Control methods and systems for intelligent dust-free pulverization in pharmaceutical manufacturing

Abstract

This application relates to the field of pharmaceutical production control technology, and discloses a control method and system for intelligent dust-free pulverization in pharmaceutical production. The method includes: collecting material characteristic parameters to construct a feature vector and matching it with a feature library to generate fingerprint data; selecting a prediction model and configuring initial parameters accordingly; collecting real-time operating data and inputting it into the model to obtain a predicted concentration sequence and calculating a feedforward control quantity; executing control and performing feedback correction. This application achieves active dust suppression based on time-series prediction and adaptive model selection based on material characteristics.

Description

Technical Field

[0001] This application relates to the field of pharmaceutical production control technology, and in particular to a control method and system for intelligent dust-free pulverization in pharmaceutical production. Background Technology

[0002] The pulverization process of active pharmaceutical ingredients (APIs) and excipients is a crucial step in pharmaceutical manufacturing, directly impacting drug dissolution, bioavailability, and quality uniformity. Traditional pulverization control methods rely heavily on manual experience to set pulverization parameters, adjusting pulverizer speed, feed rate, and dust collector fan power to control pulverization efficiency and dust dispersion. In recent years, with increasingly stringent GMP standards for pharmaceutical cleanliness, pharmaceutical manufacturing workshops have placed stricter demands on dust control during pulverization. Current technologies employ dust concentration sensors to monitor dust concentration in the pulverization chamber in real time. When the detected dust concentration exceeds a safe threshold, the control system automatically adjusts the dust collector fan power or reduces the pulverizer speed to keep the dust concentration within a safe range. Some advanced control systems also incorporate PID or fuzzy control algorithms to calculate control parameters based on dust concentration deviations, forming a closed-loop feedback control.

[0003] However, existing drug pulverization control methods still have significant shortcomings. First, existing methods employ a passive response mode of "initiating control after detecting excessive dust levels." From the moment the dust concentration sensor detects the excessive signal to the control system responding and the actuator adjusting its actions, there is a control lag of 2 to 5 seconds. During this lag, the dust concentration continues to rise or even significantly exceeds the limit, failing to meet the stringent dust control requirements for highly active drugs such as cephalosporins and penicillins. Second, existing control methods mostly use a single, universal control model, failing to fully consider the significant differences in the physical properties of different drug materials. Microcrystalline cellulose and raw drug crystals differ greatly in parameters such as hardness, brittleness, and moisture content, but existing methods use the same control strategy and parameters for all materials. This results in poor control of materials that easily generate dust, while over-controlling materials that do not easily generate dust, leading to energy waste. Furthermore, although existing PID control algorithms or fuzzy control algorithms can adjust according to real-time deviations, these algorithms are weakly adaptable to the nonlinear characteristics and time-varying properties of the crushing process. When material batches change or the wear condition of the equipment changes, the control parameters cannot be automatically optimized and adjusted, and the control accuracy gradually decreases. Summary of the Invention

[0004] This application provides a control method and system for intelligent dust-free pulverization in pharmaceutical production, which solves the problems in the prior art of pulverization control, such as lag response leading to instantaneous dust exceedance, inability to adaptively match control strategies for different materials, inaccurate prediction of new materials, and long-term degradation of model performance. It realizes active dust suppression based on time-series prediction and adaptive model selection based on material characteristics.

[0005] Firstly, this application provides a control method for intelligent dust-free pulverization in pharmaceutical manufacturing, the control method comprising:

[0006] Step S1: Collect the moisture content, hardness, bulk density and brittleness coefficient of the material to be crushed to construct a material feature vector, and perform cosine similarity calculation between the material feature vector and the material feature library to generate material feature fingerprint data;

[0007] Step S2: Select a dust concentration prediction model and configure initial parameters for crushing control based on the material characteristic fingerprint data;

[0008] Step S3: Collect real-time operating parameters of the crushing process and form a multimodal feature vector sequence with the material feature vector. Input the multimodal feature vector sequence into the dust concentration prediction model to obtain the predicted dust concentration sequence. Iterate through the predicted dust concentration sequence to determine whether there is a predicted value that exceeds the safety threshold. When there is a predicted value that exceeds the standard, calculate the speed reduction, dust removal power increase and feed rate reduction coefficient based on the deviation between the predicted value that exceeds the standard and the safety threshold as feedforward control quantities.

[0009] Step S4: Adjust the parameters of the crushing equipment according to the feedforward control quantity, and collect the actual dust concentration and the predicted dust concentration sequence to calculate the deviation and perform feedback correction.

[0010] Secondly, this application provides a control system for intelligent dust-free pulverization in pharmaceutical production, the control system comprising:

[0011] The calculation module is used to collect the moisture content, hardness, bulk density and brittleness coefficient of the material to be crushed to construct a material feature vector, and to perform cosine similarity calculation between the material feature vector and the material feature library to generate material feature fingerprint data.

[0012] The configuration module is used to select a dust concentration prediction model and configure the initial parameters for crushing control based on the material characteristic fingerprint data.

[0013] The input module is used to collect real-time operating parameters of the crushing process and form a multimodal feature vector sequence with the material feature vector. The multimodal feature vector sequence is input into the dust concentration prediction model to obtain a predicted dust concentration sequence. The predicted dust concentration sequence is traversed to determine whether there is a predicted value that exceeds the safety threshold. When there is a predicted value that exceeds the standard, the speed reduction, dust removal power increase and feed rate reduction coefficient are calculated based on the deviation between the predicted value and the safety threshold as feedforward control quantities.

[0014] The correction module is used to adjust the parameters of the crushing equipment according to the feedforward control quantity, and to collect the actual dust concentration and the predicted dust concentration sequence to calculate the deviation and perform feedback correction.

[0015] Thirdly, a control device for intelligent dust-free pulverization in pharmaceutical production is provided, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the control device for intelligent dust-free pulverization in pharmaceutical production to execute the aforementioned control method for intelligent dust-free pulverization in pharmaceutical production.

[0016] Fourthly, a computer-readable storage medium is provided, wherein instructions are stored therein, which, when executed on a computer, cause the computer to perform the aforementioned intelligent dust-free pulverization control method for pharmaceutical production.

[0017] The technical solution provided in this application constructs a material feature vector by collecting the moisture content, hardness, bulk density, and brittleness coefficient of the material to be crushed before the crushing operation begins. This material feature vector is then compared with a material feature library using cosine similarity calculation to generate material feature fingerprint data. This solves the problem of poor material adaptability caused by using a single general control strategy for all materials in existing technologies. Different dust concentration prediction models are selected based on the similarity values ​​in the material feature fingerprint data. A dedicated model is used for materials with high similarity, model fusion is performed for materials with medium similarity, and a general model is used for materials with low similarity. This achieves accurate matching between the prediction model and material characteristics, significantly improving the prediction accuracy for different materials. Simultaneously, the initial speed of the crushing motor, the initial power of the dust collector fan, and the initial feed rate are configured according to the specific values ​​in the material feature vector, ensuring that the control parameters are in an optimal state from the start of crushing. This avoids the problem of unstable dust control in the early stages of crushing caused by using fixed initial parameters in existing technologies. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are 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 This is a schematic diagram of one embodiment of the intelligent dust-free pulverization control method for pharmaceutical production in this application.

[0020] Figure 2 This is a schematic diagram of one embodiment of the intelligent dust-free pulverizing control system for pharmaceutical production in this application.

[0021] Figure 3 This is a schematic block diagram of the structure of the intelligent dust-free pulverizing control equipment for pharmaceutical production in an embodiment of the present invention. Detailed Implementation

[0022] This application provides a control method and system for intelligent dust-free pulverization in pharmaceutical manufacturing. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0023] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the intelligent dust-free pulverization control method for pharmaceutical production in this application includes:

[0024] Step S1: Collect the moisture content, hardness, bulk density and brittleness coefficient of the material to be crushed to construct a material feature vector, and calculate the cosine similarity between the material feature vector and the material feature library to generate material feature fingerprint data;

[0025] Specifically, the moisture content is determined by measuring the intensity of the moisture absorption peak in a near-infrared spectroscopy analyzer; the hardness value is calculated by measuring the diagonal length of the indentation after applying a standard load with a Vickers hardness tester; the bulk density is obtained by measuring the volume after filling a standard container with a certain mass of material; and the brittleness coefficient is calculated by the mass ratio of crushed particles after the material is rotated and tumbled in a brittleness tester. Cosine similarity calculation involves performing a dot product operation between the current material feature vector and each standard vector in the feature library, while simultaneously calculating the modulus of each vector. The difference between the two is used to obtain the similarity value. The closer this value is to 1, the more similar the material characteristics. The material feature fingerprint data records the top three matching material numbers with the highest similarity and their corresponding values.

[0026] Step S2: Select the dust concentration prediction model based on the material characteristic fingerprint data and configure the initial parameters for crushing control;

[0027] Specifically, a hierarchical selection strategy is implemented based on the similarity values ​​in the material feature fingerprint data. High similarity directly calls the dedicated LSTM prediction model for the corresponding material. Medium similarity extracts the top three matching models, normalizes their similarity values, and uses them as weight coefficients. The network parameters of the three models are then weighted and summed to generate a fusion model. Low similarity calls a general model that covers all materials and is trained together. In the initial parameter configuration for crushing control, the initial speed of the crushing motor is calculated based on hardness and brittleness coefficient. Materials with high hardness require a reduced speed to decrease dust generation, while brittle materials require a slightly increased speed to ensure crushing efficiency. The initial power of the dust collector fan is calculated based on moisture content and bulk density. Dry materials with low moisture content generate a large amount of dust and require a higher dust collector power, while materials with high bulk density require a stronger negative pressure to overcome their own weight. The initial feed rate is calculated based on bulk density and brittleness coefficient. Materials with high density and low brittleness require a faster feed rate, while fragile materials require a slower feed rate to control sudden dust spikes.

[0028] Step S3: Collect real-time operating parameters of the crushing process and material feature vectors to form a multimodal feature vector sequence. Input the multimodal feature vector sequence into the dust concentration prediction model to obtain the predicted dust concentration sequence. Traverse the predicted dust concentration sequence to determine whether there are predicted values ​​that exceed the safety threshold. When there are predicted values ​​that exceed the standard, calculate the reduction in rotational speed, the increase in dust removal power, and the reduction coefficient of feed rate based on the deviation between the predicted value and the safety threshold as feedforward control quantities.

[0029] Specifically, the construction of the multimodal feature vector sequence involves collecting the actual speed, power, and current of the crushing motor to reflect the crushing intensity; the power and air volume of the dust removal fan to reflect the dust removal capacity; the negative pressure value inside the crushing chamber to reflect the airflow state; the speed of the feeding screw to reflect the feeding speed; and the material filling rate to reflect the degree of material accumulation inside the chamber. These eight real-time operating parameters, together with the four material characteristic parameters in step S1, constitute a 12-dimensional input feature vector. Thirty consecutive time steps are collected to form the temporal input. In the forward computation of the LSTM prediction model, the 128 neural units of the first LSTM layer extract the temporal features. Each neural unit uses a forget gate to determine how much historical information to retain, an input gate to determine how much current information to receive, and an output gate to determine how much hidden state to output. The Dropout layer randomly discards some neuron connections to prevent overfitting. The 64 neural units of the second LSTM layer perform deep temporal feature extraction. The 32 neural units of the fully connected layer map the temporal features to feature representations relevant to the prediction task. The output layer generates dust concentration prediction values ​​for the next six time steps. In the calculation of the feedforward control quantity, the first out-of-limit value is found by iterating through 6 predicted values. The difference between this value and the safety threshold is calculated and divided by the safety threshold to obtain the out-of-limit ratio. The speed reduction is obtained by multiplying the out-of-limit ratio by the speed adjustment coefficient. The dust removal power increase is obtained by multiplying the out-of-limit ratio by the power adjustment coefficient. The feed speed reduction coefficient is obtained by subtracting the product of the out-of-limit ratio and the feed adjustment coefficient from the unit value.

[0030] Step S4: Adjust the parameters of the crushing equipment according to the feedforward control quantity, and collect the actual dust concentration and predicted dust concentration sequence to calculate the deviation and perform feedback correction.

[0031] Specifically, the feedforward control action adjusts the speed of the crushing motor by a variable frequency drive to reduce the speed, adjusts the power of the dust collector fan by increasing the power of the dust collector, and adjusts the speed of the feed screw by a factor that reduces the feed rate. In the feedback correction, the laser scattering dust concentration sensor generates scattered light by irradiating dust particles with a laser beam. The detector receives the intensity of the scattered light and converts it into a dust concentration value. The deviation is obtained by subtracting the measured concentration from the predicted value of the prediction model at the corresponding time. When the absolute value of the deviation exceeds the set value, the PID controller is activated. The proportional term calculates the control quantity based on the current deviation, the integral term eliminates steady-state error based on historical accumulated deviations, and the derivative term suppresses overshoot based on the rate of change of deviation. The sum of these three terms yields the control quantity, which is then multiplied by a correction factor and multiplied by the adjusted dust collector power to obtain the final corrected power.

[0032] In one specific embodiment, step S1 includes:

[0033] The moisture content of the material to be crushed was determined by a near-infrared spectroscopy analyzer, the hardness value of the material to be crushed was determined by a Vickers hardness tester, the bulk density of the material to be crushed was determined by a density meter, and the brittleness coefficient of the material to be crushed was determined by a brittleness tester.

[0034] Moisture content, hardness, bulk density, and brittleness coefficient are used to construct a material characteristic vector;

[0035] Calculate the vector dot product and Euclidean norm of the material feature vector with each standard feature vector in the material feature library, and obtain the cosine similarity value based on the ratio of the vector dot product to the Euclidean norm.

[0036] Extract the IDs and corresponding similarity values ​​of the three matching materials with the highest similarity scores to generate material feature fingerprint data.

[0037] Specifically, the generation process of material feature fingerprint data essentially achieves rapid identification and matching of material types through similarity measurement in a multi-dimensional feature space. In this process, the calculation of cosine similarity fully considers the directional characteristics of the material feature vectors, rather than focusing solely on the absolute differences in numerical values. This makes the method inherently robust to dimensional differences in material feature values. When the cosine similarity between the feature vector of the material to be pulverized and the standard feature vector is close to 1, it indicates that the two are highly consistent in direction in the feature space, meaning that the comprehensive characteristics of the materials are highly similar. At this point, the pulverization process parameters corresponding to the matched material can be directly adopted. Extracting the top three highest similarity matching results instead of selecting only a single optimal match is to retain a certain margin of error in subsequent decisions. By weighted fusion of process parameters from multiple similar materials, it is possible to better cope with the minor fluctuations in actual material characteristics and the impact of measurement errors.

[0038] The construction of the material feature library relies on long-term accumulation of feature determination and process parameter optimization practices for a large number of known materials. Each standard feature vector corresponds to a validated optimal combination of crushing process parameters. During the similarity value screening process, the similarity threshold for the first three matching results is typically set to be no less than 0.85 to ensure that the matched material and the material to be crushed have sufficient similarity in physical properties. The matching material number contained in the material feature fingerprint data can directly index the corresponding historical process parameters in the material feature library, while the similarity value serves as a weighting coefficient for subsequent parameter fusion. The higher the similarity of the matched material, the greater the influence of its process parameters on the final decision. This feature-matching-based method avoids the blindness of traditional trial-and-error methods in adjusting process parameters, significantly shortens the optimization cycle of new material crushing processes, and ensures the objectivity and repeatability of parameter selection through mathematical methods.

[0039] In one specific embodiment, step S2 includes:

[0040] When the highest similarity value in the material feature fingerprint data is greater than or equal to 0.85, the dedicated dust concentration prediction model for the material corresponding to the highest similarity value is invoked.

[0041] When the highest similarity value is between 0.60 and 0.85, the prediction models of the first three matching materials in the material feature fingerprint data and their corresponding similarity values ​​are extracted. The similarity values ​​are used as weight coefficients to perform weighted fusion on the prediction models of the first three matching materials to obtain the fusion prediction model.

[0042] When the highest similarity score is less than 0.60, the general basic prediction model is invoked;

[0043] The initial speed of the crushing motor is calculated based on the hardness value and brittleness coefficient in the material characteristic vector. The initial power of the dust removal fan is calculated based on the moisture content and bulk density in the material characteristic vector. The initial value of the feed rate is calculated based on the bulk density and brittleness coefficient in the material characteristic vector.

[0044] Specifically, the strategy of hierarchically calling the prediction model is based on the direct impact of material feature similarity on prediction accuracy. Adaptive optimization of the prediction model selection is achieved by setting different similarity threshold ranges. When the highest similarity value reaches 0.85 or higher, it indicates that the material to be crushed and the matching material almost overlap in the feature space. At this point, directly calling the dedicated prediction model for that matching material can obtain the most accurate dust concentration prediction result, because this dedicated model has been deeply trained and its parameters optimized using historical data of this type of material. In the medium similarity range of 0.60 to 0.85, a dedicated model for a single matching material may not be fully suitable for the characteristics of the material to be crushed. A weighted fusion strategy can comprehensively utilize the prediction capabilities of multiple similar material models. By using similarity values ​​as weighting coefficients, a linear combination of model outputs is achieved, enabling the fused prediction model to maintain prediction accuracy while possessing better generalization ability and anti-interference characteristics. When the highest similarity value is below 0.60, it indicates that there is a lack of reference samples in the material feature library that are highly similar to the material to be crushed. At this time, the general basic prediction model trained on a large number of material types is called. Although the prediction accuracy is relatively reduced, it can provide a basically reliable dust concentration estimate and avoid the selection of process parameters due to model matching failure.

[0045] The calculation process of the initial speed of the crushing motor, the initial power of the dust collector fan, and the initial feed speed reflects the inherent correlation between the physical properties of the material and the crushing process parameters. Hardness and brittleness coefficient jointly determine the ease of crushing the material. Materials with higher hardness require greater mechanical energy input to achieve effective crushing, while materials with a higher brittleness coefficient are more prone to fragmentation under the same energy input. Therefore, the calculation of the initial speed of the crushing motor needs to comprehensively consider the synergistic effect of both, and determine the appropriate speed range by establishing a nonlinear mapping relationship between hardness and brittleness coefficient. Moisture content has a significant impact on dust generation and dust suspension characteristics in the airflow. Materials with higher moisture content generate relatively less dust, but dust particles are more likely to adhere to the equipment surface. Bulk density reflects the volume occupancy of the material during the crushing process and the airflow resistance. The initial power of the dust collector fan needs to be comprehensively determined based on the estimated dust load according to the moisture content and the required airflow and pressure calculated according to the bulk density. The calculation of the initial value of the feeding speed takes into account both bulk density and brittleness coefficient. Materials with high bulk density have a larger mass per unit volume, so the feeding speed needs to be controlled to avoid overloading the equipment. Materials with high brittleness coefficient have a fast crushing speed, so the feeding speed can be appropriately increased to improve production efficiency. The balance between the two determines the stable operation and capacity level of the crushing system.

[0046] In one specific embodiment, step S3 involves collecting real-time operating parameters of the crushing process and material feature vectors to form a multimodal feature vector sequence, including:

[0047] The actual speed, power and current of the crushing motor are collected; the actual power and air volume of the dust removal fan are collected; the real-time negative pressure value in the crushing chamber is collected; the actual speed of the feeding screw is collected; and the material filling rate in the crushing chamber is collected.

[0048] The actual rotational speed, actual power, actual current, actual power of the dust collector fan, actual air volume, real-time negative pressure value, actual rotational speed of the feed screw, material filling rate, and moisture content, hardness value, bulk density, and brittleness coefficient in the material feature vector are combined to form a 12-dimensional input feature vector.

[0049] A multimodal feature vector sequence is formed by continuously collecting 12-dimensional input feature vectors at 30 time steps.

[0050] Specifically, the acquisition of real-time operating parameters covers key dynamic information during the operation of the crushing system. These parameters directly reflect the energy consumption pattern, material flow state, and system load level of the material during the crushing process. The actual speed, actual power, and actual current of the crushing motor form a complete description of the motor's operating status. The actual speed reflects the resistance of the material in the crushing chamber to the blades, the actual power reflects the energy conversion efficiency of the crushing process, and the actual current reveals the motor's load level and potential overload risk. The actual power and actual air volume of the dust collector fan jointly characterize the working status of the dust collection system. When the amount of dust generated in the crushing chamber increases, maintaining the same air volume requires a larger power input to overcome the increased resistance caused by the increased dust load. The real-time negative pressure value in the crushing chamber is a direct indicator of the dust collection effect. Too low a negative pressure value will lead to dust overflow and environmental pollution, while too high a negative pressure value will increase energy consumption and may prematurely suck out insufficiently crushed material. The actual speed of the feed screw determines the rate at which material enters the crushing chamber, while the material filling rate reflects the instantaneous amount of material in the crushing chamber. The matching relationship between the two directly affects the crushing efficiency and system stability.

[0051] The construction of a 12-dimensional input feature vector achieves a deep integration of the inherent properties of the material and the dynamic response of the system. The first four dimensions of the material feature vector describe the physical properties of the material to be crushed, while the latter eight dimensions of real-time operating parameters capture the dynamic performance of these physical properties during the actual crushing process. This feature combination method enables the prediction model to not only perform static inference based on the properties of the material itself, but also to make dynamic corrections based on the real-time operating status of the system, significantly improving the accuracy and timeliness of dust concentration prediction. Continuously collecting feature vectors from 30 time steps forms a multimodal feature vector sequence, introducing information in the time dimension, enabling the model to capture the dynamic evolution trend and temporal dependencies during the crushing process. In the actual crushing process, changes in dust concentration are not instantaneous responses to operating parameters at a certain moment, but are influenced by the cumulative effect of preceding time steps. For example, a continuous increase in the material filling rate in the crushing chamber leads to a gradual increase in the dust generation rate, and fluctuations in the power of the dust collector fan will also have a delayed impact on dust concentration in subsequent time steps. Through serialized feature representation, the predictive model can learn these complex temporal correlation patterns, thereby enabling early prediction of dust concentration change trends and providing a reliable decision-making basis for the feedforward adjustment of the control system.

[0052] In one specific embodiment, step S3, inputting the multimodal feature vector sequence into the dust concentration prediction model to obtain the predicted dust concentration sequence, includes:

[0053] The multimodal feature vector sequence is input into the input layer of the LSTM dust concentration prediction model;

[0054] Temporal features are extracted from the multimodal feature vector sequence through the first LSTM layer, which contains 128 neural units and uses the tanh activation function to output the first hidden state sequence.

[0055] The first hidden state sequence is fed into a second LSTM layer after passing through a Dropout layer with a dropout rate of 0.3. The second LSTM layer contains 64 neurons and outputs the second hidden state sequence.

[0056] The second hidden state sequence is input into a fully connected layer, which contains 32 neurons and uses the ReLU activation function to output a mapped feature vector.

[0057] The mapped feature vector is input to the output layer to obtain the predicted dust concentration sequence for the next 6 time steps.

[0058] Specifically, the LSTM dust concentration prediction model employs a two-layer recurrent neural network architecture, achieving deep modeling of the temporal dynamics of the pulverization process through a hierarchical feature extraction mechanism. The 128 neurons in the first LSTM layer are responsible for capturing short-term fluctuation patterns and local temporal dependencies in the multimodal feature vector sequence. The tanh activation function constrains the output value of the cell state between -1 and +1, enabling the model to effectively handle the positive and negative trends of feature values. During the computation of the first LSTM layer, the forget gate determines which historical information needs to be retained based on the current input features and the hidden state of the previous time step. The input gate controls the degree of reception of new information at the current time step, and the output gate filters out the cell state components most relevant to the current prediction task. The synergistic effect of these three gating mechanisms allows the network to selectively memorize key information and filter out irrelevant noise in a long sequence of 30 time steps. The first hidden state sequence contains the temporal feature representation extracted by the first layer, but there may be a risk of overfitting. Therefore, regularization is performed by a Dropout layer with a dropout rate of 0.3. This layer randomly drops 30% of the neuron connections during the training phase, forcing the network to learn more robust feature representations and preventing the model from over-relying on certain specific feature combination patterns.

[0059] The second LSTM layer receives the first hidden state sequence after regularization. Its 64 neurons are half the number in the first layer, reflecting the progressive relationship of feature abstraction levels. This layer is mainly responsible for extracting higher-level temporal patterns, such as the periodic fluctuations in dust concentration and the coupled response relationships between multiple operating parameters. After being processed by two layers of LSTM, the original 12-dimensional operating feature sequence is compressed and mapped into a more compact high-order semantic representation. This representation is no longer a simple combination of feature values, but rather an abstract pattern containing the dynamic evolution mechanism of the crushing system. The fully connected layer, as the feature mapping stage, performs a nonlinear transformation on the final output of the second hidden state sequence through 32 neurons. The introduction of the ReLU activation function breaks the symmetry of the tanh activation in the LSTM layer, enhancing the model's ability to fit complex nonlinear mapping relationships, especially providing better representation of the asymmetric response characteristics of dust concentration in different operating ranges. The output layer decodes the 32-dimensional mapped feature vector into a predicted dust concentration sequence for the next 6 time steps. This multi-step prediction method can not only provide the concentration estimate for the next moment, but also provide the concentration change trend within a time window, reserving enough response time for the control system. This allows the adjustment of the dust collector fan power and feed rate to intervene in advance before the dust concentration actually rises, achieving an organic combination of feedforward control and feedback control.

[0060] In one specific embodiment, the predicted dust concentration sequence is traversed to determine whether there are predicted values ​​exceeding the safety threshold. When an exceeding predicted value exists, the reduction in rotational speed, the increase in dust removal power, and the reduction coefficient in feed rate are calculated as feedforward control quantities based on the deviation between the exceeding predicted value and the safety threshold. This includes:

[0061] Set a safe upper limit for dust concentration as a safety threshold;

[0062] Iterate through the predicted dust concentration sequence to determine if there are any predicted values ​​that exceed the safety threshold;

[0063] When there is a predicted value that exceeds the safety threshold, the predicted value corresponding to the earliest time when the exceedance occurs is recorded as the exceedance predicted value.

[0064] The excess difference is obtained by subtracting the safety threshold from the predicted excess value. The excess difference is divided by the safety threshold to obtain the excess ratio. The excess ratio is multiplied by the speed adjustment coefficient to obtain the speed reduction. The excess ratio is multiplied by the power adjustment coefficient to obtain the dust removal power increase. The excess ratio is multiplied by the feed adjustment coefficient to obtain the feed speed reduction ratio. The feed speed reduction coefficient is obtained by subtracting the feed speed reduction ratio from the unit value.

[0065] Specifically, the dust concentration safety threshold is set based on occupational health standards and explosion-proof safety requirements of the crushing workplace. This threshold is typically determined by combining the lower explosive limit concentration of the material and the time-weighted average permissible concentration, ensuring that the dust concentration of the crushing system remains within a safe and controllable range under any operating condition. The traversal judgment of the predicted dust concentration sequence enables proactive identification of potential safety risks within future time windows. Compared to traditional real-time monitoring feedback control modes, this prediction-based proactive intervention strategy can initiate control actions before the dust concentration actually exceeds the limit, avoiding the short-term concentration exceedance problem caused by the inherent lag of feedback control. Recording the predicted value corresponding to the earliest time of exceedance as the basis for control decisions reflects the safety-first design philosophy. Even if the predicted value at subsequent times may show a higher degree of exceedance, the control system prioritizes responding to the earliest risk signal, minimizing the time interval from risk prediction to control intervention. This is crucial for preventing instantaneous hazards such as dust explosions.

[0066] The calculation of the deviation from the standard and the proportion of deviation transforms the absolute deviation of the predicted concentration into a standardized expression of the relative deviation, enabling the adjustment range of the control parameters to have a direct proportional response to the severity of the risk. The calculation of the speed reduction establishes a linear mapping between the proportion of deviation and the adjustment of the crushing motor speed through a speed adjustment coefficient. Reducing the speed directly reduces the amount of material crushed per unit time, suppressing the dust generation rate at its source. The value of the speed adjustment coefficient needs to comprehensively consider the motor's speed range and the acceptable decrease in crushing efficiency. The calculation of the increase in dust removal power enhances the dust collection capacity of the dust removal system. By increasing the fan power, the suction volume and negative pressure intensity are increased, accelerating the dust discharge speed from the crushing chamber. The setting of the power adjustment coefficient needs to take into account both the upper limit of the fan's rated power and the pressure loss characteristics of the pipeline system. The two-step calculation of the feed rate reduction ratio and the feed rate reduction coefficient enables proportional adjustment of the feed rate. The feed rate adjustment coefficient determines the sensitivity of the feed rate to the risk of exceeding the limit. The reduction coefficient, obtained by subtracting the reduction ratio from the unit value, can be directly used as a multiplication factor for the current feed rate, achieving a smooth reduction in the feed rate without abrupt speed changes. The synergistic adjustment of the three types of control parameters constitutes a multi-dimensional dust concentration suppression strategy. It reduces dust generation during material crushing, enhances dust removal during dust collection, and avoids excessive dust diffusion caused by excessive material accumulation in the crushing chamber by controlling the material supply rate. The combined effect of these three factors ensures that the dust concentration can be effectively controlled before the predicted exceedance occurs.

[0067] In one specific embodiment, step S4 includes:

[0068] The speed of the crushing motor is adjusted by the frequency converter according to the amount of speed reduction; the power of the dust removal fan is adjusted by the frequency converter according to the amount of dust removal power increase; and the speed of the feeding screw is adjusted by the frequency converter according to the feeding speed reduction coefficient.

[0069] The actual dust concentration inside the pulverizing chamber is collected using a laser scattering dust concentration sensor.

[0070] Calculate the deviation between the actual dust concentration and the predicted dust concentration at the corresponding time in the predicted dust concentration sequence;

[0071] When the absolute value of the deviation is greater than 1.0 mg / When the deviation is input into the PID controller, the control quantity is calculated. The dust removal power correction coefficient is calculated based on the control quantity. The dust removal power correction coefficient is multiplied by the adjusted dust removal fan power to obtain the corrected dust removal fan power.

[0072] Specifically, the frequency converter, acting as an actuator, precisely converts control commands into motor operating parameters. By adjusting the output frequency to change the synchronous speed of the motor, it achieves stepless adjustment of the crushing motor speed, dust collector fan power, and feed screw speed. The reduction in crushing motor speed is set according to the calculated speed reduction amount. After receiving the speed command, the frequency converter adjusts the ratio between the output frequency and voltage amplitude through a closed-loop vector control algorithm to ensure the motor maintains stable torque output during speed reduction, avoiding impact vibration between the blades and the material caused by sudden speed changes. The increase in dust collector fan power is achieved by increasing the output frequency of the frequency converter. When converting the power increase into a frequency increment, the power-speed cube relationship of the fan needs to be considered; that is, the fan power is proportional to the cube of the speed. Therefore, power adjustment requires a cube root calculation of the frequency increment to obtain an accurate speed adjustment value. The feed screw speed is adjusted by multiplying the feed speed reduction coefficient by the current speed to calculate the target speed. The frequency converter drives the feed motor to smoothly transition to the new speed set point through V / F control mode or vector control mode. The response time of the entire adjustment process is usually controlled within 2 to 5 seconds, which ensures both timely control and avoids mechanical shock.

[0073] The laser scattering dust concentration sensor, based on Mie scattering theory, inverts the dust mass concentration by measuring the intensity of light scattered by dust particles from a laser beam. The sensor is installed on the side wall of the pulverizing chamber or in the exhaust duct, with a sampling period set to 1 second to ensure it can capture rapid changes in dust concentration. The deviation between the actual and predicted dust concentration reflects the accuracy of the prediction model and the actual performance of the control. This deviation may stem from batch fluctuations in material characteristics, sensor measurement errors, or external interference. When the absolute value of the deviation exceeds the threshold of 1.0 mg / m³, it indicates a significant difference between the prediction model's output and the actual operating conditions, requiring a feedback correction mechanism for compensation. After receiving the deviation signal, the PID controller's proportional element generates an immediate response based on the magnitude of the deviation, the integral element accumulates historical deviations to eliminate steady-state errors, and the derivative element predicts the trend of deviation changes to suppress overshoot. The weighted sum of the outputs of these three elements yields the control quantity, which characterizes the additional dust removal capacity adjustment required to eliminate the current deviation.

[0074] The dust removal power correction coefficient is calculated by the ratio of the control quantity to the reference power. When the actual dust concentration is higher than the predicted value, the correction coefficient is greater than 1, indicating that the dust removal capacity needs to be further enhanced. When the actual concentration is lower than the predicted value, the correction coefficient is less than 1 but not lower than the lower limit of 0.9 to maintain the necessary dust removal margin. Multiplying the correction coefficient by the adjusted dust removal fan power achieves the superposition and fusion of feedforward control and feedback control. Feedforward control provides coarse adjustment based on the predictive model, while feedback control provides fine adjustment based on actual measurements. The synergistic effect of the two constitutes a dual closed-loop control structure, which significantly improves the dynamic performance and steady-state accuracy of dust concentration control. The corrected dust removal fan power is then sent back to the dust removal fan through the frequency converter, forming a closed control loop. The complete chain of the entire control process, from prediction to adjustment to correction, ensures that the pulverizing system can stably control the dust concentration below the safety threshold under various operating disturbances. At the same time, by online tuning of PID parameters, it adapts to changes in different material characteristics and process conditions, achieving adaptive optimization of dust concentration control.

[0075] The control method for intelligent dust-free pulverization in pharmaceutical production, as described above in the embodiments of this application, is as follows: Figure 2 As shown, the intelligent dust-free pulverizing control system for pharmaceutical production in this application embodiment is described below. One embodiment of the intelligent dust-free pulverizing control system for pharmaceutical production in this application embodiment includes:

[0076] The calculation module is used to collect the moisture content, hardness, bulk density and brittleness coefficient of the material to be crushed to construct a material feature vector, and to perform cosine similarity calculation between the material feature vector and the material feature library to generate material feature fingerprint data.

[0077] The configuration module is used to select a dust concentration prediction model and configure the initial parameters for crushing control based on the material characteristic fingerprint data.

[0078] The input module is used to collect real-time operating parameters of the crushing process and form a multimodal feature vector sequence with the material feature vector. The multimodal feature vector sequence is input into the dust concentration prediction model to obtain a predicted dust concentration sequence. The predicted dust concentration sequence is traversed to determine whether there is a predicted value that exceeds the safety threshold. When there is a predicted value that exceeds the standard, the speed reduction, dust removal power increase and feed rate reduction coefficient are calculated based on the deviation between the predicted value and the safety threshold as feedforward control quantities.

[0079] The correction module is used to adjust the parameters of the crushing equipment according to the feedforward control quantity, and to collect the actual dust concentration and the predicted dust concentration sequence to calculate the deviation and perform feedback correction.

[0080] Reference Figure 3 This invention also provides a control device for intelligent dust-free pulverization in pharmaceutical production. This control device can be a server, and its internal structure can be as follows: Figure 3 As shown. The intelligent dust-free pulverizing control equipment for pharmaceutical production includes a processor, memory, display screen, input device, network interface, and database connected via a system bus. The processor, designed as a computer, provides computing and control capabilities. The memory of the intelligent dust-free pulverizing control equipment includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database of the intelligent dust-free pulverizing control equipment for pharmaceutical production stores the data corresponding to this embodiment. The network interface of the intelligent dust-free pulverizing control equipment for pharmaceutical production is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements the above-described method.

[0081] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A control method for intelligent dust-free pulverization in pharmaceutical production, characterized in that, The method includes: Step S1: Collect the moisture content, hardness, bulk density and brittleness coefficient of the material to be crushed to construct a material feature vector, and perform cosine similarity calculation between the material feature vector and the material feature library to generate material feature fingerprint data; Step S2: Select a dust concentration prediction model and configure initial parameters for crushing control based on the material characteristic fingerprint data; Step S3: Collect real-time operating parameters of the crushing process and form a multimodal feature vector sequence with the material feature vector. Input the multimodal feature vector sequence into the dust concentration prediction model to obtain the predicted dust concentration sequence. Iterate through the predicted dust concentration sequence to determine whether there is a predicted value that exceeds the safety threshold. When there is a predicted value that exceeds the standard, calculate the speed reduction, dust removal power increase and feed rate reduction coefficient based on the deviation between the predicted value that exceeds the standard and the safety threshold as feedforward control quantities. Step S4: Adjust the parameters of the crushing equipment according to the feedforward control quantity, and collect the actual dust concentration and the predicted dust concentration sequence to calculate the deviation and perform feedback correction.

2. The control method for intelligent dust-free pulverization in pharmaceutical production according to claim 1, characterized in that, Step S1 includes: The moisture content of the material to be crushed was determined by a near-infrared spectroscopy analyzer, the hardness value of the material to be crushed was determined by a Vickers hardness tester, the bulk density of the material to be crushed was determined by a density meter, and the brittleness coefficient of the material to be crushed was determined by a brittleness tester. The moisture content, hardness value, bulk density, and brittleness coefficient are used to construct a material characteristic vector; The dot product and Euclidean norm of the material feature vector are calculated with each standard feature vector in the material feature library, and the cosine similarity value is obtained based on the ratio of the dot product to the Euclidean norm. Extract the IDs and corresponding similarity values ​​of the three matching materials with the highest similarity scores to generate material feature fingerprint data.

3. The control method for intelligent dust-free pulverization in pharmaceutical production according to claim 1, characterized in that, Step S2 includes: When the highest similarity value in the material feature fingerprint data is greater than or equal to 0.85, the dedicated dust concentration prediction model for the material corresponding to the highest similarity value is invoked. When the highest similarity value is between 0.60 and 0.85, the prediction models and corresponding similarity values ​​of the first three matching materials in the material feature fingerprint data are extracted, and the similarity values ​​are used as weight coefficients to perform weighted fusion on the prediction models of the first three matching materials to obtain a fusion prediction model. When the highest similarity value is less than 0.60, the general basic prediction model is invoked; The initial speed of the crushing motor is calculated based on the hardness value and brittleness coefficient in the material feature vector. The initial power of the dust removal fan is calculated based on the moisture content and bulk density in the material feature vector. The initial value of the feed rate is calculated based on the bulk density and brittleness coefficient in the material feature vector.

4. The control method for intelligent dust-free pulverization in pharmaceutical production according to claim 1, characterized in that, In step S3, the real-time operating parameters of the crushing process are collected and combined with the material feature vector to form a multimodal feature vector sequence, including: The actual speed, power and current of the crushing motor are collected; the actual power and air volume of the dust removal fan are collected; the real-time negative pressure value in the crushing chamber is collected; the actual speed of the feeding screw is collected; and the material filling rate in the crushing chamber is collected. The actual rotational speed, actual power, actual current, actual power of the dust collector fan, actual air volume, real-time negative pressure value, actual rotational speed of the feed screw, and material filling rate are combined with the moisture content, hardness value, bulk density, and brittleness coefficient in the material feature vector to form a 12-dimensional input feature vector. The 12-dimensional input feature vectors are continuously collected at 30 time steps to form a multimodal feature vector sequence.

5. The control method for intelligent dust-free pulverization in pharmaceutical production according to claim 4, characterized in that, In step S3, inputting the multimodal feature vector sequence into the dust concentration prediction model to obtain the predicted dust concentration sequence includes: The multimodal feature vector sequence is input into the input layer of the LSTM dust concentration prediction model; Temporal features are extracted from the multimodal feature vector sequence through a first LSTM layer. The first LSTM layer contains 128 neural units and uses the tanh activation function to output the first hidden state sequence. The first hidden state sequence is passed through a Dropout layer with a dropout rate of 0.3 and then fed into a second LSTM layer containing 64 neurons, which outputs a second hidden state sequence. The second hidden state sequence is input into a fully connected layer, which contains 32 neurons that use the ReLU activation function and output a mapped feature vector. The mapped feature vector is input to the output layer to obtain the predicted dust concentration sequence for the next 6 time steps.

6. The control method for intelligent dust-free pulverization in pharmaceutical production according to claim 5, characterized in that, In step S3, the step of traversing the predicted dust concentration sequence to determine whether there are predicted values ​​exceeding the safety threshold is performed. When there are predicted values ​​exceeding the threshold, the reduction in rotational speed, the increase in dust removal power, and the reduction coefficient in feed rate are calculated as feedforward control quantities based on the deviation between the predicted value exceeding the threshold and the safety threshold. This includes: Set a safe upper limit for dust concentration as a safety threshold; The predicted dust concentration sequence is traversed to determine whether there are any predicted values ​​that exceed the safety threshold; When there is a predicted value that exceeds the safety threshold, the predicted value corresponding to the earliest time when the exceedance occurs is recorded as the exceedance predicted value. The excess difference is obtained by subtracting the safety threshold from the predicted excess value. The excess difference is then divided by the safety threshold to obtain the excess ratio. The excess ratio is multiplied by the speed adjustment coefficient to obtain the speed reduction. The excess ratio is multiplied by the power adjustment coefficient to obtain the dust removal power increase. The excess ratio is multiplied by the feed adjustment coefficient to obtain the feed speed reduction ratio. The feed speed reduction coefficient is obtained by subtracting the feed speed reduction ratio from the unit value.

7. The control method for intelligent dust-free pulverization in pharmaceutical production according to claim 1, characterized in that, Step S4 includes: The speed of the crushing motor is adjusted by the frequency converter according to the decrease in rotational speed; the power of the dust removal fan is adjusted by the frequency converter according to the increase in dust removal power; and the speed of the feeding screw is adjusted by the frequency converter according to the decrease in feeding speed coefficient. The actual dust concentration inside the pulverizing chamber is collected using a laser scattering dust concentration sensor. Calculate the deviation between the actual dust concentration and the predicted dust concentration at the corresponding time in the predicted dust concentration sequence; When the absolute value of the deviation is greater than 1.0 mg / When the deviation is input into the PID controller, the control quantity is calculated. The dust removal power correction coefficient is calculated based on the control quantity. The dust removal power correction coefficient is multiplied by the adjusted dust removal fan power to obtain the corrected dust removal fan power.

8. A control system for intelligent dust-free pulverization in pharmaceutical production, characterized in that, A control method for implementing intelligent dust-free pulverization for pharmaceutical production as described in any one of claims 1-7, wherein the control system for intelligent dust-free pulverization for pharmaceutical production comprises: The calculation module is used to collect the moisture content, hardness, bulk density and brittleness coefficient of the material to be crushed to construct a material feature vector, and to perform cosine similarity calculation between the material feature vector and the material feature library to generate material feature fingerprint data. The configuration module is used to select a dust concentration prediction model and configure the initial parameters for crushing control based on the material characteristic fingerprint data. The input module is used to collect real-time operating parameters of the crushing process and form a multimodal feature vector sequence with the material feature vector. The multimodal feature vector sequence is input into the dust concentration prediction model to obtain a predicted dust concentration sequence. The predicted dust concentration sequence is traversed to determine whether there is a predicted value that exceeds the safety threshold. When there is a predicted value that exceeds the standard, the speed reduction, dust removal power increase and feed rate reduction coefficient are calculated based on the deviation between the predicted value and the safety threshold as feedforward control quantities. The correction module is used to adjust the parameters of the crushing equipment according to the feedforward control quantity, and to collect the actual dust concentration and the predicted dust concentration sequence to calculate the deviation and perform feedback correction.

Images