Probiotic packaging stability control method based on deep learning

Abstract

The invention discloses a probiotic packaging stability control method based on deep learning, and the method comprises the following steps: collecting multi-source data in a packaging process, and carrying out the preprocessing of the multi-source data; performing time sequence feature analysis on the standardized structured data set, and extracting multi-scale time sequence features; performing target segmentation and multi-dimensional structure feature extraction on the standardized microscopic image data set; feature fusion is carried out, a probiotic packaging stability feature space is constructed, and a stable feature vector is generated; inputting a time sequence prediction network, and carrying out time sequence modeling and prediction processing; performing control parameter adjustment and constraint optimization based on a difference value between a prediction result and stability reference data; and executing and updating the packaging control adjustment scheme according to feedback to complete a packaging control closed loop. According to the method, deep learning and multi-source data analysis are fused, intelligent prediction and adaptive control of the probiotic packaging process are realized, and the method has the advantages of high stability and high precision.

Description

Technical Field

[0001] This invention relates to the fields of bioengineering and intelligent manufacturing technology, and in particular to a method for controlling the stability of probiotic encapsulation based on deep learning. Background Technology

[0002] Currently, probiotic encapsulation technology mainly employs physicochemical methods such as spray drying, emulsification curing, or co-precipitation to improve storage resistance and activity retention by forming an encapsulation wall structure outside the bacteria. Existing encapsulation processes largely rely on fixed process parameters and empirical control, lacking real-time perception and modeling analysis of the dynamic characteristics and microstructural changes during the encapsulation process. This makes it difficult to accurately reflect the comprehensive impact of process disturbances, material properties, and bacterial physiological characteristics on encapsulation stability, resulting in significant fluctuations in encapsulation structure integrity and viability.

[0003] While existing technologies include process optimization methods based on statistics or single neural networks, most only perform single-dimensional modeling of process parameters or quality results, failing to achieve fusion analysis of microscopic structure data and time-series process data, and lacking dynamic constraint mechanisms oriented towards biological mechanisms. Furthermore, traditional control systems have not established a closed-loop control structure of prediction-decision-execution-feedback, making it impossible to adaptively adjust key parameters such as drying temperature, airflow rate, and encapsulating agent ratio, thus failing to meet the high requirements for stability and consistency in the probiotic encapsulation process.

[0004] Therefore, how to provide a deep learning-based method for controlling the stability of probiotic encapsulation is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a probiotic encapsulation stability control method based on deep learning. This invention integrates deep learning and multi-source data analysis to achieve intelligent prediction and adaptive control of the probiotic encapsulation process, and has the advantages of high stability and high precision.

[0006] A method for controlling the stability of probiotic encapsulation based on deep learning according to an embodiment of the present invention includes the following steps:

[0007] Multi-source data were collected during the packaging process, preprocessed separately, and standardized structured datasets and standardized microscopic image datasets were generated.

[0008] Temporal feature analysis is performed on standardized structured datasets to extract multi-scale temporal features that reflect the dynamic changes in the encapsulation process;

[0009] Target segmentation and multidimensional structural feature extraction are performed on a standardized microscopic image dataset to generate encapsulation morphology features that characterize the morphology and structural properties of probiotic encapsulation particles;

[0010] By fusing encapsulation morphology features with multi-scale temporal features, a probiotic encapsulation stability feature space is constructed, generating a stable feature vector.

[0011] The stable feature vector is input into the time series prediction network for time series modeling and prediction processing, and the prediction results of probiotic encapsulation stability are output.

[0012] Based on the difference between the prediction results and the stability reference data, control parameters are adjusted and constraints are optimized to generate an encapsulated control adjustment scheme.

[0013] The packaging control adjustment plan is sent to the packaging equipment for execution control, the operating status and control parameter response of the packaging process are monitored, and the packaging control adjustment plan is updated based on the execution feedback to complete the packaging control closed loop.

[0014] Optionally, the multi-source data includes structured data and microscopic image data. The structured data includes process parameters, material parameters, strain characteristic parameters, and process output data. The microscopic structure data includes particle size distribution, porosity, surface morphology, and embedding integrity. The preprocessing includes performing time alignment, missing value filling, normalization, and feature encoding operations on the structured data, and performing size unification, noise reduction, contrast adjustment, and enhancement processing on the microscopic image data.

[0015] Optionally, the extraction of the multi-scale temporal features specifically includes:

[0016] The standardized structured dataset is segmented into time series data, and the packaging process is divided into continuous time segments according to the sampling time. Each time segment corresponds to a process stage in the packaging process, and a stage data matrix is ​​established.

[0017] A time-series sliding window operation is performed on the phased data matrix to calculate the changing trend and fluctuation amplitude of process parameters within each time segment. The process parameters in adjacent windows are weighted and summed, with the weighting factor increasing sequentially with the time segment to obtain the dynamic feature sequence of the packaging process.

[0018] Normalization and scaling transformation are performed on the dynamic feature sequence to transform process parameters of different dimensions into dimensionless feature values ​​under a unified scale, thereby generating a primary timing feature set for the packaging process.

[0019] Multi-scale feature extraction is performed on the primary temporal feature set. By dividing the time window into layers, the encapsulation process response features under long and short time scales are calculated to obtain multi-scale temporal response features.

[0020] The multi-scale time-series response features are subjected to correlation aggregation processing to calculate the time-series correlation coefficient and interaction strength between various process parameters. Based on the packaging process stage identifier, weighted integration is performed to extract multi-scale time-series feature data that reflects the dynamic change law of the packaging process.

[0021] Optionally, the generation of the packaging morphology features specifically includes:

[0022] Target segmentation was performed on the standardized microscopic image dataset to obtain binary masks for the encapsulated particle region, the embedded wall material region, and the pore region. Adaptive thresholding and morphological operations were used to refine the boundaries and remove small noise points to obtain the region annotation results.

[0023] Extract particle size distribution features from the regional labeling results, calculate the equivalent circle diameter for each particle, and statistically analyze the fixed bin particle size histogram and percentile index to obtain particle size distribution description data.

[0024] In the region annotation results, the embedded wall material region and the pore region are selected as regions of interest. The pixel data in the region of interest are calculated, and the ratio of the pore pixel area to the total pixel area of ​​the corresponding region is used as the porosity value. Based on the distance transformation and skeleton analysis method, the wall material thickness distribution data is extracted, and the average value, extreme value and dispersion index of the wall material thickness are calculated to obtain the porosity and thickness description data.

[0025] Surface morphology and texture features are extracted from the surface portion of the encapsulated particle region in the region annotation results. Gray-level co-occurrence statistics are calculated for the gray-level image. Local binary mode histogram, edge pixel density and main direction ratio are calculated for the same region to obtain surface texture description data.

[0026] In the regional labeling results, identify the regions of embedded bacteria and the corresponding embedded wall material regions. Determine the degree of circumferential continuous coverage for each region of embedded bacteria. Use the circumferential continuous ratio as the integrity index. Calculate the proportion of bacteria that are completely covered and partially exposed to obtain the embedding integrity description data.

[0027] The particle size distribution description data, porosity and thickness description data, surface texture description data, and encapsulation integrity description data are normalized and connected to generate encapsulation morphology features that characterize the morphology and structural properties of probiotic encapsulated particles.

[0028] Optionally, the generation of the stable feature vector includes: performing index alignment and dimension matching processing on multi-scale temporal features and packaging morphology features; using a feature projection method to map different features to a unified dimensional space; calculating the correlation matrix between corresponding features; determining the set of fusion weight coefficients based on the correlation matrix; performing weighted fusion on multi-scale temporal features and packaging morphology features and performing decorrelation constraint and noise suppression operations to obtain a fused high-dimensional feature representation; performing feature connection and dimensionality reduction processing on the high-dimensional feature representation; mapping the fused features to a low-dimensional embedding space; and generating a stable feature vector in the low-dimensional embedding space by representing the principal components that characterize the differences in packaging structure and the correlation of process features.

[0029] Optionally, the output of the prediction result specifically includes:

[0030] Receive stable feature vectors, perform category encoding on the stage identifiers of the encapsulation process and generate stage embedding vectors, embed and map the encapsulation morphology features and align and concatenate them with the stage embedding vectors to obtain the input sequence arranged by time index;

[0031] The input sequence is length aligned by padding the feature sequences at different time steps to a uniform length, and a mask matrix is ​​generated for the missing positions and then batch assembled to obtain a standardized input sequence.

[0032] The standardized input sequence is input into the time series prediction network. The input sequence is decomposed in the time series decomposition module to obtain trend components, periodic components and residual components. The time series prediction network adopts the ETSformer network structure, including a time series decomposition module, an encoding and decoding module and a prediction branch module.

[0033] A bioactivity decay constraint is applied to the trend component, the constraint including: setting the value of the trend component to not increase at the time index after the drying and curing stage, the change in adjacent time indices not exceeding the upper limit determined by the bacterial characteristic parameters and environmental parameters, and the half-life within the preset stage is configured by a combination of oxygen resistance and water activity threshold, drying temperature and airflow rate.

[0034] The constrained trend component, the decomposed periodic component, and the residual component are input into the encoding and decoding module of the time series prediction network. Time step feature encoding, context relationship modeling, and multi-head attention weighting are performed on each component to obtain component-level context feature representations.

[0035] The component-level contextual feature representation is input into the prediction branch module, feature alignment and connection are performed on each component, fusion weights are calculated based on component correlation and weighted fusion is performed to generate intermediate prediction results of the target sequence.

[0036] The intermediate prediction results are scaled and unit mapped, and inverse normalized based on the statistics from the training phase to obtain the predicted data of viable cell rate change and encapsulation structure integrity index at each time index.

[0037] The predicted data at each time index are aggregated over time and the final value is selected to output the prediction results of probiotic encapsulation stability.

[0038] Optionally, the application of the bioactivity decay constraint specifically includes:

[0039] Receive the trend component sequence and the packaging process stage identifier, divide the trend component sequence into time indexes according to the stage identifiers, extract the time indexes of the drying and curing stage and thereafter, and obtain the time index set.

[0040] Apply monotonicity constraints to the trend component sequences in the time index set, perform monotonicity constraint processing, and limit the trend value at the current time index to no greater than the trend value at the previous time index;

[0041] The oxygen tolerance parameters, water activity threshold, drying temperature, airflow rate, and ambient humidity of the strain were dimensionless, and the upper limit of the trend change was calculated based on the processed data.

[0042] ;

[0043] in, , These are the weighting coefficients, This is a parameter for the oxygen tolerance of the strain. The water activity threshold. The drying temperature, For airflow rate, Ambient humidity;

[0044] Based on the upper limit of trend change amplitude, the absolute value of the trend difference between adjacent time indices is calculated, the resulting change is compared with the upper limit of trend change amplitude, and the change exceeding the upper limit is subjected to amplitude clipping to limit the range of change of trend components between adjacent time indices, thus obtaining trend component data with amplitude constraints.

[0045] Based on the amplitude-constrained trend component data, calculate the half-life of the trend component:

[0046] ;

[0047] in, The half-life of the trend component is k, k1 is the proportionality constant, and the exponent parameter is k. R is an empirical coefficient. o This is a parameter for the oxygen tolerance of the strain. The water activity threshold. The drying temperature, The airflow rate;

[0048] Based on the half-life, an exponential decay mapping is performed on the trend component, and the updated trend component satisfies:

[0049] ;

[0050] in, These are the values ​​of the trend components after exponential decay mapping. This is the initial trend value for the drying and curing stage. Indexed by the current time. The half-life of the trend component is given by denoted as 'e', ​​where 'e' is the base of the natural exponent.

[0051] The trend components, after being processed by monotonicity constraints, variation range limits, and exponential decay mapping, are aligned according to the time index order, and the constrained trend components are output.

[0052] Optionally, the generation of the packaging control adjustment scheme specifically includes:

[0053] Obtain the prediction results, read the corresponding stability reference data, calculate the viability deviation and structural integrity deviation respectively, and sum the deviations according to the set weight coefficients to obtain the comprehensive deviation.

[0054] Based on the stage identifiers of the packaging process, the comprehensive deviation is mapped to the staged control target deviation, and the allowable value range, adjustment step and execution cycle of the packaging control adjustment data are established according to the mapping results.

[0055] Based on the control law that combines proportional and integral elements, the initial adjustment amount is calculated for the encapsulated control adjustment data respectively.

[0056] Parameter coupling compensation is performed on the initial adjustment amount, and cross correction is performed using the coefficients of the coupling compensation matrix;

[0057] Constraint processing is applied to the adjustment amount after coupling compensation, including upper and lower limit clipping, single adjustment amplitude limit and adjacent period change rate limit, and smoothing filtering is applied to the adjustment sequence.

[0058] The target setpoint, adjustment step, and execution cycle of the constrained encapsulated control adjustment data are encapsulated to generate a control adjustment scheme.

[0059] The beneficial effects of this invention are:

[0060] This invention proposes a deep learning-based method for controlling the stability of probiotic encapsulation. By fusing microscopic structural data and time-series process data, it achieves intelligent modeling and dynamic control of the entire probiotic encapsulation process. This method utilizes multi-source data feature extraction and time-series analysis techniques to capture the dynamic changes in the encapsulation process from multi-dimensional data such as process parameters, material properties, and strain characteristics. Simultaneously, by combining microscopic image segmentation and morphological feature extraction, it achieves precise quantification of structural characteristics such as particle size distribution, porosity, surface texture, and encapsulation integrity, providing fine-grained feature support for subsequent stability prediction. By establishing a feature system that links process and structure bidirectionally, the model can comprehensively reflect the true evolution of the encapsulation process.

[0061] This invention, based on feature modeling, introduces an improved time-series prediction network to jointly predict changes in probiotic activity and the integrity of the packaging structure during the packaging process. This network employs time-series decomposition and bioactivity decay constraints, embedding process stage information and morphological feature vectors into the model input, thereby achieving high-precision modeling of the packaging stability evolution trend. Compared to traditional empirical formulas or single statistical models, this invention can dynamically identify trend changes, periodic fluctuations, and random disturbances at different process stages, exhibiting stronger temporal adaptability and physical interpretability. Through this prediction model, feedforward prediction and real-time adjustment of packaging quality can be achieved, improving the controllability of the packaging process and product consistency.

[0062] Furthermore, this invention establishes a packaging control adjustment mechanism based on prediction results. Combining proportional-integral control laws and parameter coupling compensation methods, it adaptively adjusts key process parameters such as drying temperature, airflow rate, encapsulating agent ratio, and mixing intensity. Through dynamic weight allocation and constraint optimization of deviations, it ensures smooth convergence of parameter adjustments and conforms to actual equipment operating conditions, forming a closed-loop control system of prediction-decision-execution-feedback. This method significantly improves the problems of adjustment lag, large parameter fluctuations, and insufficient packaging stability in traditional packaging processes. It achieves high-precision intelligent control of the probiotic packaging process, improves the viable bacteria retention rate and packaging structure integrity, and provides technical support for the large-scale and stable production of probiotic products. Attached Figure Description

[0063] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0064] Figure 1 This is a flowchart of a probiotic encapsulation stability control method based on deep learning proposed in this invention. Detailed Implementation

[0065] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0066] refer to Figure 1 A deep learning-based method for controlling the stability of probiotic encapsulation includes the following steps:

[0067] Multi-source data were collected during the packaging process, preprocessed separately, and standardized structured datasets and standardized microscopic image datasets were generated.

[0068] Temporal feature analysis is performed on standardized structured datasets to extract multi-scale temporal features that reflect the dynamic changes in the encapsulation process;

[0069] Target segmentation and multidimensional structural feature extraction are performed on a standardized microscopic image dataset to generate encapsulation morphology features that characterize the morphology and structural properties of probiotic encapsulation particles;

[0070] By fusing encapsulation morphology features with multi-scale temporal features, a probiotic encapsulation stability feature space is constructed, generating a stable feature vector.

[0071] The stable feature vector is input into the time series prediction network for time series modeling and prediction processing, and the prediction results of probiotic encapsulation stability are output.

[0072] Based on the difference between the prediction results and the stability reference data, control parameters are adjusted and constraints are optimized to generate an encapsulated control adjustment scheme.

[0073] The packaging control adjustment plan is sent to the packaging equipment for execution control, the operating status and control parameter response of the packaging process are monitored, and the packaging control adjustment plan is updated based on the execution feedback to complete the packaging control closed loop.

[0074] In this embodiment, the multi-source data includes structured data and microscopic image data. The structured data includes process parameters, material parameters, strain characteristic parameters, and process output data. The process parameters include temperature, humidity, stirring rate, drying temperature, airflow rate, pressure, and feed / discharge rate. The material parameters include encapsulating agent type, wall material to core material mass ratio, system viscosity, and solution solid content. The strain characteristic parameters include strain number, initial viable cell count, oxygen tolerance, and water activity threshold. The process output data includes encapsulation yield, discharge moisture content, and storage activity retention rate. The microscopic structure data includes particle size distribution, porosity, surface morphology, and encapsulation integrity. The preprocessing includes performing time alignment, missing value filling, normalization, and feature encoding operations on the structured data, and performing size unification, noise reduction, contrast adjustment, and enhancement processing on the microscopic image data.

[0075] In this embodiment, the extraction of multi-scale temporal features specifically includes:

[0076] The standardized structured dataset is processed by time series segmentation. The encapsulation process is divided into continuous time segments according to the sampling time. Each time segment corresponds to a process stage in the encapsulation process. A stage data matrix is ​​established. The encapsulation process includes the raw material mixing stage, the encapsulation formation stage, the drying and curing stage, the cooling and stabilization stage, and the material discharge and storage stage.

[0077] A time-series sliding window operation is performed on the phased data matrix to calculate the changing trend and fluctuation amplitude of process parameters within each time segment. The process parameters in adjacent windows are weighted and summed, with the weighting factor increasing sequentially with the time segment to obtain the dynamic feature sequence of the packaging process.

[0078] Normalization and scaling transformation are performed on the dynamic feature sequence to transform process parameters of different dimensions into dimensionless feature values ​​under a unified scale, thereby generating a primary timing feature set for the packaging process.

[0079] Multi-scale feature extraction is performed on the primary time-series feature set. By dividing the time window into layers, the packaging process response features under long time scale and short time scale are calculated respectively to obtain multi-scale time-series response features. The long time scale reflects the overall process trend, and the short time scale reflects the instantaneous disturbance change.

[0080] The multi-scale time-series response features are subjected to correlation aggregation processing to calculate the time-series correlation coefficient and interaction strength between various process parameters. Based on the packaging process stage identifier, weighted integration is performed to extract multi-scale time-series feature data that reflects the dynamic change law of the packaging process.

[0081] In this embodiment, the generation of the packaging morphology features specifically includes:

[0082] Target segmentation was performed on the standardized microscopic image dataset to obtain binary masks for the encapsulated particle region, the embedded wall material region, and the pore region. Adaptive thresholding and morphological operations were used to refine the boundaries and remove small noise points to obtain the region annotation results.

[0083] The region labeling results are obtained by: performing pixel-level processing on each image of the standardized microscopic image dataset, calculating a local threshold map based on the brightness gradient and grayscale distribution, using an adaptive thresholding method to distinguish between the encapsulated particle region and the background region, generating an initial binary image, performing structural correction on the initial binary image using morphological dilation and erosion operations, removing isolated noise points and smoothing particle edges, calculating connected regions on the corrected image, filtering encapsulated particle regions based on area and shape features, further identifying embedded wall material regions and pore regions through edge detection and connected region analysis, and finally labeling the encapsulated particle region, embedded wall material region, and pore region as different pixel categories, outputting the corresponding binary mask map, and synthesizing the mask set into the region labeling results;

[0084] Extract particle size distribution features from the regional labeling results, calculate the equivalent circle diameter for each particle, and statistically analyze the fixed bin particle size histogram and percentile index to obtain particle size distribution description data.

[0085] The particle size distribution description data is obtained by: performing pixel area calculation for each packaged particle region in the region labeling results, recording the total number of pixels of the particles, converting the pixel area into the actual area, calculating the equivalent circle diameter based on the area value, summarizing the equivalent circle diameters of all particles, performing interval statistics on the diameter data according to the preset particle size binning interval, generating a particle size frequency histogram, further calculating the statistical characteristic values ​​of the particle size distribution to reflect the distribution characteristics of particle size in the overall sample, and combining the particle size frequency histogram with the statistical characteristic values ​​to obtain the particle size distribution description data. The statistical characteristic values ​​include the average particle size, median particle size, standard deviation, and percentile index. The equivalent circle diameter is defined as the diameter of a circle with the same area as the corresponding particle.

[0086] In the region annotation results, the embedded wall material region and the pore region are selected as regions of interest. The pixel data in the region of interest are calculated, and the ratio of the pore pixel area to the total pixel area of ​​the corresponding region is used as the porosity value. Based on the distance transformation and skeleton analysis method, the wall material thickness distribution data is extracted, and the average value, extreme value and dispersion index of the wall material thickness are calculated to obtain the porosity and thickness description data.

[0087] Surface morphology and texture features are extracted from the surface portion of the encapsulated particle region in the region annotation results. Gray-level co-occurrence statistics are calculated for the gray-level image. Local binary mode histogram, edge pixel density and main direction ratio are calculated for the same region to obtain surface texture description data. The gray-level co-occurrence statistics include contrast, energy, homogeneity and correlation statistics.

[0088] In the regional labeling results, identify the regions of embedded bacteria and the corresponding embedded wall material regions. Determine the degree of circumferential continuous coverage for each region of embedded bacteria. Use the circumferential continuous ratio as the integrity index. Calculate the proportion of bacteria that are completely covered and partially exposed to obtain the embedding integrity description data.

[0089] The particle size distribution description data, porosity and thickness description data, surface texture description data, and encapsulation integrity description data are normalized and connected to generate encapsulation morphology features that characterize the morphology and structural properties of probiotic encapsulated particles.

[0090] In this embodiment, the generation of the stable feature vector includes: performing index alignment and dimension matching processing on multi-scale temporal features and packaging morphology features; using a feature projection method to map different features to a unified dimensional space; calculating the correlation matrix between corresponding features; determining the set of fusion weight coefficients based on the correlation matrix; performing weighted fusion on multi-scale temporal features and packaging morphology features and performing decorrelation constraint and noise suppression operations to obtain a fused high-dimensional feature representation; performing feature connection and dimensionality reduction processing on the high-dimensional feature representation; mapping the fused features to a low-dimensional embedding space; and generating a stable feature vector in the low-dimensional embedding space by representing the principal components that characterize the differences in packaging structure and the correlation of process features.

[0091] In this embodiment, the output of the prediction result specifically includes:

[0092] Receive stable feature vectors, perform category encoding on the stage identifiers of the encapsulation process and generate stage embedding vectors, embed and map the encapsulation morphology features and align and concatenate them with the stage embedding vectors to obtain the input sequence arranged by time index;

[0093] The input sequence is length aligned by padding the feature sequences at different time steps to a uniform length, and a mask matrix is ​​generated for the missing positions and then batch assembled to obtain a standardized input sequence.

[0094] The standardized input sequence is input into the time series prediction network. The input sequence is decomposed in the time series decomposition module to obtain trend components, periodic components and residual components. The time series prediction network adopts the ETSformer network structure, including a time series decomposition module, an encoding and decoding module and a prediction branch module.

[0095] A bioactivity decay constraint is applied to the trend component, the constraint including: setting the value of the trend component to not increase at the time index after the drying and curing stage, the change in adjacent time indices not exceeding the upper limit determined by the bacterial characteristic parameters and environmental parameters, and the half-life within the preset stage is configured by a combination of oxygen resistance and water activity threshold, drying temperature and airflow rate.

[0096] The constrained trend component, the decomposed periodic component, and the residual component are input into the encoding and decoding module of the time series prediction network. Time step feature encoding, context relationship modeling, and multi-head attention weighting are performed on each component to obtain component-level context feature representations.

[0097] The component-level context feature representation is input into the prediction branch module, feature alignment and connection are performed on each component, fusion weights are calculated based on component correlation and weighted fusion is performed to generate intermediate prediction results of the target sequence. The target sequence includes a sequence of viable cell rate change values ​​and a sequence of encapsulation structure integrity indicators.

[0098] The intermediate prediction results are scaled and unit mapped, and inverse normalized based on the statistics from the training phase to obtain the predicted data of viable cell rate change and encapsulation structure integrity index at each time index.

[0099] The predicted data at each time index are aggregated over time and the final value is selected to output the prediction results of probiotic encapsulation stability.

[0100] In this embodiment, the application of the bioactivity attenuation constraint specifically includes:

[0101] Receive the trend component sequence and the packaging process stage identifier, divide the trend component sequence into time indexes according to the stage identifiers, extract the time indexes of the drying and curing stage and thereafter, and obtain the time index set.

[0102] Apply monotonicity constraints to the trend component sequences in the time index set, perform monotonicity constraint processing, and limit the trend value at the current time index to no greater than the trend value at the previous time index;

[0103] The oxygen tolerance parameters, water activity threshold, drying temperature, airflow rate, and ambient humidity of the strain were dimensionless, and the upper limit of the trend change was calculated based on the processed data.

[0104] ;

[0105] in, , These are the weighting coefficients, This is a parameter for the oxygen tolerance of the strain. The water activity threshold. The drying temperature, For airflow rate, Ambient humidity;

[0106] Based on the upper limit of trend change amplitude, the absolute value of the trend difference between adjacent time indices is calculated, the resulting change is compared with the upper limit of trend change amplitude, and the change exceeding the upper limit is subjected to amplitude clipping to limit the range of change of trend components between adjacent time indices, thus obtaining trend component data with amplitude constraints.

[0107] Based on the amplitude-constrained trend component data, calculate the half-life of the trend component:

[0108] ;

[0109] in, The half-life of the trend component is k, k1 is the proportionality constant, and the exponent parameter is k. R is an empirical coefficient. o This is a parameter for the oxygen tolerance of the strain. The water activity threshold. The drying temperature, The airflow rate;

[0110] Based on the half-life, an exponential decay mapping is performed on the trend component, and the updated trend component satisfies:

[0111] ;

[0112] in, These are the values ​​of the trend components after exponential decay mapping. This is the initial trend value for the drying and curing stage. Indexed by the current time. The half-life of the trend component is given by denoted as 'e', ​​where 'e' is the base of the natural exponent.

[0113] The trend components, after being processed by monotonicity constraints, variation range limits, and exponential decay mapping, are aligned according to the time index order, and the constrained trend components are output.

[0114] In this embodiment, the generation of the packaging control adjustment scheme specifically includes:

[0115] Obtain the prediction results, read the corresponding stability reference data, calculate the viability deviation and structural integrity deviation respectively, and sum the deviations according to the set weight coefficients to obtain the comprehensive deviation. The stability reference data includes the viability reference value, the packaging structural integrity reference value, the stage tolerance range and the process condition constraint reference for each packaging stage.

[0116] Based on the stage identifiers of the packaging process, the comprehensive deviation is mapped to the stage control target deviation, and the allowable value range, adjustment step and execution cycle of the packaging control adjustment data are established according to the mapping results. The packaging control adjustment data includes drying temperature, airflow rate, encapsulant ratio and mixing intensity.

[0117] According to the control law combining proportional and integral elements, the initial adjustment amount is calculated for the packaging control adjustment data. The coefficients of the proportional element and the integral element are set according to the type of each control parameter, and are used to adjust the response sensitivity and cumulative correction amplitude of the packaging control adjustment data.

[0118] Parameter coupling compensation is performed on the initial adjustment amount, and cross correction is performed using the coefficients of the coupling compensation matrix. Each element of the coupling compensation matrix is ​​the coupling coefficient between different control parameters, which is used to characterize the mutual influence strength between the encapsulated control adjustment data.

[0119] Constraint processing is applied to the adjustment amount after coupling compensation, including upper and lower limit clipping, single adjustment amplitude limit and adjacent period change rate limit, and smoothing filtering is applied to the adjustment sequence. The smoothing filtering adopts the moving average method with a fixed window length.

[0120] The target setpoint, adjustment step, and execution cycle of the constrained encapsulated control adjustment data are encapsulated to generate a control adjustment scheme.

[0121] Example 1:

[0122] To verify the feasibility of this invention in practice, it was applied to the packaging process control of a probiotic powder production line. This production line, located in East China, is mainly used for the spray drying and microcapsule encapsulation of lactic acid bacteria probiotics. Traditional packaging control methods rely on experience to adjust drying temperature, airflow rate, and encapsulating agent ratio, resulting in large fluctuations in packaging stability, low viable cell retention, and lag in process response. This leads to significant batch-to-batch product differences and makes continuous quality optimization difficult. This invention aims to use a deep learning model to model and predict multi-source data in the packaging process, achieving intelligent control and adaptive optimization of probiotic packaging stability, thereby improving the accuracy of the packaging process and product stability.

[0123] In practical applications, multiple types of sensors and microscopic image acquisition devices are first deployed at key stages of the production line to collect data on drying temperature, airflow rate, humidity, stirring speed, system viscosity, discharge moisture content, and microscopic images. The collected structured data undergoes time alignment, missing value imputation, and normalization to form a standardized structured dataset; the microscopic image data is then standardized by size unification, noise reduction, and enhancement to form a standardized microscopic image dataset. Subsequently, the system performs time-series segmentation and multi-scale feature extraction on the standardized structured dataset to extract the dynamic changes in the encapsulation process at different stages. Simultaneously, an improved image segmentation algorithm is used to identify and extract features from the encapsulation particles, wall material, and pore regions in the microscopic images, obtaining multi-dimensional structural features such as particle size distribution, porosity, wall material thickness, and encapsulation integrity.

[0124] The system fuses multi-scale temporal features with microstructural features to generate a stable feature vector that comprehensively reflects changes in the packaging state. This vector is then input into a time series prediction model based on the ETSformer network structure for analysis. The model performs decomposition, encoding, and prediction operations on the input sequence, extracting trend, periodic, and residual components. A bioactivity decay constraint is introduced into the trend component to ensure that the viability rate changes non-increasingly after the drying and curing stage. The model's encoding and decoding modules establish time dependencies, and the system outputs the viability rate changes and packaging structure integrity prediction results at different time indices.

[0125] After obtaining the prediction results, the system calculates the deviation between the prediction results and the packaging stability reference data, and automatically generates a packaging control adjustment scheme through the control parameter adjustment module. The control parameters include drying temperature, airflow rate, encapsulant ratio, and mixing intensity. The system dynamically corrects the control parameters based on the deviation and uses a proportional-integral control law for real-time adjustment. The interaction relationships between the control parameters are corrected by a coupling compensation matrix to avoid overcompensation or conflicts during the parameter adjustment process. The adjusted parameters are then constrained and smoothed before being sent to the packaging equipment for execution, achieving adaptive optimization control.

[0126] After continuous operation for a period of time during the production process, this method significantly improves the operational stability of the packaging equipment. The system can maintain the integrity of the packaging structure and the stability of the viability rate under different batches and environmental conditions. Data comparison shows that after adopting the deep learning control method of this invention, the number of adjustments in the packaging process is reduced, the system response speed is improved, the fluctuation of packaging stability indicators is significantly reduced, and the shelf-life activity retention effect of the packaged products is continuously improved.

[0127] Table 1. Performance Comparison of the Invention and Traditional Probiotic Encapsulation Stability Control Methods

[0128] Indicator Categories Traditional control methods Method of the present invention Average packaging yield (%) 87.2 94.6 Stable deviation of discharge moisture content (%) 6.1 2.3 Viable cell retention rate (%) 72.5 88.9 Package structure integrity index (%) 81.4 93.7 Inter-batch stability fluctuation coefficient (%) 9.6 3.5 Process adjustment response time (min) 18.4 6.9 Activity retention rate during storage (%) 68.3 84.5

[0129] As can be clearly seen from Table 1, the method of the present invention is superior to the traditional method in many indicators.

[0130] In terms of average encapsulation yield, the method of this invention achieves 94.6%, which is 8.48% higher than the 87.2% yield of traditional manual experience control. This improvement is mainly due to the high-dimensional modeling capability of deep learning models for multi-source data in the encapsulation process. Through joint analysis of process parameters and morphological characteristics, more accurate matching of encapsulation ratio and drying conditions is achieved, thereby reducing ineffective energy consumption and encapsulation loss.

[0131] Regarding the stability deviation of the output moisture content, the method of this invention reduces the fluctuation range from 6.1% to 2.3%, a reduction of up to 62.30%. This result demonstrates that the system can monitor the changes in drying temperature, airflow rate, and humidity in real time, and utilizes the time-dependent modeling of the ETSformer prediction network to identify deviation trends in advance. This allows the proportional-integral adjustment module to correct the control signal in advance, significantly improving the stability of the drying process.

[0132] The most significant improvement was in the viable cell retention rate, which increased from 72.5% under traditional control to 88.9%, an increase of 22.55%. This is mainly due to the introduction of a bioactivity decay constraint in this invention, ensuring that the trend component maintains a non-increasing pattern after the drying and curing stage, preventing cell inactivation caused by excessive drying or temperature peaks. Simultaneously, the system dynamically adjusts the drying temperature and airflow rate based on the oxygen tolerance and water activity threshold of the bacterial strain, achieving precise control of environmental adaptability.

[0133] Regarding the integrity of the encapsulation structure, the integrity index of the method of this invention reaches 93.7%, which is 15.13% higher than that of the traditional method (81.4%). When analyzing the encapsulation morphology characteristics, the prediction branch of the ETSformer network achieves high-precision prediction of the changes in the thickness and porosity of the embedded wall material through the fusion of multiple components of trend, period, and residual, enabling better control of material flowability and curing conditions during the encapsulation formation stage.

[0134] The batch-to-batch stability fluctuation coefficient decreased from 9.6% to 3.5%, and the process adjustment response time was shortened from 18.4 minutes to 6.9 minutes, representing improvements of 63.54% and 62.50%, respectively. These indicators collectively reflect a significant enhancement in the system's closed-loop control capability. Because the model can identify abnormal trends early and automatically generate packaging control adjustment schemes, operators no longer need frequent manual intervention, greatly reducing human judgment errors and adjustment delays.

[0135] Finally, in terms of activity retention during storage, the method of this invention achieves 84.5%, which is 23.71% higher than the traditional method. This is because, throughout the entire encapsulation process, this invention not only achieves dynamic optimal control of short-term process parameters, but also suppresses stress concentration and moisture migration in the microstructure through long-term time-series prediction, resulting in a more compact and stable encapsulation structure for the final product.

[0136] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for controlling the stability of probiotic encapsulation based on deep learning, characterized in that, Includes the following steps: Multi-source data were collected during the packaging process, preprocessed separately, and standardized structured datasets and standardized microscopic image datasets were generated. Temporal feature analysis is performed on standardized structured datasets to extract multi-scale temporal features that reflect the dynamic changes in the encapsulation process; Target segmentation and multidimensional structural feature extraction are performed on a standardized microscopic image dataset to generate encapsulation morphology features that characterize the morphology and structural properties of probiotic encapsulation particles; By fusing encapsulation morphology features with multi-scale temporal features, a probiotic encapsulation stability feature space is constructed, generating a stable feature vector. The stable feature vector is input into the time series prediction network for time series modeling and prediction processing, and the prediction results of probiotic encapsulation stability are output. Based on the difference between the prediction results and the stability reference data, control parameters are adjusted and constraints are optimized to generate an encapsulated control adjustment scheme. The packaging control adjustment plan is sent to the packaging equipment for execution control, the operating status and control parameter response of the packaging process are monitored, and the packaging control adjustment plan is updated based on the execution feedback to complete the packaging control closed loop.

2. The method for controlling the stability of probiotic encapsulation based on deep learning according to claim 1, characterized in that, The multi-source data includes structured data and microscopic image data. The structured data includes process parameters, material parameters, strain characteristic parameters, and process output data. The microscopic structure data includes particle size distribution, porosity, surface morphology, and embedding integrity. The preprocessing includes performing time alignment, missing value filling, normalization, and feature encoding operations on the structured data, and performing size unification, noise reduction, contrast adjustment, and enhancement processing on the microscopic image data.

3. The method for controlling the stability of probiotic encapsulation based on deep learning according to claim 1, characterized in that, The extraction of the multi-scale temporal features specifically includes: The standardized structured dataset is segmented into time series data, and the packaging process is divided into continuous time segments according to the sampling time. Each time segment corresponds to a process stage in the packaging process, and a stage data matrix is ​​established. A time-series sliding window operation is performed on the phased data matrix to calculate the changing trend and fluctuation amplitude of process parameters within each time segment. The process parameters in adjacent windows are weighted and summed, with the weighting factor increasing sequentially with the time segment to obtain the dynamic feature sequence of the packaging process. Normalization and scaling transformation are performed on the dynamic feature sequence to transform process parameters of different dimensions into dimensionless feature values ​​under a unified scale, thereby generating a primary timing feature set for the packaging process. Multi-scale feature extraction is performed on the primary temporal feature set. By dividing the time window into layers, the encapsulation process response features under long and short time scales are calculated to obtain multi-scale temporal response features. The multi-scale time-series response features are subjected to correlation aggregation processing to calculate the time-series correlation coefficient and interaction strength between various process parameters. Based on the packaging process stage identifier, weighted integration is performed to extract multi-scale time-series feature data that reflects the dynamic change law of the packaging process.

4. The method for controlling the stability of probiotic encapsulation based on deep learning according to claim 1, characterized in that, The generation of the packaging morphology features specifically includes: Target segmentation was performed on the standardized microscopic image dataset to obtain binary masks for the encapsulated particle region, the embedded wall material region, and the pore region. Adaptive thresholding and morphological operations were used to refine the boundaries and remove small noise points to obtain the region annotation results. Extract particle size distribution features from the regional labeling results, calculate the equivalent circle diameter for each particle, and statistically analyze the fixed bin particle size histogram and percentile index to obtain particle size distribution description data. In the region annotation results, the embedded wall material region and the pore region are selected as regions of interest. The pixel data in the region of interest are calculated, and the ratio of the pore pixel area to the total pixel area of ​​the corresponding region is used as the porosity value. Based on the distance transformation and skeleton analysis method, the wall material thickness distribution data is extracted, and the average value, extreme value and dispersion index of the wall material thickness are calculated to obtain the porosity and thickness description data. Surface morphology and texture features are extracted from the surface portion of the encapsulated particle region in the region annotation results. Gray-level co-occurrence statistics are calculated for the gray-level image. Local binary mode histogram, edge pixel density and main direction ratio are calculated for the same region to obtain surface texture description data. In the regional labeling results, identify the regions of embedded bacteria and the corresponding embedded wall material regions. Determine the degree of circumferential continuous coverage for each region of embedded bacteria. Use the circumferential continuous ratio as the integrity index. Calculate the proportion of bacteria that are completely covered and partially exposed to obtain the embedding integrity description data. The particle size distribution description data, porosity and thickness description data, surface texture description data, and encapsulation integrity description data are normalized and connected to generate encapsulation morphology features that characterize the morphology and structural properties of probiotic encapsulated particles.

5. The method for controlling the stability of probiotic encapsulation based on deep learning according to claim 1, characterized in that, The generation of the stable feature vector includes: performing index alignment and dimension matching processing on multi-scale temporal features and packaging morphology features; using a feature projection method to map different features to a unified dimensional space; calculating the correlation matrix between corresponding features; determining the set of fusion weight coefficients based on the correlation matrix; performing weighted fusion of multi-scale temporal features and packaging morphology features and performing decorrelation constraint and noise suppression operations to obtain a fused high-dimensional feature representation; performing feature connection and dimensionality reduction processing on the high-dimensional feature representation; mapping the fused features to a low-dimensional embedding space; and generating a stable feature vector by representing the principal components that characterize the differences in packaging structure and the correlation of process features in the low-dimensional embedding space.

6. The method for controlling the stability of probiotic encapsulation based on deep learning according to claim 1, characterized in that, The output of the prediction results specifically includes: Receive stable feature vectors, perform category encoding on the stage identifiers of the encapsulation process and generate stage embedding vectors, embed and map the encapsulation morphology features and align and concatenate them with the stage embedding vectors to obtain the input sequence arranged by time index; The input sequence is length aligned by padding the feature sequences at different time steps to a uniform length, and a mask matrix is ​​generated for the missing positions and then batch assembled to obtain a standardized input sequence. The standardized input sequence is input into the time series prediction network. The input sequence is decomposed in the time series decomposition module to obtain trend components, periodic components and residual components. The time series prediction network adopts the ETSformer network structure, including a time series decomposition module, an encoding and decoding module and a prediction branch module. A bioactivity decay constraint is applied to the trend component, the constraint including: setting the value of the trend component to not increase at the time index after the drying and curing stage, the change in adjacent time indices not exceeding the upper limit determined by the bacterial characteristic parameters and environmental parameters, and the half-life within the preset stage is configured by a combination of oxygen resistance and water activity threshold, drying temperature and airflow rate. The constrained trend component, the decomposed periodic component, and the residual component are input into the encoding and decoding module of the time series prediction network. Time step feature encoding, context relationship modeling, and multi-head attention weighting are performed on each component to obtain component-level context feature representations. The component-level contextual feature representation is input into the prediction branch module, feature alignment and connection are performed on each component, fusion weights are calculated based on component correlation and weighted fusion is performed to generate intermediate prediction results of the target sequence. The intermediate prediction results are scaled and unit mapped, and inverse normalized based on the statistics from the training phase to obtain the predicted data of viable cell rate change and encapsulation structure integrity index at each time index. The predicted data at each time index are aggregated over time and the final value is selected to output the prediction results of probiotic encapsulation stability.

7. The method for controlling the stability of probiotic encapsulation based on deep learning according to claim 6, characterized in that, The application of the bioactivity attenuation constraint specifically includes: Receive the trend component sequence and the packaging process stage identifier, divide the trend component sequence into time indexes according to the stage identifiers, extract the time indexes of the drying and curing stage and thereafter, and obtain the time index set. Apply monotonicity constraints to the trend component sequences in the time index set, perform monotonicity constraint processing, and limit the trend value at the current time index to no greater than the trend value at the previous time index; The oxygen tolerance parameters, water activity threshold, drying temperature, airflow rate, and ambient humidity of the strain were dimensionless, and the upper limit of the trend change was calculated based on the processed data. ； in, , These are the weighting coefficients, This is a parameter for the oxygen tolerance of the strain. The water activity threshold. The drying temperature, For airflow rate, Ambient humidity; Based on the upper limit of trend change amplitude, the absolute value of the trend difference between adjacent time indices is calculated, the resulting change is compared with the upper limit of trend change amplitude, and the change exceeding the upper limit is subjected to amplitude clipping to limit the range of change of trend components between adjacent time indices, thus obtaining trend component data with amplitude constraints. Based on the amplitude-constrained trend component data, calculate the half-life of the trend component: ； in, The half-life of the trend component is k, k1 is the proportionality constant, and the exponent parameter is k. R is an empirical coefficient. o This is a parameter for the oxygen tolerance of the strain. The water activity threshold. The drying temperature, The airflow rate; Based on the half-life, an exponential decay mapping is performed on the trend component, and the updated trend component satisfies: ； in, These are the values ​​of the trend components after exponential decay mapping. This is the initial trend value for the drying and curing stage. Indexed by the current time. The half-life of the trend component is given by denoted as 'e', ​​where 'e' is the base of the natural exponent. The trend components, after being processed by monotonicity constraints, variation range limits, and exponential decay mapping, are aligned according to the time index order, and the constrained trend components are output.

8. The method for controlling the stability of probiotic encapsulation based on deep learning according to claim 1, characterized in that, The generation of the packaging control adjustment scheme specifically includes: Obtain the prediction results, read the corresponding stability reference data, calculate the viability deviation and structural integrity deviation respectively, and sum the deviations according to the set weight coefficients to obtain the comprehensive deviation. Based on the stage identifiers of the packaging process, the comprehensive deviation is mapped to the staged control target deviation, and the allowable value range, adjustment step and execution cycle of the packaging control adjustment data are established according to the mapping results. Based on the control law that combines proportional and integral elements, the initial adjustment amount is calculated for the encapsulated control adjustment data respectively. Parameter coupling compensation is performed on the initial adjustment amount, and cross correction is performed using the coefficients of the coupling compensation matrix; Constraint processing is applied to the adjustment amount after coupling compensation, including upper and lower limit clipping, single adjustment amplitude limit and adjacent period change rate limit, and smoothing filtering is applied to the adjustment sequence. The target setpoint, adjustment step, and execution cycle of the constrained encapsulated control adjustment data are encapsulated to generate a control adjustment scheme.

Images