Traditional Chinese medicine concentration process optimization system based on deep learning

Abstract

The invention provides a traditional Chinese medicine concentration process optimization system based on deep learning. The traditional Chinese medicine concentration process optimization system comprises a data acquisition module, a preprocessing module, a feature extraction module, a process optimization module, a real-time control module, a quality evaluation module and a feedback adjustment module. The data acquisition module acquires sensor data such as temperature and pressure and image data; the data preprocessing module eliminates noise, aligns a time sequence and standardizes data; the feature extraction module extracts dynamic features through a time-frequency fusion algorithm and a deep learning model; the process optimization module constructs a multi-objective optimization model and generates an adjustment instruction; the real-time control module adjusts the equipment according to the instruction; the quality evaluation module is used for detecting the product quality on line and calculating the standard-reaching rate; and the feedback adjustment module updates the deep learning model parameters according to the evaluation result. The intelligent level and the product quality of traditional Chinese medicine production can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of traditional Chinese medicine concentration technology, and more specifically, to a traditional Chinese medicine concentration process optimization system based on deep learning. Background Art

[0002] Traditional Chinese medicine (TCM) concentration is a critical step in the pharmaceutical manufacturing process. Its purpose is to remove water from the medicinal solution by evaporation, thereby increasing the concentration and stability of the active ingredients. Traditional TCM concentration processes rely primarily on empirical control, with manual monitoring of parameters such as temperature, pressure, and time to control the concentration process. However, due to the complex composition of TCM, the properties of the medicinal solution vary depending on factors such as the type of herbal medicine, origin, and collection time. This makes it difficult to precisely control the concentration process using traditional methods, resulting in unstable product quality. Furthermore, manual monitoring is not only inefficient but also susceptible to human factors, making it difficult to achieve standardization and automation for large-scale production. With the advancement of modern technology, automated control technologies have begun to be applied to TCM concentration processes. Sensors monitor key parameters in the concentration process in real time and automatically adjust the operating status of the equipment, improving the stability and consistency of the concentration process. However, most of these automated systems rely on fixed control strategies, lack the ability to dynamically optimize the process, and are unable to adapt to changes in the medicinal solution properties, thus limiting further improvements in concentration efficiency and product quality.

[0003] In the process of implementing the embodiments of the present invention, the inventors found that there are at least the following problems or defects in the existing technology: the traditional Chinese medicine concentration process relies on experience-based operations, and the product quality is unstable; the automated control system lacks dynamic optimization capabilities and has difficulty adapting to changes in the properties of the medicinal solution, which limits the improvement of concentration efficiency and product quality. Summary of the Invention

[0004] The present invention provides a traditional Chinese medicine concentration process optimization system based on deep learning, comprising:

[0005] Data acquisition module, used to obtain real-time sensor data and image data of temperature, pressure, flow rate and medicinal material composition during the Chinese medicine concentration process;

[0006] Data preprocessing module, used to eliminate noise, align multi-source time series and standardize the raw data;

[0007] Feature extraction module, used to extract process dynamic features through time-frequency fusion algorithm and deep learning model;

[0008] Process optimization module, which is used to build a multi-objective optimization model based on characteristic data and generate adjustment instructions for concentration process parameters;

[0009] A real-time control module is used to dynamically adjust the heating power, vacuum level, and feed rate of the concentration equipment according to optimization instructions;

[0010] The quality assessment module is used to detect the active ingredient concentration, viscosity and color of the concentrated product online and calculate the process compliance rate;

[0011] Feedback adjustment module, used to iteratively update deep learning model parameters based on quality assessment results and historical optimization data.

[0012] Furthermore, the data acquisition module includes:

[0013] Multimodal sensor unit, used to synchronously collect output signals from temperature sensor, pressure transmitter and spectrum analyzer;

[0014] High-speed image acquisition unit, used to capture high-definition image sequences of the boiling state of the chemical liquid in the concentration kettle;

[0015] The data verification unit is used to eliminate abnormal collected values based on the physical constraints of the device.

[0016] Furthermore, the data preprocessing module includes:

[0017] The adaptive filtering unit uses the following formula to eliminate high-frequency noise in the sensor signal:

[0018] y t =αx t +(1-α)y t-1

[0019] Among them, y t is the filtered signal at time t, x t is the original signal at time t, α is the dynamic smoothing coefficient, and its value range is [0.2, 0.8];

[0020] Timing alignment unit, used to map data with different sampling frequencies to a unified time axis through cubic spline interpolation;

[0021] Normalization unit, scaling the data to the [-1, 1] interval by batch.

[0022] Furthermore, the specific steps of the time-frequency fusion algorithm include:

[0023] Step 1: Perform wavelet packet decomposition on the sensor signal to extract the energy ratio characteristics of each frequency band

[0024]

[0025] Among them, W k is the kth wavelet coefficient, and N is the total number of coefficients;

[0026] Step 2: Use a 3D convolutional network to extract spatial-temporal features F from the image sequence v, and calculate the texture complexity of the keyframe

[0027]

[0028] Among them, p i is the probability of the i-th gray level in the gray level histogram, and M is the total number of gray levels;

[0029] Step 3: Fusion of time domain and frequency domain features through attention mechanism

[0030]

[0031] in, is the energy characteristic of the jth frequency band, is the jth space-time feature, w j is the dynamic weight of the jth feature, J is the total number of fusion features, Represents a feature concatenation operation.

[0032] Furthermore, the objective function of the multi-objective optimization model is:

[0033]

[0034] Among them, C 实际 is the real-time concentration of active ingredients, C 目标 is the preset concentration threshold, ΔP is the pressure fluctuation variance, T(t) is the temperature deviation function, λ1, λ2, λ3 are dynamic penalty coefficients, and t0 and t1 are the process start and end times, respectively.

[0035] Furthermore, the update rule of the dynamic penalty coefficient is:

[0036]

[0037] in, is the updated i-th penalty coefficient, is the i-th penalty coefficient before update, L is the process compliance rate loss function, η is the learning rate, i∈{1,2,3}, and satisfies constraints.

[0038] Furthermore, the real-time control module includes:

[0039] A fuzzy decision-making unit is used to generate a fuzzy set of control variables based on the optimization instructions and the current state of the equipment;

[0040] The anti-saturation regulation unit uses the following formula to limit the actuator output range:

[0041]

[0042] Among them, u实际 is the actual output value of the actuator, u 指令 is the output value of the optimization instruction, u max is the maximum allowable output value of the actuator, u min is the minimum allowable output value of the actuator;

[0043] The dynamic compensation unit is used to predict and correct the actuator hysteresis effect based on the historical control error.

[0044] Furthermore, the quality assessment module performs the following operations:

[0045] Online detection unit, which measures the concentration of active ingredients in the drug solution in real time through near-infrared spectrometer;

[0046] Physical property analysis unit uses the fusion data of rotational viscometer and colorimeter to calculate the comprehensive quality score

[0047] Q=ω1S 浓度 +ω2S 粘度 +ω3S 色差

[0048] Among them, S 浓度 Score for active ingredient concentration, 粘度 S is the viscosity score, 色差 is the color difference score, ω1, ω2, ω3 are the preset weight coefficients;

[0049] The compliance judgment unit triggers a process abnormality alarm when Q<0.9.

[0050] Furthermore, the iterative update process of the feedback adjustment module includes:

[0051] A parameter correction unit that adjusts the loss function weights of the deep learning model based on the quality score deviation;

[0052] The feature attenuation unit uses the following formula to eliminate invalid process features:

[0053]

[0054] in, is the updated weight of the i-th feature, is the weight of the i-th feature before updating, τ is the decay coefficient, and Δt is the duration that the feature is not effectively activated;

[0055] The model reconstruction unit reinitializes the network hidden layer when the cumulative quality deviation exceeds a threshold.

[0056] Furthermore, the dynamic weight in the time-frequency fusion algorithm is calculated as follows:

[0057]

[0058] in, is the energy characteristic of the jth frequency band, is the j-th spatial-temporal visual feature, σ(·) is the Sigmoid function, tanh(·) is the hyperbolic tangent function, and K is the total number of fused features.

[0059] The above-described embodiments of the present invention have at least the following beneficial effects: the system can achieve intelligent optimization of the traditional Chinese medicine concentration process. Using deep learning technology, various data from the concentration process are analyzed and processed in real time, allowing precise adjustment of process parameters to improve concentration efficiency and product quality. Furthermore, the system can also perform online quality testing of the concentrated product and automatically adjust process parameters based on the test results, ensuring stable and consistent product quality.

[0060] Furthermore, the system can continuously optimize deep learning models through a feedback adjustment mechanism, improving their accuracy and adaptability, further enhancing system performance and effectiveness. This intelligent optimization system can effectively address issues existing in traditional Chinese medicine concentration processes and improve the modernization and competitiveness of Chinese medicine production. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] The above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily apparent by reading the following detailed description with reference to the accompanying drawings, in which several embodiments of the present invention are shown by way of example and not limitation, in which:

[0062] Figure 1 A schematic diagram of the structure of a deep learning-based traditional Chinese medicine concentration process optimization system provided in one embodiment of the present invention. DETAILED DESCRIPTION

[0063] The principles and spirit of the present invention will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided solely to enable those skilled in the art to better understand and implement the present invention, and are not intended to limit the scope of the present invention in any way. Rather, these embodiments are provided to make the present invention more thorough and complete, and to fully convey the scope of the present invention to those skilled in the art.

[0064] Those skilled in the art will appreciate that the embodiments of the present invention may be implemented as a system, apparatus, device, method, or computer program product. Therefore, the present invention may be implemented in the following forms: entirely in hardware, entirely in software (including firmware, resident software, microcode, etc.), or in a combination of hardware and software.

[0065] It should be noted that any number of elements in the drawings is for illustration only and not for limitation, and any naming is only for distinction and does not have any limiting meaning.

[0066] Reference below Figure 1 , Figure 1 This is a schematic diagram of the structure of a Chinese medicine concentration process optimization system based on deep learning provided by one embodiment of the present invention. Figure 1 As shown, a Chinese medicine concentration process optimization system 100 based on deep learning includes:

[0067] The data acquisition module 101 is used to obtain sensor data and image data of temperature, pressure, flow rate and medicinal material components in real time during the Chinese medicine concentration process;

[0068] The data preprocessing module 102 is used to perform noise elimination, multi-source time series alignment and standardization on the original data;

[0069] Feature extraction module 103, used to extract process dynamic features through time-frequency fusion algorithm and deep learning model;

[0070] The process optimization module 104 is used to build a multi-objective optimization model based on the characteristic data and generate concentration process parameter adjustment instructions;

[0071] Real-time control module 105, for dynamically adjusting the heating power, vacuum degree and feed rate of the concentration equipment according to the optimization instructions;

[0072] The quality assessment module 106 is used to detect the active ingredient concentration, viscosity and color of the concentrated product online and calculate the process compliance rate;

[0073] The feedback adjustment module 107 is used to iteratively update the deep learning model parameters based on the quality assessment results and historical optimization data.

[0074] The present invention relates to a deep learning-based system for optimizing the concentration process of traditional Chinese medicine (TCM). The system comprises multiple modules that work together to optimize the TCM concentration process. The system includes a data acquisition module that captures key data from the concentration process, providing a foundation for subsequent processing; a data preprocessing module that processes raw data to ensure data quality; a feature extraction module that extracts dynamic process features using advanced algorithms; a process optimization module that constructs an optimization model based on feature data; a real-time control module that adjusts equipment according to optimization instructions; a quality assessment module that monitors product quality online; and a feedback adjustment module that updates model parameters based on the assessment results. These modules work together to effectively improve the efficiency and product quality of the TCM concentration process.

[0075] Specifically, the multimodal sensor unit in the data acquisition module synchronously collects temperature, pressure, and spectrum analyzer output signals to ensure comprehensive and accurate data. The high-speed image acquisition unit captures high-definition image sequences of the boiling liquid in the concentration kettle, providing rich visual information for feature extraction. The data verification unit eliminates anomalous acquisition values based on the physical constraints of the equipment to ensure data reliability. The adaptive filtering unit in the data preprocessing module uses a specific formula to eliminate high-frequency noise in sensor signals. The time alignment unit uses cubic spline interpolation to map data with different sampling frequencies to a unified time axis. The normalization unit scales data batches to specific intervals. These processing steps ensure data standardization and usability. The time-frequency fusion algorithm in the feature extraction module uses wavelet packet decomposition, a three-dimensional convolutional network, and an attention mechanism to extract dynamic process features, providing key information for process optimization. The multi-objective optimization model in the process optimization module is based on a specific objective function, and the dynamic penalty coefficient is updated based on the process compliance loss function to ensure the adaptability and effectiveness of the optimization model. The fuzzy decision-making unit, anti-saturation adjustment unit, and dynamic compensation unit in the real-time control module work together to dynamically adjust the operating status of the concentration equipment according to optimization instructions, ensuring equipment stability and responsiveness. The online detection unit, physical property analysis unit, and compliance determination unit in the quality assessment module work together to monitor the quality of the concentrated product in real time and trigger an alarm if quality does not meet standards. The parameter correction unit, feature attenuation unit, and model reconstruction unit in the feedback adjustment module continuously iteratively update the deep learning model parameters based on quality assessment results and historical optimization data, improving the model's accuracy and stability.

[0076] Preferably, the multimodal sensor unit of the data acquisition module can include high-precision temperature sensors, pressure transmitters, and spectrum analyzers to ensure the accuracy and reliability of collected data. The high-speed image acquisition unit can utilize a high-resolution industrial camera to capture clear images of boiling liquid medicine. The data verification unit can set a reasonable threshold range to effectively eliminate abnormal collected values based on the physical constraints of the equipment. In the adaptive filtering unit of the data preprocessing module, the dynamic smoothing coefficient can be set to approximately 0.5 to achieve good filtering effects. The time alignment unit can utilize a high-precision cubic spline interpolation algorithm to ensure data accuracy on a unified time axis. The normalization unit can scale the data batches to the [-1, 1] interval to meet subsequent processing requirements. In the time-frequency fusion algorithm of the feature extraction module, wavelet packet decomposition can select appropriate wavelet basis functions and decomposition levels to extract energy contribution features for each frequency band. The three-dimensional convolutional network can utilize a pretrained model to improve the extraction of spatial-temporal features. The attention mechanism can utilize a dynamic weight calculation method to effectively integrate time and frequency domain features. In the multi-objective optimization model of the process optimization module, the parameters of the objective function can be set according to actual process requirements, and the dynamic penalty coefficient update rule can use an appropriate learning rate to ensure rapid model convergence and stability. The fuzzy decision-making unit of the real-time control module can use a fuzzy logic control algorithm to generate a reasonable fuzzy set of control variables based on the optimization instructions and the current equipment state. The anti-saturation adjustment unit can set the maximum and minimum allowable output values of the actuator to prevent actuator saturation. The dynamic compensation unit can use an appropriate historical control error prediction algorithm to correct for actuator hysteresis. The online detection unit of the quality assessment module can use a near-infrared spectrometer to measure the active ingredient concentration of the drug solution in real time. The physical property analysis unit can use a rotational viscometer and colorimeter to calculate the comprehensive quality score. The compliance judgment unit can set a reasonable quality score threshold to ensure product quality stability. The parameter correction unit of the feedback adjustment module can adjust the loss function weights of the deep learning model based on quality score deviations. The feature attenuation unit can set an appropriate attenuation coefficient to eliminate invalid process features. The model reconstruction unit can reinitialize the network hidden layer to restore model performance when the cumulative quality deviation exceeds a threshold.

[0077] In some embodiments, the data acquisition module includes:

[0078] Multimodal sensor unit, used to synchronously collect output signals from temperature sensor, pressure transmitter and spectrum analyzer;

[0079] High-speed image acquisition unit, used to capture high-definition image sequences of the boiling state of the chemical liquid in the concentration kettle;

[0080] The data verification unit is used to eliminate abnormal collected values based on the physical constraints of the device.

[0081] It should be noted that the data acquisition module in this system comprises a multimodal sensor unit, a high-speed image acquisition unit, and a data verification unit. These units work together to ensure the comprehensiveness and accuracy of the collected data. The multimodal sensor unit simultaneously collects temperature, pressure, and spectrum analyzer output signals, providing basic data for subsequent data processing. The high-speed image acquisition unit captures a high-definition image sequence of the boiling liquid in the concentration kettle, providing visual information for feature extraction. The data verification unit eliminates anomalous collected values based on the physical constraints of the equipment to ensure data reliability. The implementation of these functions is crucial to the stable operation of the entire system.

[0082] Specifically, the temperature sensor, pressure transmitter, and spectrum analyzer in the multimodal sensor unit need to have high precision and high stability to ensure the accuracy of the collected data. The accuracy of the temperature sensor should reach ±0.1°C, the accuracy of the pressure transmitter should reach ±0.5kPa, and the wavelength range of the spectrum analyzer should cover the characteristic absorption peaks of the active ingredients of traditional Chinese medicine. The frame rate of the high-speed image acquisition unit should be no less than 30fps, and the resolution should be no less than 1080p to ensure that the boiling state of the medicinal liquid can be clearly captured. The outlier rejection algorithm of the data verification unit should be based on the physical constraints of the equipment, such as the reasonable range of temperature and pressure, and the intensity range of the spectral signal. The setting of these parameters can ensure that the collected data is both comprehensive and accurate, providing a reliable foundation for subsequent data processing and analysis.

[0083] Preferably, the multimodal sensor unit can adopt a redundant design, that is, multiple sensors are configured for each sensor type to improve the reliability of the data and the fault tolerance of the system. For example, two temperature sensors and two pressure transmitters can be configured to improve the measurement accuracy through data fusion algorithms. The high-speed image acquisition unit can use an industrial-grade camera with high dynamic range and low noise characteristics to adapt to the complex lighting conditions in the concentration kettle. The data verification unit can use machine learning algorithms, such as support vector machines or neural networks, to detect anomalies in the collected data and further improve the reliability of the data. In addition, the data acquisition module can be equipped with data caching and backup mechanisms to ensure that data is not lost in the event of equipment failure or network interruption. These optimization measures can further improve the stability of the system and the accuracy of the data.

[0084] In some embodiments, the data preprocessing module includes:

[0085] The adaptive filtering unit uses the following formula to eliminate high-frequency noise in the sensor signal:

[0086] y t =αx t +(1-α)y t-1

[0087] Among them, y t is the filtered signal at time t, x t is the original signal at time t, α is the dynamic smoothing coefficient, and its value range is [0.2, 0.8];

[0088] Timing alignment unit, used to map data with different sampling frequencies to a unified time axis through cubic spline interpolation;

[0089] Normalization unit, scaling the data to the [-1, 1] interval by batch.

[0090] It's important to note that the data preprocessing module in this system is responsible for noise removal, multi-source timing alignment, and normalization of the raw data to ensure data quality and usability. The adaptive filtering unit uses a specific formula to remove high-frequency noise from the sensor signal, the timing alignment unit maps data with different sampling frequencies onto a unified time axis, and the normalization unit scales the data to a specific interval. These processing steps are crucial for subsequent feature extraction and process optimization, effectively improving system stability and accuracy.

[0091] Specifically, the dynamic smoothing coefficient α in the adaptive filtering unit ranges from [0.2 to 0.8] and can be adjusted based on the actual noise level. The time series alignment unit uses a cubic spline interpolation algorithm to accurately align data with different sampling frequencies to a unified time axis, ensuring data synchronization. The normalization unit scales data batches to the range [-1, 1], which helps improve the training efficiency and stability of deep learning models. Furthermore, the data preprocessing module can also include data cleaning and outlier processing functions to further improve data quality. For example, outliers can be removed by setting a threshold, or statistical methods can be used to smooth the data.

[0092] Preferably, the adaptive filtering unit can adopt a variety of filtering algorithms, such as Kalman filtering or wavelet transform, to adapt to different types of noise. The timing alignment unit can adopt a more efficient interpolation algorithm, such as spline interpolation or polynomial interpolation, to improve the accuracy of data alignment. The normalization unit can adopt other scaling methods, such as Z-score standardization or Min-Max standardization, to meet the needs of different models. In addition, the data preprocessing module can also include a data enhancement function to increase the diversity of the data by adding noise or performing data transformation, thereby improving the generalization ability of the model. These optimization measures can further improve the effect of data preprocessing and provide more reliable data support for subsequent feature extraction and process optimization.

[0093] In some embodiments, the specific steps of the time-frequency fusion algorithm include:

[0094] Step 1: Perform wavelet packet decomposition on the sensor signal to extract the energy ratio characteristics of each frequency band

[0095]

[0096] Among them, W k is the kth wavelet coefficient, and N is the total number of coefficients;

[0097] Step 2: Use a 3D convolutional network to extract spatial-temporal features F from the image sequence v , and calculate the texture complexity of the keyframe

[0098]

[0099] Among them, p i is the probability of the i-th gray level in the gray level histogram, and M is the total number of gray levels;

[0100] Step 3: Fusion of time domain and frequency domain features through attention mechanism

[0101]

[0102] in, is the energy characteristic of the jth frequency band, is the jth space-time feature, w j is the dynamic weight of the jth feature, J is the total number of fusion features, Represents a feature concatenation operation.

[0103] It's important to note that the time-frequency fusion algorithm in this system is used to extract key features from sensor signals and image sequences, which are crucial for optimizing the traditional Chinese medicine concentration process. Specifically, the algorithm effectively fuses time-domain and frequency-domain features through wavelet packet decomposition, a three-dimensional convolutional network, and an attention mechanism. This process not only captures the dynamic changes in the boiling state of the medicinal liquid but also extracts deep-level features related to process quality, providing strong support for subsequent process optimization.

[0104] Specifically, during the wavelet packet decomposition step, the sensor signal is decomposed into multiple frequency bands, and the energy contribution characteristics of each frequency band are obtained by calculating the square sum of the wavelet coefficients. For example, a 4-layer decomposition using the db4 wavelet basis function can be used to extract the energy characteristics of different frequency bands. A three-dimensional convolutional network is used to process image sequences, extract spatial-temporal features, and calculate the texture complexity of key frames. Texture complexity is calculated based on the grayscale histogram, using a formula to calculate the probability distribution of grayscale levels, thereby obtaining the texture complexity value. The attention mechanism fuses time-domain and frequency-domain features through dynamic weights. The dynamic weights are calculated based on the weighted sum of the energy characteristics of each frequency band and the spatial-temporal characteristics, achieving effective feature fusion.

[0105] Preferably, the number of layers of wavelet packet decomposition and the wavelet basis function can be selected according to the actual signal characteristics to optimize the extraction effect of energy features. The structural parameters of the three-dimensional convolutional network, such as the convolution kernel size, network depth, etc., can be adjusted according to the characteristics of the image data to improve the accuracy of feature extraction. The dynamic weight calculation in the attention mechanism can use more complex functions, such as radial basis functions or Gaussian functions, to enhance the flexibility and effectiveness of feature fusion. In addition, in order to improve the robustness of the algorithm, data enhancement techniques can be introduced in the feature extraction process, such as adding noise, performing geometric transformations, etc., to enhance the model's adaptability to different working conditions.

[0106] In some embodiments, the objective function of the multi-objective optimization model is:

[0107]

[0108] Among them, C 实际 is the real-time concentration of active ingredients, C 目标 is the preset concentration threshold, ΔP is the pressure fluctuation variance, T(t) is the temperature deviation function, λ1, λ2, λ3 are dynamic penalty coefficients, and t0 and t1 are the process start and end times, respectively.

[0109] It's important to note that the multi-objective optimization model in this system aims to precisely optimize the Traditional Chinese Medicine (TCM) concentration process by comprehensively considering multiple key process indicators. With the objective function at its core, the model incorporates key parameters such as active ingredient concentration, pressure fluctuation, and temperature deviation. By introducing dynamic penalty coefficients, it enables real-time adjustment and optimization of the process. This process not only improves concentration efficiency but also ensures stable and consistent product quality, significantly enhancing the economic benefits and market competitiveness of the entire production process.

[0110] Specifically, the objective function consists of three components: active ingredient concentration deviation, pressure fluctuation variance, and the integral of temperature deviation. The active ingredient concentration deviation is calculated as the difference between the actual concentration and a preset threshold. The pressure fluctuation variance reflects the pressure stability during the process. The integral of the temperature deviation measures the accuracy of temperature control. Dynamic penalty coefficients α1, α2, and α3 are updated based on the process compliance loss function to dynamically adjust the weights of different process indicators. These parameters can be flexibly configured according to actual production needs to achieve optimal optimization results.

[0111] Preferably, the parameters in the objective function can be weighted based on the importance and sensitivity of the process. For example, deviation in active ingredient concentration can be assigned a higher weight to ensure stable product quality. The dynamic penalty coefficient update rule can use an adaptive algorithm to automatically adjust based on changes in process compliance rates to improve the model's adaptability and robustness. Furthermore, the multi-objective optimization model can incorporate other process indicators, such as energy consumption and production cycle time, to achieve more comprehensive process optimization.

[0112] In some embodiments, the update rule of the dynamic penalty coefficient is:

[0113]

[0114] in, is the updated i-th penalty coefficient, is the i-th penalty coefficient before update, L is the process compliance rate loss function, η is the learning rate, i∈{1,2,3}, and satisfies constraints.

[0115] It's important to note that the dynamic penalty coefficient update mechanism in this system dynamically adjusts the penalty coefficients in the multi-objective optimization model based on changes in the process compliance rate. This mechanism, through a combination of the learning rate and the process compliance rate loss function, enables adaptive updates of the penalty coefficients, thereby improving the model's optimization effectiveness and adaptability. This dynamic penalty coefficient update rule ensures the model's stability and effectiveness under varying process conditions, effectively addressing process uncertainties and changes.

[0116] Specifically, in the update rule of the dynamic penalty coefficient, the learning rate η controls the step size of the penalty coefficient update. The value range is usually between 0 and 1, for example, it can be set to 0.1 or 0.2 to ensure the stability and convergence speed of the update process. The process compliance rate loss function L reflects the deviation between the process compliance rate and the target value, and can be calculated and updated based on historical data. The update formula of the penalty coefficient is:

[0117]

[0118] in, and are the penalty coefficients before and after the update, respectively, and i represents the sequence number of the penalty coefficient. In addition, the update of the penalty coefficient must also meet the normalization constraint, that is, the sum of all penalty coefficients is 1, to ensure the rationality and effectiveness of the model.

[0119] Preferably, the learning rate η can be adaptively adjusted according to the changing trend of the process compliance rate. For example, when the process compliance rate is low, the learning rate can be increased to speed up the update speed of the penalty coefficient; when the process compliance rate is high, the learning rate can be reduced to improve the stability of the model. The process compliance rate loss function L can take a variety of forms, such as mean square error, absolute error or logarithmic loss, etc., to adapt to different process requirements and data characteristics. In addition, in order to prevent the penalty coefficient from taking negative values or exceeding a reasonable range, constraints can be added during the update process, such as setting upper and lower limits of the penalty coefficient to ensure the rationality and effectiveness of the model. In some embodiments, the real-time control module includes:

[0120] A fuzzy decision-making unit is used to generate a fuzzy set of control variables based on the optimization instructions and the current state of the equipment;

[0121] The anti-saturation regulation unit uses the following formula to limit the actuator output range:

[0122]

[0123] Among them, u 实际 is the actual output value of the actuator, u 指令 is the output value of the optimization instruction, u max is the maximum allowable output value of the actuator, u min is the minimum allowable output value of the actuator;

[0124] The dynamic compensation unit is used to predict and correct the actuator hysteresis effect based on the historical control error.

[0125] It should be noted that the real-time control module in this system is responsible for dynamically adjusting the operating parameters of the concentration equipment according to optimization instructions to achieve precise control of the traditional Chinese medicine concentration process. This module includes a fuzzy decision-making unit, an anti-saturation adjustment unit, and a dynamic compensation unit. These units work together to ensure stable and efficient operation of the equipment under different operating conditions. Through the intelligent control strategy of the fuzzy decision-making unit, combined with the output limitation of the anti-saturation adjustment unit and the hysteresis correction of the dynamic compensation unit, the real-time control module can effectively cope with various complex situations in the process, ensuring the smooth progress of the concentration process and the stability of product quality.

[0126] Specifically, the fuzzy decision-making unit uses a fuzzy logic control algorithm to generate a fuzzy set of control variables based on the optimization instructions and the current state of the equipment. This unit implements intelligent control of the equipment by setting a series of fuzzy rules, such as: if the temperature is too high and the pressure is normal, then appropriately reduce the heating power. This unit transforms complex control logic into simple fuzzy rules. The anti-saturation regulation unit prevents control instructions from exceeding the physical limitations of the equipment by limiting the output range of the actuator. Specifically, when the output value of the optimization instruction exceeds the maximum or minimum allowable output value of the actuator, the anti-saturation regulation unit adjusts the instruction value to within the allowable range to ensure safe operation of the equipment. The dynamic compensation unit uses a predictive algorithm based on historical control errors to correct for actuator hysteresis, thereby improving the response speed and accuracy of the control system. For example, by analyzing control errors over a period of time, the dynamic compensation unit can predict the current hysteresis and incorporate the corresponding compensation value into the control instruction, thereby achieving real-time correction of hysteresis.

[0127] Preferably, the fuzzy decision-making unit can adopt an adaptive fuzzy control algorithm to automatically adjust the fuzzy rules and membership functions according to changes in the process, thereby improving the adaptability and robustness of the control strategy. The anti-saturation adjustment unit can introduce a soft saturation strategy to gradually reduce the gain when approaching the saturation point to achieve a smoother control transition. The dynamic compensation unit can adopt a data-driven prediction model, such as a neural network or a support vector machine, to learn the laws of hysteresis characteristics based on historical data, thereby achieving more accurate hysteresis correction. In addition, the real-time control module can also be equipped with fault diagnosis and fault-tolerant control functions. When a device failure or abnormality is detected, it can automatically switch to a safe mode to ensure the continuity and stability of the production process.

[0128] In some embodiments, the quality assessment module performs the following operations:

[0129] Online detection unit, which measures the concentration of active ingredients in the drug solution in real time through near-infrared spectrometer;

[0130] Physical property analysis unit uses the fusion data of rotational viscometer and colorimeter to calculate the comprehensive quality score

[0131] Q=ω1S 浓度 +ω2S 粘度 +ω3S 色差

[0132] Among them, S 浓度 Score for active ingredient concentration, 粘度 S is the viscosity score, 色差 is the color difference score, ω1, ω2, ω3 are the preset weight coefficients;

[0133] The compliance judgment unit triggers a process abnormality alarm when Q<0.9.

[0134] It's important to note that the quality assessment module in this system is used to monitor the quality of the concentrated product online and calculate the process compliance rate based on the test results. This module uses a near-infrared spectrometer to measure the active ingredient concentration of the liquid in real time and calculates a comprehensive quality score using data from a rotational viscometer and a colorimeter. If the quality score falls below a preset threshold, the system triggers a process anomaly alarm. This mechanism ensures that the quality of the concentrated product remains within controllable limits, allowing for the timely identification and resolution of quality issues, thereby improving the stability and reliability of the production process.

[0135] Specifically, the near-infrared spectrometer in the online detection unit can quickly and non-destructively measure the concentration of active ingredients in the drug solution with a measurement accuracy of ±0.5%. The physical property analysis unit measures the viscosity of the drug solution using a rotational viscometer and the color of the drug solution using a colorimeter to calculate the overall quality score. The formula for calculating the overall quality score is: Q = w1×C + w2×V + w3×C col , where C represents the active ingredient concentration score, V represents the viscosity score, and C col represents the color difference score, while w1, w2, and w3 are preset weight coefficients that are adjusted based on actual process requirements. The compliance assessment unit sets the quality score threshold to 0.9. When the quality score falls below this threshold, the system determines that the process is abnormal and triggers an alarm.

[0136] Preferably, the online detection unit can use a variety of spectral techniques, such as UV-visible spectroscopy or Raman spectroscopy, to improve the accuracy and reliability of active ingredient concentration measurements. The physical property analysis unit can introduce more physical parameters, such as density or surface tension, to more comprehensively evaluate the quality of the drug solution. In addition, the calculation formula for the comprehensive quality score can be adjusted according to actual needs, such as using weighted average or other more complex fusion algorithms. When a process anomaly alarm is triggered, the system can automatically record the relevant data and generate a detailed alarm report for subsequent analysis and improvement.

[0137] In some embodiments, the iterative update process of the feedback adjustment module includes:

[0138] A parameter correction unit that adjusts the loss function weights of the deep learning model based on the quality score deviation;

[0139] The feature attenuation unit uses the following formula to eliminate invalid process features:

[0140]

[0141] in, is the updated weight of the i-th feature, is the weight of the i-th feature before updating, τ is the decay coefficient, and Δt is the duration that the feature is not effectively activated;

[0142] The model reconstruction unit reinitializes the network hidden layer when the cumulative quality deviation exceeds a threshold.

[0143] It should be noted that the feedback adjustment module in this system is responsible for iteratively updating the parameters of the deep learning model based on quality assessment results and historical optimization data. This module includes a parameter correction unit, a feature decay unit, and a model reconstruction unit. Through the collaborative work of these units, continuous optimization and improvement of the model is achieved. The parameter correction unit adjusts the model's loss function weights based on quality score deviations, the feature decay unit eliminates invalid process features, and the model reconstruction unit reinitializes the network's hidden layers when the cumulative quality deviation exceeds a threshold. This mechanism can effectively improve the model's accuracy and adaptability, ensuring that the system maintains good performance under different operating conditions.

[0144] Specifically, the parameter correction unit adjusts the loss function weight of the deep learning model based on the quality score deviation ΔQ. The specific adjustment method is: multiply the weight parameter in the loss function by (1+η×ΔQ), where η is the learning rate, ranging from 0.01 to 0.1. The feature decay unit eliminates features that are not effectively activated by calculating the weight decay coefficient of the feature. The specific formula is: in and are the weights after feature update and before update, β is the decay coefficient, t i is the duration that the feature is not effectively activated. When the cumulative quality deviation exceeds the preset threshold, the model reconstruction unit reinitializes the weight parameters of the network hidden layer to restore the model performance.

[0145] Preferably, the parameter correction unit can adopt an adaptive learning rate η, dynamically adjusting it based on the changing trend of the quality score deviation to improve the convergence speed and stability of the model. The feature decay unit can set a threshold for feature activation. When the activation value of a feature falls below this threshold, the feature is considered ineffectively activated and decayed. The model reconstruction unit can adopt a partial initialization strategy, reinitializing only the weight parameters of some hidden layers to reduce the forgetting of existing knowledge during model reconstruction. In addition, the feedback adjustment module can also introduce a model verification mechanism to verify the model after it is updated to ensure improved model performance.

[0146] In some embodiments, the dynamic weight in the time-frequency fusion algorithm is calculated as follows:

[0147]

[0148] in, is the energy characteristic of the jth frequency band, is the j-th spatial-temporal visual feature, σ(·) is the Sigmoid function, tanh(·) is the hyperbolic tangent function, and K is the total number of fused features.

[0149] It should be noted that the dynamic weight calculation method of the time-frequency fusion algorithm in this system uses a specific formula to assign a dynamic weight to each feature based on the energy characteristics and spatial-temporal visual characteristics of each frequency band. This mechanism can adaptively highlight the features that have a greater impact on process optimization, thereby improving the effect of feature fusion and the accuracy of the model. Dynamic weight α j The calculation formula is

[0150]

[0151] Here, σ is the sigmoid function, and tanh is the hyperbolic tangent function. By nonlinearly combining energy features with visual features, we can flexibly adjust feature weights to better capture key information in the traditional Chinese medicine concentration process.

[0152] Specifically, in the time-frequency fusion algorithm, the calculation process of the dynamic weight is as follows: First, the energy characteristics E of each frequency band are j Normalization is performed to obtain the normalized energy feature value; then, the normalized energy feature value is input into the Sigmoid function to obtain a value between 0 and 1, which represents the relative importance of the energy feature; at the same time, the spatial-temporal visual feature C j Apply the hyperbolic tangent function and map it to the nonlinear space to extract the sensitive information of the feature; finally, multiply the output of the Sigmoid function and the hyperbolic tangent function element by element, and calculate the weighted sum of all frequency bands to obtain the dynamic weight α j For example, when the energy characteristics and visual characteristics of a certain frequency band are both high, the corresponding dynamic weight α j It will also be larger, thus being given a higher weight in subsequent feature fusion.

[0153] Preferably, to further improve the accuracy of dynamic weight calculation, a variant of the attention mechanism, such as the self-attention mechanism or the cross-attention mechanism, can be introduced into the feature fusion process. The self-attention mechanism can focus on the features of each frequency band and automatically adjust the weights of each feature point; the cross-attention mechanism can establish more complex associations between energy features and visual features, better reflecting the interaction between the two. In addition, deep neural networks, such as multi-layer perceptrons or graph neural networks, can be used to perform end-to-end learning of dynamic weights, automatically adjusting the parameters in the weight calculation formula based on a large amount of historical data, thereby improving the model's adaptability and generalization performance.

[0154] The above-described embodiments of the present invention have the following beneficial effects: The system can achieve intelligent optimization of the traditional Chinese medicine concentration process. Using deep learning technology, various data from the concentration process are analyzed and processed in real time, thereby precisely adjusting process parameters and improving concentration efficiency and product quality. Furthermore, the system can detect the quality of the concentrated product online and automatically adjust process parameters based on the test results to ensure the stability and consistency of product quality. Furthermore, the system can continuously optimize the deep learning model through a feedback adjustment mechanism, improving the model's accuracy and adaptability, further enhancing the system's performance and effectiveness. This intelligent optimization system can effectively address problems existing in traditional traditional Chinese medicine concentration processes and improve the modernization and competitiveness of traditional Chinese medicine production. The system can be widely used in traditional Chinese medicine pharmaceutical companies, helping them improve production efficiency, reduce costs, and enhance product quality, thereby enhancing market competitiveness. Furthermore, the system can provide strong technical support for the research and development of traditional Chinese medicine concentration processes, promoting technological innovation and development in the traditional Chinese medicine pharmaceutical industry.

[0155] Furthermore, the storage medium of the embodiment of the present application stores program instructions that can implement all the above methods, wherein the program instructions can be stored in the above storage medium in the form of a software product, including a number of instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) or a processor to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, or a terminal device such as a computer, a server, a mobile phone, or a tablet.

[0156] The above descriptions are merely some preferred embodiments of the present invention and an illustration of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present invention is not limited to the technical solutions formed by a specific combination of the above-mentioned technical features, but should also encompass other technical solutions formed by any combination of the above-mentioned technical features or their equivalents without departing from the above-mentioned inventive concept. For example, a technical solution formed by mutually replacing the above-mentioned features with (but not limited to) technical features having similar functions disclosed in the embodiments of the present invention.

Claims

1. A Chinese medicine concentration process optimization system based on deep learning, characterized in that: Includes the following modules: Data acquisition module, used to obtain real-time sensor data and image data of temperature, pressure, flow rate and medicinal material composition during the Chinese medicine concentration process; Data preprocessing module, used to eliminate noise, align multi-source time series and standardize the raw data; Feature extraction module, used to extract process dynamic features through time-frequency fusion algorithm and deep learning model; Process optimization module, which is used to build a multi-objective optimization model based on characteristic data and generate adjustment instructions for concentration process parameters; A real-time control module is used to dynamically adjust the heating power, vacuum level, and feed rate of the concentration equipment according to optimization instructions; The quality assessment module is used to detect the active ingredient concentration, viscosity and color of the concentrated product online and calculate the process compliance rate; Feedback adjustment module, used to iteratively update deep learning model parameters based on quality assessment results and historical optimization data.

2. The system according to claim 1, wherein: The data acquisition module includes: Multimodal sensor unit, used to synchronously collect output signals from temperature sensor, pressure transmitter and spectrum analyzer; High-speed image acquisition unit, used to capture high-definition image sequences of the boiling state of the chemical liquid in the concentration kettle; The data verification unit is used to eliminate abnormal collected values based on the physical constraints of the device.

3. The system according to claim 1, wherein: The data preprocessing module includes: The adaptive filtering unit uses the following formula to eliminate high-frequency noise in the sensor signal: y t =αx t +(1-α)y t-1 Among them, y t is the filtered signal at time t, x t is the original signal at time t, α is the dynamic smoothing coefficient; Timing alignment unit, used to map data with different sampling frequencies to a unified time axis through cubic spline interpolation; Normalization unit, scaling the data to the [-1, 1] interval by batch.

4. The system according to claim 3, characterized in that The specific steps of the time-frequency fusion algorithm include: Step 1: Perform wavelet packet decomposition on the sensor signal to extract the energy ratio characteristics of each frequency band: Among them, W k is the kth wavelet coefficient, and N is the total number of coefficients; Step 2: Use a 3D convolutional network to extract spatial-temporal features F from the image sequence v , and calculate the texture complexity of the keyframe: Among them, p i is the probability of the i-th gray level in the gray level histogram, and M is the total number of gray levels; Step 3: Fusion of time domain and frequency domain features through attention mechanism: in, is the energy characteristic of the jth frequency band, is the jth space-time feature, w j is the dynamic weight of the jth feature, J is the total number of fusion features, Represents a feature concatenation operation.

5. The system according to claim 1, wherein: The objective function of the multi-objective optimization model is: Among them, C 实际 is the real-time concentration of active ingredients, C 目标 is the preset concentration threshold, ΔP is the pressure fluctuation variance, T(t) is the temperature deviation function, λ1, λ2, λ3 are dynamic penalty coefficients, and t0 and t1 are the process start and end times, respectively.

6. The system according to claim 5, characterized in that The updating rule of the dynamic penalty coefficient is: in, is the updated i-th penalty coefficient, is the i-th penalty coefficient before update, L is the process compliance rate loss function, η is the learning rate, i∈{1,2,3}, and satisfies constraints.

7. The system according to claim 1, wherein: The real-time control module includes: A fuzzy decision-making unit is used to generate a fuzzy set of control variables based on the optimization instructions and the current state of the equipment; The anti-saturation regulation unit uses the following formula to limit the actuator output range: Among them, u 实际 is the actual output value of the actuator, u 指令 is the output value of the optimization instruction, u max is the maximum allowable output value of the actuator, u min is the minimum allowable output value of the actuator; The dynamic compensation unit is used to predict and correct the actuator hysteresis effect based on the historical control error.

8. The system according to claim 1, wherein: The quality assessment module performs the following operations: Online detection unit, which measures the concentration of active ingredients in the drug solution in real time through near-infrared spectrometer; Physical property analysis unit uses the fusion data of rotational viscometer and colorimeter to calculate the comprehensive quality score Q=ω1S 浓度 +ω2S 粘度 +ω3S 色差 Among them, S 浓度 Score for active ingredient concentration, 粘度 S is the viscosity score, 色差 is the color difference score, ω1, ω2, ω3 are the preset weight coefficients; The compliance judgment unit triggers a process abnormality alarm when Q<0.

9.

9. The system according to claim 8, characterized in that The iterative update process of the feedback adjustment module includes: A parameter correction unit that adjusts the loss function weights of the deep learning model based on the quality score deviation; The feature attenuation unit uses the following formula to eliminate invalid process features: in, is the updated weight of the i-th feature, is the weight of the i-th feature before updating, τ is the decay coefficient, and Δt is the duration that the feature is not effectively activated; The model reconstruction unit reinitializes the network hidden layer when the cumulative quality deviation exceeds a threshold.

10. The system according to claim 1, wherein: The calculation method of dynamic weight in the time-frequency fusion algorithm is: in, is the energy characteristic of the jth frequency band, is the j-th spatial-temporal visual feature, σ(·) is the Sigmoid function, tanh(·) is the hyperbolic tangent function, and K is the total number of fused features.

Images