Intelligent distillation optimization method and system based on artificial intelligence

Abstract

The invention relates to the technical field of intelligent distillation control, and discloses an intelligent distillation optimization method and system based on artificial intelligence. The method comprises the following steps: acquiring spectral data of a fermented grain raw material and initial operation parameters of distillation equipment, generating a characteristic distribution state of the fermented grain raw material through a convolutional neural network, and triggering distillation process monitoring according to a comparison result of the characteristic distribution state and a preset threshold range; after monitoring is triggered, pressure fluctuation data, fraction temperature gradient data and wine steam component concentration data of a distillation tower are synchronously collected, and multi-dimensional time sequence data are generated through timestamp alignment processing; generating a target intrinsic mode function through variational mode decomposition, and removing the intrinsic mode function with the maximum composite entropy total value to obtain an effective distillation signal; abnormal control parameters are identified and eliminated based on the effective distillation signals, and an optimization control parameter set is formed; and performing iterative optimization through a cattle herd optimization algorithm to generate optimal control parameters, and adjusting operation parameters of the distillation equipment according to the optimal control parameters.

Description

Technical Field

[0001] This invention relates to the field of intelligent distillation control technology, specifically to an intelligent distillation optimization method and system based on artificial intelligence. Background Technology

[0002] In the winemaking and related distillation industries, distillation is a crucial process determining product quality and production efficiency, and its process control has long relied on traditional experience-based operating methods. In traditional distillation, operators primarily rely on their accumulated experience, manually adjusting operating parameters by observing the apparent parameters of the distillation equipment. This method has significant limitations. Experience-based judgment is subjective; different operators have different judgment standards, leading to difficulties in ensuring the stability of the distillation process and significant fluctuations in product quality. Manual adjustments also have a lag; when abnormal parameters occur during distillation, operators cannot react accurately in a timely manner, potentially missing the optimal adjustment opportunity, thus affecting product yield and quality. With the development of industrial automation technology, some distillation production scenarios have begun to introduce automated control systems. These systems collect key parameters during the distillation process using sensors and adjust these parameters according to preset fixed algorithms. However, these automated control systems still have shortcomings. The distillation process is significantly affected by the characteristics of the mash raw materials. Different batches of mash raw materials have different characteristics, and fixed algorithms cannot dynamically adjust the control strategy according to changes in raw material characteristics, resulting in poor adaptability of the system to changes in raw material characteristics. Various interference factors exist during the distillation process, and the data collected by sensors is easily affected. Traditional automated control systems lack effective data processing mechanisms, making it difficult to accurately identify valid signals and interference signals, thus affecting the accuracy of parameter adjustments. The application of artificial intelligence (AI) technology in the industrial field is becoming increasingly widespread, providing new ideas for optimizing distillation processes. However, research on applying AI technology to distillation optimization is still in its early stages. Most existing technologies only analyze single parameters in the distillation process (such as fraction temperature), failing to comprehensively consider the correlation between the characteristics of the mash raw material and multiple parameters of the distillation process, resulting in limited optimization effects. Furthermore, in terms of data processing, existing technologies have not effectively solved the problems of synchronizing and denoising multi-dimensional time-series data, making it difficult to accurately extract key information from the distillation process and providing a reliable basis for optimizing distillation parameters. Therefore, the distillation industry urgently needs an intelligent optimization method that can combine the characteristics of the mash raw material, accurately process multi-dimensional distillation data, and dynamically adjust control strategies to improve the stability of the distillation process and product quality. Summary of the Invention

[0003] The purpose of this invention is to provide an intelligent distillation optimization method and system based on artificial intelligence to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides an intelligent distillation optimization method based on artificial intelligence, the method comprising: Obtain spectral data of the mash raw material and the initial operating parameters of the distillation equipment; Based on the spectral data, a convolutional neural network is used to generate the characteristic distribution state of the fermentation mash raw materials; Distillation process monitoring is triggered based on the comparison between the characteristic distribution state and the preset threshold range; When the distillation process monitoring is triggered, data on distillation column pressure fluctuations, distillate temperature gradients, and alcohol vapor concentrations are collected simultaneously. The distillation column pressure fluctuation data, distillation temperature gradient data, and alcohol vapor component concentration data are subjected to timestamp alignment processing to generate multi-dimensional time-series data. Based on the multi-dimensional time-series data, a target number of intrinsic mode functions are generated through variational mode decomposition. The effective distillation signal is obtained by removing the intrinsic mode function with the largest total composite entropy from the target intrinsic mode functions; Based on the effective distillation signal, abnormal control parameters are identified by comparing discrete data, and the abnormal control parameters are eliminated and replaced to obtain an optimized control parameter set. Based on the optimized control parameter set, the optimal control parameters are generated iteratively using the cattle herd optimization algorithm. Adjust the operating parameters of the distillation equipment according to the optimal control parameters.

[0005] Preferably, the step of generating the characteristic distribution state of the fermentation mash raw materials based on the spectral data using a convolutional neural network includes: Set the initial kernel size and initial number of pooling layers for the convolutional neural network; Based on the spectral data, frequency domain feature vectors are extracted, and the frequency domain feature vectors are convolved with an initial convolution kernel to generate a primary feature map. Max pooling is performed on the primary feature map to generate a compressed feature map; Based on the compressed feature map, a feature distribution state vector is generated through a fully connected layer; Based on the matching degree between the feature distribution state vector and the historical feature library, the initial convolution kernel size and the initial number of pooling layers are corrected to generate the target convolution kernel size and the target number of pooling layers; The target convolution kernel size and target pooling layer number are used to perform secondary feature extraction on the spectral data to generate the characteristic distribution state of the mash raw materials.

[0006] Preferably, the time-stamp alignment processing of the distillation column pressure fluctuation data, fraction temperature gradient data, and alcohol vapor concentration data includes: Acquire the first sampling time series of the distillation column pressure fluctuation data, the second sampling time series of the distillation temperature gradient data, and the third sampling time series of the alcohol vapor component concentration data; Using the first sampling time series as a reference, the second and third sampling time series are aligned to the first sampling time series using linear interpolation. Generate multi-dimensional time series data with a unified timestamp based on the aligned time series.

[0007] Preferably, the step of generating a target number of intrinsic mode functions based on the multi-dimensional time-series data through variational mode decomposition includes: Set the initial number of decomposition layers and the initial penalty factor for variational mode decomposition; The fitness function value is generated based on the multi-dimensional time series data and the initial intrinsic mode functions. Based on the fitness function value, the initial decomposition layer number and initial penalty factor are optimized using the particle swarm optimization algorithm to generate the target decomposition layer number and target penalty factor; Variational mode decomposition is performed on the multi-dimensional time series data using the target decomposition level and the target penalty factor to generate target intrinsic mode functions.

[0008] Preferably, the step of generating fitness function values ​​based on the multi-dimensional time-series data and the initial intrinsic mode functions includes: The multi-dimensional time series data and each eigenmode function of the initial eigenmode function are divided into N sub-signals; Calculate the weighted sum of the envelope entropy and sample entropy of each segment of the multi-dimensional time series data to generate the composite entropy of the time series data; Calculate the weighted sum of the envelope entropy and sample entropy of each sub-signal segment of each intrinsic mode function to generate the composite entropy of the mode function; The fitness function value is generated based on the correlation coefficient between the composite entropy of the time series data and the composite entropy of the modal function.

[0009] Preferably, the step of removing the intrinsic mode function with the largest total composite entropy from the target intrinsic mode functions includes: Calculate the total composite entropy of each eigenmode function in the target intrinsic mode functions; Identify the eigenmode function with the largest total composite entropy; The effective distillation signal is generated by superimposing all eigenmode functions before the eigenmode function with the largest total composite entropy value and all eigenmode functions after it.

[0010] Preferably, the step of identifying abnormal control parameters by comparing discrete data based on the effective distillation signal includes: Set a dynamic threshold coefficient, and calculate the average value and standard deviation based on the effective distillation signal; The effective data interval is generated by using the product of the average value and the standard deviation multiplied by the dynamic threshold coefficient as the boundary; Traverse the discrete data in the valid distillation signal and mark the discrete data that does not belong to the valid data range as abnormal control parameters.

[0011] Preferably, the removal and replacement of the abnormal control parameters includes: The arithmetic mean of the previous and next data points of the abnormal control parameter is used as the replacement value; Replace the abnormal control parameters with the alternative values ​​to generate an optimized control parameter set.

[0012] Preferably, the step of generating optimal control parameters through iterative optimization using a cattle herd optimization algorithm based on the optimized control parameter set includes: The optimized control parameter set is used as the initial population for the cattle herd optimization algorithm; The fitness function is the weighted sum of distillation efficiency and energy consumption ratio. The population position is iteratively updated using the cattle herd optimization algorithm until the fitness function value converges. The population-optimal solution at convergence is used as the optimal control parameter. The step of adjusting the operating parameters of the distillation equipment according to the optimal control parameters includes: The pressure, temperature, and flow parameters in the optimal control parameters are respectively mapped to the pressure controller, temperature sensor, and flow regulating valve of the distillation column; The operating parameters of the distillation equipment are adjusted in real time based on the mapping results.

[0013] Preferably, the present invention also includes an artificial intelligence-based intelligent distillation optimization system, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor, when executing the computer program, implements the steps of the above-described artificial intelligence-based intelligent distillation optimization method.

[0014] Compared with the prior art, the beneficial effects of the present invention are: This AI-based intelligent distillation optimization method effectively overcomes the limitations of traditional distillation control methods by analyzing the characteristics of the mash raw materials and the multi-parameter collaborative analysis of the distillation process. In the raw material characteristic analysis stage, a convolutional neural network is used to process the spectral data of the mash raw materials to generate the characteristic distribution state of the raw materials. This accurately captures the characteristic differences between different batches of mash raw materials, providing a basis for dynamic adjustments to the subsequent distillation process. It avoids the rigidity of control strategies caused by neglecting changes in raw material characteristics in traditional methods, enabling the distillation process to better adapt to the characteristics of the raw materials and reducing product quality fluctuations caused by raw material differences. In terms of distillation process monitoring and data processing, this method, upon triggering distillation process monitoring, simultaneously collects distillation column pressure fluctuation data, fraction temperature gradient data, and alcohol vapor component concentration data, and performs timestamp alignment processing to generate multi-dimensional time-series data. This solves the problems of asynchronous parameter acquisition and single data dimension in traditional methods. Multi-dimensional time-series data can comprehensively reflect the true state of the distillation process, providing rich information support for subsequent data analysis and parameter optimization. Simultaneously, by generating target intrinsic mode functions through variational mode decomposition and removing the intrinsic mode function with the largest total composite entropy to obtain effective distillation signals, this method effectively filters out interference signals during the distillation process, improves data accuracy and reliability, avoids parameter judgment errors caused by interference signals, and provides a precise data foundation for optimizing distillation parameters. In the parameter optimization stage, this method identifies abnormal control parameters through discrete data comparison, and then eliminates or replaces them to form an optimized control parameter set. This allows for the timely detection of parameter anomalies during the distillation process, preventing these abnormal parameters from having a lasting impact on the process. Subsequently, the optimized control parameter set is iteratively optimized using the bovine herd optimization algorithm to generate the optimal control parameters. Compared to traditional fixed algorithms or single-parameter optimization methods, the bovine herd optimization algorithm has a stronger global optimization capability. It can find the optimal parameter combination that best suits the current distillation state in complex distillation systems with multiple parameters working together, achieving precise control of the distillation process. This method enables intelligent and dynamic control of the distillation process, eliminating reliance on operator experience and reducing the impact of human factors, thus improving process stability. Simultaneously, through precise raw material characteristic analysis and multi-dimensional data processing, the distillation equipment can always operate at optimal parameters, contributing to improved product quality consistency and reducing raw material waste and energy consumption caused by improper parameters. Furthermore, the application of this method reduces operator workload, lowers manual labor intensity, and increases production efficiency, providing a feasible technical path for the intelligent upgrading of the distillation industry. It promotes the transformation of distillation production from traditional experience-driven to data- and intelligent algorithm-driven approaches, adapting to the needs of modern industrial intelligent development. Attached Figure Description

[0015] Figure 1 This is a schematic diagram illustrating the working principle of the AI-based intelligent distillation optimization method described in this invention. Figure 2 A flowchart for generating the characteristic distribution state of fermentation mash raw materials based on spectral data; Figure 3 This is a flowchart for performing timestamp alignment processing on multiple types of data. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Please see Figure 1 This invention provides an intelligent distillation optimization method and system based on artificial intelligence. The method includes acquiring spectral data of the raw materials (fermented mash) and initial operating parameters of the distillation equipment. The spectral data can be acquired using a near-infrared spectrometer or a Raman spectrometer. The initial operating parameters include preset pressure, temperature curves, and feed flow rate of the distillation column. Based on the spectral data, a convolutional neural network is used to generate the characteristic distribution state of the raw materials. The convolutional neural network uses a multi-layer structure to process the frequency domain information in the spectral data and outputs feature vectors representing the components of the raw materials. Distillation process monitoring is triggered based on the comparison between the characteristic distribution state and a preset threshold range. The preset threshold range is set based on historical high-quality fermented mash data. When the feature vector exceeds the threshold, the monitoring process is initiated. When distillation process monitoring is triggered, distillation column pressure fluctuation data, fraction temperature gradient data, and alcohol vapor concentration data are simultaneously acquired. Pressure fluctuation data is acquired through high-frequency sampling by a pressure sensor, temperature gradient data is recorded by multiple temperature sensors, and alcohol vapor concentration is measured by an online gas chromatograph. Timestamp alignment is performed on the distillation column pressure fluctuation data, fraction temperature gradient data, and alcohol vapor concentration data to generate multi-dimensional time-series data. Timestamp alignment ensures that data points with different sampling rates are aligned on the same time base.

[0018] Based on multi-dimensional time-series data, a target number of intrinsic mode functions (EMFs) are generated through variational mode decomposition (VMD). VMD decomposes the time-series data into a series of EMFs, each representing a different frequency component. The EMF with the largest total composite entropy is removed from the target number of EMFs to obtain the effective distillation signal. Composite entropy calculation combines envelope entropy and sample entropy to quantify the disorder of the mode functions. Abnormal control parameters are identified by comparing discrete data with the effective distillation signal. These abnormal control parameters are then removed and replaced to obtain an optimized control parameter set. Abnormal parameter identification is based on statistical boundary judgment. The optimal control parameters are generated iteratively using a bovine herd optimization algorithm, which searches using distillation efficiency and energy consumption ratio as fitness functions. The operating parameters of the distillation equipment are adjusted according to the optimal control parameters, including adjusting the pressure controller, temperature setpoint, and flow valve opening.

[0019] Example 1: See Figure 2 In the process of generating the characteristic distribution state of the mash raw materials, the initial convolutional kernel size of the convolutional neural network is set according to the number of specific bands of the collected spectral data. For example, when processing data acquired by a near-infrared spectrometer, the width of the initial convolutional kernel is usually set to 5. This size can effectively capture the frequency domain variation pattern of the local range in the spectral curve. The initial pooling layer is dynamically adjusted according to the length of the input spectral data. If the number of data points is large, three pooling layers are set to gradually compress the data dimension. If the number of data points is relatively small, two pooling layers are selected to balance information retention and computational efficiency. Extracting the frequency domain feature vector based on the spectral data involves converting the original light intensity-wavelength sequence to the frequency domain through a fast Fourier transform, generating a set of complex sequences containing amplitude and phase information. When convolving this frequency domain vector with the initial convolutional kernel, a sliding window with a stride of 1 is used to traverse and calculate along the frequency axis. The data in each window is multiplied by the weight of the convolutional kernel, summed, and a bias term is added. Finally, a feature map containing local frequency response features is output. Each element of this feature map represents the activation intensity of the original spectrum in the corresponding frequency region.

[0020] When performing max pooling on the primary feature map, a 2x2 sampling window is used. This window slides across the feature map with a stride of 2 and extracts the maximum value within each region. This operation reduces the spatial resolution of the feature map while highlighting salient features, reducing subsequent computation while enhancing the spatial invariance of the features. The compressed feature map obtained after pooling not only has reduced dimensionality but also exhibits more robust feature responses. The process of generating a feature distribution state vector from the compressed feature map through a fully connected layer involves flattening the two-dimensional compressed feature map into a one-dimensional vector and inputting it into a fully connected network with multiple hidden layers. Each neuron in the fully connected layer is connected to all neurons in the previous layer and transformed using a nonlinear activation function such as ReLU. Finally, the output layer generates a fixed-dimensional feature distribution state vector, which comprehensively represents the overall compositional characteristics of the mash raw materials. When correcting the initial convolutional network parameters based on the matching degree between the feature distribution state vector and the historical feature library, the historical feature library stores the feature vectors of previously successful batches of fermented mash as a reference standard. The matching degree calculation uses the cosine similarity algorithm to measure the directional consistency between the current feature vector and the historical vector. If the similarity is lower than the preset threshold, it indicates that the initial convolutional kernel size or pooling layer number setting is insufficient to effectively extract features. At this time, the system will automatically increase the convolutional kernel size to expand the receptive field or increase the number of pooling layers to enhance the abstraction ability, thereby generating a more optimized target convolutional kernel size and target pooling layer number.

[0021] Secondary feature extraction of spectral data using the target convolution kernel size and target pooling layer number is a reprocessing step. The original spectral data is convolved again using updated network parameters. Adjusting the convolution kernel size captures a wider or narrower frequency range of features, while changing the number of pooling layers adjusts the feature compression level. The new feature map generated by the secondary convolution is compressed by pooling and then input into the fully connected layer. The final output feature distribution state vector more accurately represents the intrinsic properties of the mash raw materials. This vector is used to compare with a preset threshold range to determine whether to trigger the distillation monitoring process. The entire feature extraction process fully considers the characteristics of spectral data and ensures reliable feature representations under different batch data conditions through an adaptive parameter adjustment mechanism.

[0022] Example 2: See Figure 3The first step in timestamp alignment is to acquire the first sampling time series of distillation column pressure fluctuation data, the second sampling time series of distillation temperature gradient data, and the third sampling time series of alcohol vapor concentration data. These data come from different sensor systems installed on the distillation equipment. Pressure sensors typically collect data at millisecond intervals to form a high-density time series, while temperature sensors record temperature changes on each tray at a second-level frequency. However, online gas chromatographs generate relatively sparse data points due to analysis cycle limitations. This multi-source heterogeneous sampling mechanism leads to time base differences in the original data. Using the first sampling time series as the base means using the high-frequency timestamps of the pressure data as a unified time axis, because pressure changes often best reflect the dynamic processes inside the distillation column. When aligning the second and third sampling time series to the first sampling time series using linear interpolation, for each pressure data point's timestamp, the system finds two adjacent time points in the temperature series and calculates the estimated temperature value at that pressure timestamp based on the temperature values ​​corresponding to these two time points. The same method is applied to the alcohol vapor concentration data. This interpolation method can achieve data synchronization without introducing significant errors. Based on the aligned time series, multi-dimensional time series data with unified timestamps are generated, ultimately forming a regular data matrix. Each row of the matrix corresponds to a precise time point, and each column contains the pressure reading, temperature measurement value, and component concentration analysis results at that moment.

[0023] The process of generating target intrinsic mode functions (EMFs) through variational mode decomposition begins with setting the initial decomposition level and initial penalty factor. The initial decomposition level is usually set empirically based on the length and complexity of the time series data. For typical distillation process data, the level may be set between 5 and 8 to balance decomposition accuracy and computational burden. The initial penalty factor affects the bandwidth constraint of the modal components and is generally taken in the range of 1000 to 2000 to control the smoothness of the modes. When generating fitness function values ​​based on multi-dimensional time series data and the initial EMFs, the system first performs a preliminary decomposition on the original multi-dimensional time series data to obtain a set of tentative mode functions. Then, the decomposition quality is evaluated by calculating the correlation between the composite entropy of these mode functions and the original data. A higher fitness function value indicates that the decomposition result retains more of the essential characteristics of the original data. When optimizing the initial decomposition level and initial penalty factor using the particle swarm optimization algorithm based on the fitness function value, the algorithm treats each parameter combination as the position of a particle in the search space. It iteratively updates the particle's velocity and position to find the optimal parameter configuration that maximizes the fitness function. This process involves a collaborative and competitive mechanism of swarm intelligence, ultimately outputting the target decomposition level and target penalty factor. When performing variational mode decomposition on multi-dimensional time-series data using the target decomposition level and target penalty factor, the variational mode decomposition algorithm adaptively decomposes the original signal into a series of intrinsic mode functions (IMFs) by constructing and solving a constrained variational problem. Each IMF is an amplitude-modulated (AM) or frequency-modulated (FM) signal with a specific center frequency. The set of these functions can comprehensively characterize the dynamic features of different time scales in the distillation process data.

[0024] Taking the distillation process of baijiu (Chinese liquor) production as an example, a distillation system is in operation. Its pressure sensors collect data at a frequency of 100Hz, a temperature sensor array records the temperature distribution of each tray at a frequency of 10Hz, and an online gas chromatograph outputs a set of vapor component concentration data every 10 seconds. Due to differences in the sampling mechanisms of different devices, the three types of data have inconsistent time bases, leading to biases when directly combined for analysis. Timestamp alignment processing first obtains the time series of these three types of data: pressure fluctuation data forms the first time series from high-frequency acquisition, temperature gradient data forms the second time series from medium-frequency acquisition, and vapor component data forms the third time series from low-frequency acquisition. Using the first time series as the benchmark, the system selects the timestamp of the pressure data as the unified time axis, as it best reflects the instantaneous dynamic changes of the process. For the timestamp of each pressure data point, the temperature data is calculated using linear interpolation: assuming at a certain pressure sampling time... The previous record point in the temperature time series is ( , The next record point is (). , If the estimated temperature value at that moment is given, then... From the formula The calculation yielded the results. The same method was applied to the alcohol vapor composition data, but due to the larger sampling interval, special attention needed to be paid to smoothing the data changes during interpolation.

[0025] After timestamp alignment, each pressure data point corresponds to an interpolated temperature and component concentration value, forming a multi-dimensional time-series data matrix with a unified timestamp. This matrix contains three columns of data: pressure readings, temperature measurements, and component concentration values, with the same number of rows as the pressure data points and a time interval of 0.01 seconds. This well-organized data structure facilitates subsequent signal decomposition. In the variational mode decomposition stage, the initial decomposition layer number is set to 6, and the initial penalty factor is set to 1500. These parameters are based on the typical characteristics of the distillation process data. The particle swarm optimization algorithm is used to optimize these two parameters. The algorithm initializes a group of particles, each representing a combination of decomposition layer number and penalty factor. The optimal parameters are found by iteratively calculating the fitness function value under each parameter combination. The fitness function is evaluated based on the composite entropy correlation between the decomposed mode function and the original data; a higher correlation value indicates a better decomposition effect. The optimized target decomposition layer number and target penalty factor are used to perform variational mode decomposition on the multi-dimensional time-series data. This decomposition process adaptively decomposes the three-channel time-series data into several intrinsic mode functions. Each function represents a signal component in a different frequency range, and these components together characterize the dynamic characteristics of the distillation process at different time scales.

[0026] Example 3: In the process of generating fitness function values, the multi-dimensional time series data and the initial intrinsic mode functions (IMFs) are first divided into N segments. The number of segments N needs to be determined by comprehensively considering the total data length and the minimum data volume requirement for each segment. It is usually dynamically calculated based on the sampling frequency and process duration to ensure that each segment contains enough data points for effective entropy calculation. Equal-length segmentation is used. If the total data length is not divisible by N, the last segment is appropriately extended or truncated. When calculating the weighted sum of the envelope entropy and sample entropy of each segment of the multi-dimensional time series data, the envelope entropy is calculated by first extracting the signal envelope through Hilbert transform, and then calculating the information entropy value of the envelope, which reflects the uniformity of the signal energy distribution. The sample entropy measures the complexity and irregularity of the signal sequence. The randomness of the time series is quantified by calculating the probability of pattern matching. The weight coefficients of the two entropy values ​​in the weighted sum calculation are adjusted according to the specific application scenario. Usually, the weight of the envelope entropy is slightly higher than that of the sample entropy to emphasize the energy characteristics of the signal.

[0027] When calculating the weighted sum of the envelope entropy and sample entropy of each segment of the intrinsic mode function (EMF), each EMF is independently segmented and its entropy value is calculated. The composite entropy value of each EMF characterizes the complexity and energy distribution characteristics of that frequency component in different time periods. High-frequency modes typically have higher composite entropy values, while low-frequency modes have relatively lower values. When generating the fitness function value based on the correlation coefficient between the composite entropy of time-series data and the composite entropy of the EMFs, the Pearson correlation coefficient method is used to measure the linear correlation between the original data and each EMF in terms of composite entropy characteristics. A higher correlation coefficient indicates that the EMF better represents the essential characteristics of the original data. The fitness function value is defined as a combination function of the average and standard deviation of the correlation coefficients of each EMF. in: This represents the fitness function value. This represents the total number of modal functions. Indicates the first The correlation coefficient between the composite entropy of each modal function and the composite entropy of the original data. The balancing coefficient is used to adjust the weights of the mean and standard deviation in fitness. This represents the average of all correlation coefficients. When removing the intrinsic mode function (IMF) with the largest total composite entropy from the target IMFs, the total composite entropy of each IMF needs to be calculated first. This total value is obtained by weighted averaging of the composite entropy of each sub-signal segment, with weights allocated according to the importance of the time period. After identifying the IMF with the largest total composite entropy, this mode usually represents noise interference or abnormal fluctuation components; removing it helps improve signal quality. When superimposing all remaining IMFs to generate an effective distillation signal, a linear superposition method is used to reconstruct the signal, preserving the main process characteristics while removing noise interference. The reconstructed signal will be used for subsequent identification and optimization of abnormal control parameters.

[0028] Taking a typical baijiu distillation process as an example, six intrinsic mode functions (EMFs) were obtained through variational mode decomposition, representing signal components in different frequency ranges. The implementation process first divides these six EMFs and the original multi-dimensional time-series data into 10 sub-signals, each containing the same number of data points. The division takes into account the duration of the distillation process, ensuring that each sub-signal fully reflects the process characteristics of a specific time period. When calculating the envelope entropy and sample entropy of each sub-signal in the multi-dimensional time-series data, two entropy values ​​are calculated for each sub-signal. The envelope entropy calculation requires first extracting the signal envelope using Hilbert transform, and then calculating the information entropy value of the envelope, which reflects the uniformity of the signal energy distribution. The sample entropy calculation measures the signal complexity and irregularity by analyzing the probability of pattern repetition in the signal sequence. The two entropy values ​​are then weighted and summed with weights of 0.6 and 0.4 respectively to obtain the composite entropy value for each sub-signal.

[0029] The same calculations were performed on each sub-signal of each intrinsic mode function, resulting in 10 composite entropy values ​​for each mode function. These values ​​reflect the complexity and energy characteristics of the mode over different time periods. Analysis of the distribution pattern of these composite entropy values ​​reveals that high-frequency modes typically exhibit higher composite entropy values, while low-frequency modes show relatively lower values. The Pearson correlation coefficient method was used to calculate the correlation coefficient between the composite entropy of the time-series data and the composite entropy of each mode function. The composite entropy of each sub-signal of the original data was used as a baseline sequence, and correlation coefficients were calculated between this sequence and the composite entropy sequence of each sub-signal of each mode function. Higher correlation coefficients indicate that the composite entropy trend of the mode function is more consistent with the original data, and thus better represents the essential characteristics of the data. Fitness function values ​​were generated based on these correlation coefficients. These values ​​comprehensively consider the consistency between each mode function and the original data, as well as the consistency differences between the modes. Higher fitness function values ​​indicate that the decomposition results better preserve the key features of the original data.

[0030] When identifying the mode function with the largest total composite entropy from six intrinsic mode functions (EMFs), it is necessary to calculate the average of the composite entropy of all segments of the signal for each mode function. In this example, the total composite entropy of the fourth mode function is significantly higher than that of the other modes, indicating that this mode contains more noise or abnormal fluctuation components. After removing the fourth mode function, the first three mode functions are superimposed with the last two mode functions for reconstruction. The first three mode functions mainly contain low-frequency process trend information, while the last two mode functions contain some mid-frequency process dynamic characteristics. These components together constitute an effective signal reflecting the essential characteristics of the distillation process. The reconstructed effective distillation signal eliminates the influence of high-frequency noise and abnormal fluctuations, while retaining the main dynamic characteristics of the process. The entire implementation process, through composite entropy analysis and correlation evaluation, ensures the reliability of signal decomposition and reconstruction, providing effective support for the optimized control of the distillation process.

[0031] Example 4: In the implementation of abnormal control parameter identification, setting a dynamic threshold coefficient is an important initial step. The dynamic threshold coefficient is not a fixed value but is adaptively adjusted based on the overall fluctuation characteristics of the signal. When the overall fluctuation of the effective distillation signal is large, the coefficient is appropriately increased to avoid misjudging normal fluctuations as abnormalities. Conversely, when the signal is relatively stable, the coefficient is decreased to improve detection sensitivity. This dynamic adjustment mechanism is achieved by analyzing the historical statistical characteristics of the signal, allowing the threshold to adapt to different production process states. When calculating the mean and standard deviation based on the effective distillation signal, the mean reflects the central trend of the data, and the standard deviation measures the dispersion of the data. A sliding window method is used to process time-series data during calculation. The window size is set according to the sampling frequency and usually includes enough data points to ensure statistical reliability. The calculated mean and standard deviation serve as the basis for subsequent boundary determination. The effective data interval is generated by multiplying the mean and standard deviation by the dynamic threshold coefficient. The upper boundary is the product of the mean, the dynamic threshold coefficient and the standard deviation, and the lower boundary is the product of the mean and the dynamic threshold coefficient and the standard deviation. Data points falling within this interval are considered normal values, otherwise they are marked as abnormal. During the boundary calculation process, the dynamic threshold coefficient is dynamically adjusted according to the signal characteristics to ensure that the interval range can reasonably reflect the normal fluctuation range of the data. See Table 1.

[0032] Table 1: Valid Distillation Signal Data and Anomaly Identification When traversing the discrete data in the valid distillation signal and marking data that does not belong to the valid data range as abnormal control parameters, the system checks each data point one by one whether it is within the valid data range. For the example data shown in the table above, assuming the calculated average value is 10.5, the standard deviation is 2.1, and the dynamic threshold coefficient is set to 2.0, then the upper boundary is 10.5 + 2.0 × 2.1 = 14.7, and the lower boundary is 10.5 - 2.0 × 2.1 = 6.3. The fourth data point value of 15.8 exceeds the upper boundary and is therefore marked as abnormal. Similarly, the ninth data point value of 16.2 also exceeds the upper boundary and is marked as abnormal. When using the arithmetic mean of the preceding and following data points of the outlier control parameters as replacement values, for the fourth outlier data point, the arithmetic mean of its preceding data point 11.1 and its following data point 10.3, (11.1+10.3) / 2=10.7, is used as the replacement value. For the ninth outlier data point, the arithmetic mean of its preceding data point 9.5 and its following data point 9.8, (9.5+9.8) / 2=9.65, is used as the replacement value. This replacement method maintains the continuity and smoothness of the data. When generating the optimized control parameter set by replacing the outlier control parameters with replacement values, the outliers in the original data are replaced with the calculated replacement values, forming a new data sequence. This new sequence eliminates outlier fluctuations, and the optimized control parameter set will be used in the iterative optimization process of the cattle herd optimization algorithm. The entire outlier identification and replacement process effectively removes outlier interference in the data through a dynamic threshold mechanism and a nearest neighbor interpolation method, improving the quality of the control parameters and the subsequent optimization effect.

[0033] Example 5: The process of generating optimal control parameters through iterative optimization using a bovine herd optimization algorithm based on an optimized control parameter set begins by using the optimized control parameter set as the initial population for the algorithm. This optimized control parameter set contains sequences of pressure, temperature, and flow parameters after anomaly handling. Each parameter sequence is encoded as an individual in the population, with a population size set to 50 individuals to balance search efficiency and computational complexity. Each individual represents a complete set of control parameter configurations, including variables such as pressure setpoint, temperature curve parameters, and flow regulation coefficients. When the fitness function is the weighted sum of distillation efficiency and energy consumption ratio, distillation efficiency is calculated by measuring the output of qualified distillate per unit time, and energy consumption ratio is determined by monitoring the ratio of steam consumption to output. The weighting coefficients are configured according to the economic and quality requirements of the production process, typically with efficiency weighting slightly higher than energy consumption weighting to prioritize product quality. The value of the fitness function directly reflects the comprehensive performance of the control parameter combination.

[0034] When iteratively updating the population position using the cattle herd optimization algorithm, the algorithm simulates the collective movement pattern in cattle foraging behavior. Each individual adjusts its movement direction based on its own historical best position and the group's global best position. The position update formula includes inertial components, individual cognitive components, and social guidance components. During the iteration process, a random perturbation factor is introduced to avoid premature convergence. After each iteration, the fitness values ​​of all individuals are recalculated and the historical best records are updated. Iteration stops when the fitness function value converges. The convergence condition is set to the improvement of the fitness function value for several consecutive generations being less than a predetermined threshold, indicating that the algorithm has found an approximate optimal solution. At this point, the search process is terminated to prevent over-optimization.

[0035] The population optimum solution at convergence is used as the optimal control parameter. This optimum solution represents the best-performing parameter combination within the current search range, including key operational variables such as the pressure control setpoint, temperature regulation curve, and flow control parameters. These parameters have been thoroughly explored and verified by the optimization algorithm. When the pressure, temperature, and flow parameters in the optimal control parameters are mapped to the pressure controller, temperature sensor, and flow control valve of the distillation column, respectively, the pressure parameter is converted into the setpoint signal of the pressure controller, the temperature parameter is input into the setpoint curve of the temperature controller, and the flow parameter is converted into the opening command of the control valve, thus achieving precise issuance of control commands.

[0036] Based on the mapping results, the operating parameters of the distillation equipment are adjusted in real time. The pressure controller adjusts the opening of the inlet valve according to the received setpoint, the temperature controller adjusts the heating power output, and the flow regulating valve changes its opening size according to the command. The entire control loop forms a closed-loop regulation, stabilizing the distillation process in an optimized state. During implementation, key process parameters and product quality indicators are continuously monitored to ensure the continuity and stability of the optimized control effect, ultimately achieving intelligent optimized operation of the distillation process.

[0037] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An intelligent distillation optimization method based on artificial intelligence, characterized in that, include: Obtain spectral data of the mash raw material and the initial operating parameters of the distillation equipment; Based on the spectral data, a convolutional neural network is used to generate the characteristic distribution state of the fermentation mash raw materials; Distillation process monitoring is triggered based on the comparison between the characteristic distribution state and the preset threshold range; When the distillation process monitoring is triggered, data on distillation column pressure fluctuations, distillate temperature gradients, and alcohol vapor concentrations are collected simultaneously. The distillation column pressure fluctuation data, distillation temperature gradient data, and alcohol vapor component concentration data are subjected to timestamp alignment processing to generate multi-dimensional time-series data. Based on the multi-dimensional time-series data, a target number of intrinsic mode functions are generated through variational mode decomposition. The effective distillation signal is obtained by removing the intrinsic mode function with the largest total composite entropy from the target intrinsic mode functions; Based on the effective distillation signal, abnormal control parameters are identified by comparing discrete data, and the abnormal control parameters are eliminated and replaced to obtain an optimized control parameter set. Based on the optimized control parameter set, the optimal control parameters are generated iteratively using the cattle herd optimization algorithm. Adjust the operating parameters of the distillation equipment according to the optimal control parameters.

2. The intelligent distillation optimization method based on artificial intelligence according to claim 1, characterized in that, The process of generating the characteristic distribution state of the fermentation mash raw materials based on the spectral data using a convolutional neural network includes: Set the initial kernel size and initial number of pooling layers for the convolutional neural network; Based on the spectral data, frequency domain feature vectors are extracted, and the frequency domain feature vectors are convolved with an initial convolution kernel to generate a primary feature map. Max pooling is performed on the primary feature map to generate a compressed feature map; Based on the compressed feature map, a feature distribution state vector is generated through a fully connected layer; Based on the matching degree between the feature distribution state vector and the historical feature library, the initial convolution kernel size and the initial number of pooling layers are corrected to generate the target convolution kernel size and the target number of pooling layers; The target convolution kernel size and target pooling layer number are used to perform secondary feature extraction on the spectral data to generate the characteristic distribution state of the mash raw materials.

3. The intelligent distillation optimization method based on artificial intelligence according to claim 1, characterized in that, The process of performing timestamp alignment on the distillation column pressure fluctuation data, fraction temperature gradient data, and alcohol vapor concentration data includes: Acquire the first sampling time series of the distillation column pressure fluctuation data, the second sampling time series of the distillation temperature gradient data, and the third sampling time series of the alcohol vapor component concentration data; Using the first sampling time series as a reference, the second and third sampling time series are aligned to the first sampling time series using linear interpolation. Generate multi-dimensional time series data with a unified timestamp based on the aligned time series.

4. The intelligent distillation optimization method based on artificial intelligence according to claim 1, characterized in that, The generation of target intrinsic mode functions based on the multi-dimensional time-series data through variational mode decomposition includes: Set the initial number of decomposition layers and the initial penalty factor for variational mode decomposition; The fitness function value is generated based on the multi-dimensional time series data and the initial intrinsic mode functions. Based on the fitness function value, the initial decomposition layer number and initial penalty factor are optimized using the particle swarm optimization algorithm to generate the target decomposition layer number and target penalty factor; Variational mode decomposition is performed on the multi-dimensional time series data using the target decomposition level and the target penalty factor to generate target intrinsic mode functions.

5. The intelligent distillation optimization method based on artificial intelligence according to claim 4, characterized in that, The process of generating fitness function values ​​based on the multi-dimensional time-series data and the initial intrinsic mode functions includes: The multi-dimensional time series data and each eigenmode function of the initial eigenmode function are divided into N sub-signals; Calculate the weighted sum of the envelope entropy and sample entropy of each segment of the multi-dimensional time series data to generate the composite entropy of the time series data; Calculate the weighted sum of the envelope entropy and sample entropy of each sub-signal segment of each intrinsic mode function to generate the composite entropy of the mode function; The fitness function value is generated based on the correlation coefficient between the composite entropy of the time series data and the composite entropy of the modal function.

6. The intelligent distillation optimization method based on artificial intelligence according to claim 1, characterized in that, The step of removing the intrinsic mode function with the largest total composite entropy from the target intrinsic mode functions includes: Calculate the total composite entropy of each eigenmode function in the target intrinsic mode functions; Identify the eigenmode function with the largest total composite entropy; The effective distillation signal is generated by superimposing all eigenmode functions before the eigenmode function with the largest total composite entropy value and all eigenmode functions after it.

7. The intelligent distillation optimization method based on artificial intelligence according to claim 1, characterized in that, The step of identifying abnormal control parameters by comparing discrete data based on the effective distillation signal includes: Set a dynamic threshold coefficient, and calculate the average value and standard deviation based on the effective distillation signal; The effective data interval is generated by using the product of the average value and the standard deviation multiplied by the dynamic threshold coefficient as the boundary; Traverse the discrete data in the valid distillation signal and mark the discrete data that does not belong to the valid data range as abnormal control parameters.

8. The intelligent distillation optimization method based on artificial intelligence according to claim 7, characterized in that, The removal and replacement of the abnormal control parameters includes: The arithmetic mean of the previous and next data points of the abnormal control parameter is used as the replacement value; Replace the abnormal control parameters with the alternative values ​​to generate an optimized control parameter set.

9. The intelligent distillation optimization method based on artificial intelligence according to claim 1, characterized in that, The process of generating optimal control parameters through iterative optimization using the cattle herd optimization algorithm based on the optimized control parameter set includes: The optimized control parameter set is used as the initial population for the cattle herd optimization algorithm; The fitness function is the weighted sum of distillation efficiency and energy consumption ratio. The population position is iteratively updated using the cattle herd optimization algorithm until the fitness function value converges. The population-optimal solution at convergence is used as the optimal control parameter. The step of adjusting the operating parameters of the distillation equipment according to the optimal control parameters includes: The pressure, temperature, and flow parameters in the optimal control parameters are respectively mapped to the pressure controller, temperature sensor, and flow regulating valve of the distillation column; The operating parameters of the distillation equipment are adjusted in real time based on the mapping results.

10. An intelligent distillation optimization system based on artificial intelligence, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the artificial intelligence-based intelligent distillation optimization method according to any one of claims 1 to 9.

Images