correction coefficient.
6. The continuous odor removal adaptive control method based on deep learning according to claim 1, characterized in that, The process of determining the maximum feasible feed rate based on the predicted capacity range, and obtaining control parameters for adjusting the continuous deodorization process, includes: Based on the upper limit of the production capacity, which represents the upper limit of the processing capacity within a future preset time period in the predicted production capacity range, and combined with the current effective volume of the heating chamber and the preset safe filling rate, the maximum allowable material processing volume is calculated to obtain the corresponding theoretical maximum feed volume. Based on the theoretical maximum feed rate and the length of the preset time period, the feed rate is converted to obtain the upper limit of the feed rate. The upper limit of the feeding speed is compared and adjusted with the equipment's rated feeding capacity constraint to obtain the maximum feasible feeding speed that meets the equipment's operating constraints.
7. The continuous odor removal adaptive control method based on deep learning according to claim 6, characterized in that, S4 includes the following steps: Based on real-time weighing data, the change in material weight within a preset time interval is calculated to obtain the actual instantaneous production capacity at the current moment. Based on the maximum feasible feed rate, the actual instantaneous capacity is limited, and based on the residence time correction coefficient, the limited instantaneous capacity is proportionally corrected to obtain the feed adjustment reference value. Based on the aforementioned feed adjustment reference value, and combined with the mapping relationship between the feed speed of the feed device and the frequency of the inverter, the feed adjustment reference value is converted to obtain the feed frequency parameter used to control the feed inverter. Based on the dwell time correction coefficient, the reference dwell time of the heating chamber is corrected and calculated to obtain the target dwell time, and the timing parameters for transferring the heating chamber are determined according to the target dwell time. Based on the transfer timing parameters and the real-time weighing data, the allowable material weight for a single transfer is calculated to obtain the transfer weight parameters. The feed frequency inverter is controlled and adjusted according to the feed frequency parameters, and the heating chamber is controlled and adjusted according to the transfer timing parameters and transfer weight parameters.
8. The continuous odor removal adaptive control method based on deep learning according to claim 1, characterized in that, S5 includes the following steps: Based on the operating status of the heating chamber during the control and regulation process, the residence time distribution, filling rate, temperature, vacuum degree and torque fluctuation of the material in the heating chamber are calculated and collected to obtain status data; The state data is normalized and then vectorized and concatenated according to a preset feature arrangement order to obtain a state feature vector.
9. The continuous odor removal adaptive control method based on deep learning according to claim 8, characterized in that, The process of regressing the block probability onto the state feature vector to obtain the block probability value includes the following steps: Based on the statistical distribution characteristics of each dimension of the state feature vector in the historical continuous deodorization production process, a weighted linear combination calculation is performed on the state feature vector to obtain an intermediate feature value that reflects the intensity of the clumping trend. A nonlinear mapping calculation is performed on the intermediate feature value to map the intermediate feature value to a preset probability interval, thereby obtaining a block probability value that represents the current operating state.
10. A deep learning-based continuous odor removal adaptive control system, applied to the deep learning-based continuous odor removal adaptive control method described in any one of claims 1-9, characterized in that, include: The data acquisition module is used to collect sensor data from the silo and heating chamber during the continuous deodorization production process, and construct a time-series feature sequence according to a preset sliding time window. The sensor data includes weighing data, temperature data, screw current, screw speed and exhaust gas VOC concentration. The odor removal completion and production capacity prediction module is used to perform time series modeling and regression calculation on the time series feature sequence based on a recurrent neural network model containing dual output branches, so as to obtain the odor removal completion prediction value at the current moment and the production capacity range prediction value within a preset time period in the future. The control parameter generation module is used to calculate the residence time correction coefficient corresponding to the target deodorization completion based on the predicted value of deodorization completion, and to determine the maximum feasible feed rate based on the predicted value of the production capacity range, and to integrate the residence time correction coefficient and the maximum feasible feed rate as dynamic control parameters. The feeding and transfer control module is used to perform fusion calculation and processing on instantaneous production capacity based on the dynamic control parameters and combined with the real-time weighing data collected by the weighing sensor, to obtain feeding frequency parameters, heating chamber transfer timing parameters and transfer weight parameters, and to perform control and adjustment on the feeding frequency converter and heating chamber. The state feature construction module is used to vectorize the residence time distribution, filling rate, temperature, vacuum degree and torque fluctuation of the heating chamber during the control and regulation execution process to obtain the state feature vector; The agglomeration assessment and adjustment module is used to perform regression calculation on the agglomeration probability of the state feature vector to obtain the agglomeration probability value, and to perform gradient adjustment on the feed frequency parameter, transfer weight parameter and reflux and stirring control parameter according to the risk level corresponding to the agglomeration probability value.