Detection system and method for enzymolysis reaction components of brewing nourishment

Abstract

The invention relates to the technical field of nourishment detection, and discloses a system and a method for detecting enzymolysis reaction components of brewing nourishment. The method comprises the following steps: collecting dynamic data such as an enzyme activity value, a substrate concentration value and a product generation rate value in an enzymolysis reaction process in real time through a multi-source sensor; performing multi-modal feature fusion processing on the dynamic data to generate an enzymolysis reaction feature vector; classifying enzymolysis reaction state categories based on the feature vectors, and dynamically adjusting detection parameter thresholds according to different state categories; and according to the deviation value between the real-time reaction state and the parameter threshold value, correcting an enzymolysis reaction process control instruction by adopting a nonlinear optimization algorithm. According to the method, the dynamic state of the enzymolysis reaction can be captured in real time, accurate detection of enzymolysis reaction components and reasonable regulation and control of the reaction process are achieved through multi-dimensional data fusion and dynamic adjustment, and the method is suitable for efficient detection of the enzymolysis reaction of the brewing nourishment.

Description

Technical Field

[0001] This invention relates to the field of nutritional product testing technology, specifically to a detection system and method for the enzymatic hydrolysis reaction components of reconstituted nutritional products. Background Technology

[0002] In the production of instant nutritional supplements, enzymatic hydrolysis is a crucial step affecting product quality. Changes in components during this reaction directly impact nutrient retention, taste, and stability. Enzymatic hydrolysis involves the interaction between enzymes and substrates; fluctuations in enzyme activity, the rate of substrate consumption, and the dynamic changes in product formation all collectively determine the final product quality. Therefore, real-time and accurate monitoring of enzymatic hydrolysis components is a key focus in the production of instant nutritional supplements. The detection methods for enzymatic hydrolysis components in instant nutritional supplements have many limitations. Traditional detection methods mostly rely on offline sampling analysis, which involves periodically extracting samples from the reaction system and using laboratory instruments to analyze the components. This method cannot reflect the dynamic changes of the reaction in real time. By the time the test results are returned, the enzymatic hydrolysis reaction may have deviated from the ideal process, making timely intervention difficult. Furthermore, offline detection requires significant manpower and time, and the sampling process may disrupt the stability of the reaction system, affecting the continuity of the reaction. Some existing technologies attempt to use a single sensor for online detection, such as monitoring only substrate concentration or a single enzyme activity indicator. However, enzymatic reactions are complex processes involving the interaction of multiple factors, and a single indicator cannot fully reflect the true state of the reaction. For example, when the substrate concentration is within the normal range, abnormal fluctuations in enzyme activity may have already caused the product formation rate to deviate from expectations, and relying solely on a single indicator is prone to detection bias. In existing detection methods, the detection parameter thresholds are mostly fixed values, which cannot adapt to the characteristics of different stages of enzymatic hydrolysis reactions. Enzymatic hydrolysis reactions typically go through an initiation phase, a rapid reaction phase, and a plateau phase, and the reaction kinetic characteristics of each stage are significantly different. Fixed parameter thresholds are difficult to match the changing patterns of different stages, which may lead to misjudgment of abnormalities in the early stages of the reaction or failure to identify potential problems in a timely manner in the later stages of the reaction, thus affecting the accuracy and applicability of the detection. Summary of the Invention

[0003] The purpose of this invention is to provide a detection system and method for the enzymatic hydrolysis reaction components of instant nutritional supplements, so as to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides a method for detecting enzymatic hydrolysis components in instant nutritional supplements, the method comprising: Dynamic data during the enzymatic hydrolysis reaction process are collected in real time using multi-source sensors. The dynamic data includes enzyme activity value, substrate concentration value, and product formation rate value. The dynamic data is subjected to multimodal feature fusion processing to generate an enzymatic reaction feature vector; The enzyme hydrolysis reaction state categories are divided based on the feature vector of the enzyme hydrolysis reaction, and the detection parameter thresholds are dynamically adjusted according to different state categories. Based on the deviation between the real-time enzymatic hydrolysis reaction status and the parameter threshold, a nonlinear optimization algorithm is used to correct the enzymatic hydrolysis reaction process control command.

[0005] Preferably, the multi-source sensor includes an enzyme activity biosensor, a substrate concentration spectrometer, and a product formation rate meter; The dynamic data is standardized and denoised to construct a time-series-based matrix of the coordinated changes in enzyme activity, substrate concentration, and product generation rate. The cooperative change matrix is ​​input into the feature extraction network, which outputs a multidimensional feature vector containing the intensity and stability of the enzymatic reaction.

[0006] Preferably, an enzymatic hydrolysis reaction pathway association model is constructed, wherein the enzymatic hydrolysis reaction pathway association model uses a multidimensional feature vector as the input layer; The correlation weight coefficients of each dimension of the feature vector of the enzymatic hydrolysis reaction are calculated using a deep belief network. The enzyme hydrolysis reaction maturity value is generated based on the correlation weight coefficient, and the enzyme hydrolysis reaction state is divided into insufficient reaction state, optimal reaction state and over-reaction state according to the maturity value.

[0007] Preferably, the substrate residue change curve is tracked in real time under the condition of insufficient reaction; The abrupt change point of the slope of the substrate residue change curve was calculated using a discrete differential algorithm. The substrate residue fluctuation index is generated based on the slope mutation point, and the substrate residue fluctuation index is fed back to the input layer of the enzymatic reaction pathway association model.

[0008] Preferably, when the enzymatic hydrolysis reaction is in the optimal reaction state, the periodic fluctuation data of the product generation rate are collected; The dominant frequency amplitude value of periodic fluctuation data is extracted using spectral analysis. The product generation rate stability threshold is set based on the main frequency amplitude value, and the adjustment amplitude of the enzymatic hydrolysis reaction process control command is dynamically matched.

[0009] Preferably, if the enzymatic hydrolysis reaction state switches to an over-reaction state, the enzyme activity decay trajectory data is acquired in real time; By compressing the dimensions of enzyme activity decay trajectory data using a convolutional autoencoder, an enzyme activity decay feature code is generated. By coupling the enzyme activity decay characteristic encoding with the substrate residue fluctuation index, an abnormal risk value for enzymatic hydrolysis reaction is generated.

[0010] Preferably, the abnormal risk value of the enzymatic hydrolysis reaction is input into the decision tree classifier, and the enzymatic hydrolysis reaction process is paused or the rate is inhibited. When the output rate suppression command is executed, the gradient descent algorithm is used to iteratively optimize the target control range of the substrate concentration value. The optimized target control range is synchronously updated to the weight parameters of the enzymatic reaction pathway association model.

[0011] Preferably, a historical database of the enzymatic hydrolysis reaction process is established, wherein the historical database stores the mapping relationship between enzyme activity values, substrate concentration values ​​and product generation rate values ​​under different reaction states; The K-means clustering algorithm was used to classify the mapping relationships in the historical database into patterns, generating a template library of typical enzymatic hydrolysis reaction curves; The matching results between real-time dynamic data and a library of typical enzymatic hydrolysis reaction curve templates are used to calibrate the maturity value of the enzymatic hydrolysis reaction.

[0012] Preferably, a feedback closed-loop mechanism for detecting components of the enzymatic hydrolysis reaction is constructed, wherein the execution result of the rate inhibition command is used as a new sample; The hidden layer node parameters of the deep belief network are updated using an online incremental learning algorithm; The association weight coefficients of the enzymatic hydrolysis reaction feature vector are regenerated based on the updated deep belief network.

[0013] Preferably, the present invention further includes a detection system for enzymatic hydrolysis reaction components of reconstituted nutritional supplements, used to implement the above-described method for detecting enzymatic hydrolysis reaction components of reconstituted nutritional supplements, the system comprising: The dynamic data acquisition module is configured to acquire enzyme activity, substrate concentration, and product generation rate values ​​through multi-source sensors. The feature extraction and fusion module is configured to perform multimodal feature fusion on dynamic data and generate enzymatic hydrolysis reaction feature vectors; The reaction state classification module is configured to identify incompletely reacted states, optimal reaction states, and over-reacted states based on the feature vectors of enzymatic hydrolysis reactions. The parameter dynamic adjustment module is configured to adjust the detection parameter thresholds according to different reaction states and generate enzymatic hydrolysis reaction process control instructions. The feedback optimization module is configured to update the parameters of the enzymatic hydrolysis reaction path association model based on the execution results of control commands.

[0014] Compared with the prior art, the beneficial effects of the present invention are: This method for detecting the enzymatic hydrolysis components in instant nutritional supplements utilizes multi-source sensors to collect dynamic data in real time, including enzyme activity, substrate concentration, and product formation rate. This allows for the simultaneous capture of instantaneous changes in multiple key indicators during the enzymatic hydrolysis reaction. Compared to traditional single-dimensional data acquisition methods, the application of multi-source sensors covers multiple aspects of the enzymatic hydrolysis reaction, enabling the timely capture of subtle fluctuations during the reaction process, thus providing a more detailed reflection of the reaction's real-time status. This real-time and multi-dimensional data acquisition mode avoids misjudgments of the reaction progress due to missing information. Multimodal feature fusion processing of dynamic data generates enzymatic hydrolysis reaction feature vectors, integrating previously scattered data such as enzyme activity, substrate concentration, and product formation rate into correlated feature information. Multimodal feature fusion is not simply data aggregation; rather, it uses algorithms to uncover the intrinsic relationships between different data points, forming feature vectors that better represent the overall state of the enzymatic hydrolysis reaction. This integration method overcomes the limitations of single data points, allowing subsequent state judgments to be based on more comprehensive information, making the understanding of the enzymatic hydrolysis reaction state closer to reality. By classifying reaction states based on enzymatic hydrolysis reaction feature vectors and dynamically adjusting detection parameter thresholds according to different state categories, the detection system can achieve greater adaptability. Enzymatic hydrolysis reactions exhibit different state characteristics at different stages, and fixed parameter thresholds are difficult to adapt to all states. However, dynamically adjusting the thresholds according to the actual state allows the detection standard to match the real-time characteristics of the reaction, enabling the use of appropriate judgment standards at different stages such as the reaction initiation period and the rapid reaction period, reducing detection bias caused by rigid standards. Based on the deviation between the real-time reaction status and parameter thresholds, a nonlinear optimization algorithm is used to correct the control commands for the enzymatic hydrolysis reaction process, making the reaction regulation more closely aligned with actual needs. The nonlinear optimization algorithm can handle complex nonlinear relationships in the reaction, generating more reasonable correction commands based on the specific deviations, rather than simple linear adjustments. This correction method allows for more precise intervention when the enzymatic hydrolysis reaction deviates from expectations, guiding the reaction process in the ideal direction and avoiding reaction imbalances caused by improper regulation. Attached Figure Description

[0015] Figure 1 This is a schematic diagram illustrating the working principle of the method for detecting enzymatic hydrolysis components in instant nutritional supplements according to the present invention. Figure 2 Design diagram for collaborative processing of multi-source sensor data; Figure 3 Design diagram for insufficient state tracking; Figure 4 This is a design diagram for detecting abnormal overreaction states. 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 a method for detecting enzymatic hydrolysis components in instant nutritional supplements, the method comprising: Dynamic data during the enzymatic hydrolysis reaction process are acquired in real time using multi-source sensors. This dynamic data includes enzyme activity, substrate concentration, and product formation rate. The multi-source sensors are deployed within the enzymatic hydrolysis reaction vessel and synchronously acquire these three types of data at a fixed sampling frequency, forming a time-aligned raw data stream.

[0018] The dynamic data is subjected to multimodal feature fusion processing to generate enzymatic reaction feature vectors. After standardization and noise reduction, the original data stream is used to extract time series segments using a sliding window mechanism. The correlation index between enzyme activity fluctuation intensity, substrate consumption rate and product formation trend is calculated through a feature extraction network, and finally, a fusion result containing 128-dimensional feature vectors is output.

[0019] Enzymatic hydrolysis reaction states are categorized based on feature vectors, and detection parameter thresholds are dynamically adjusted according to different state categories. A deep belief network is used to classify the feature vectors, outputting the probability distributions of insufficiently reacted states, optimal reaction states, and over-reacted states. Different enzyme activity thresholds, substrate residue thresholds, and product formation rate thresholds are preset for each state category, forming a dynamic set of detection parameters.

[0020] Based on the real-time deviation between the enzymatic hydrolysis reaction status and the parameter threshold, a nonlinear optimization algorithm is used to correct the enzymatic hydrolysis reaction process control commands. When the deviation between the current status and the target threshold is detected to exceed the allowable range, the Levenberg-Marquardt algorithm is invoked to iteratively calculate the optimal control quantity, generating commands for temperature adjustment, enzyme addition rate adjustment, or stirring intensity adjustment.

[0021] Example 1: See Figure 2The system for detecting the components of the enzymatic hydrolysis reaction in reconstituted nutritional supplements is equipped with three core sensor units: An enzyme activity biosensor employs an immobilized enzyme electrode structure, with a specific protease covalently bound to its sensing surface. It monitors the current changes in the reaction system in real time via an electrochemical workstation. The current signal amplitude is positively correlated with the enzyme-catalyzed reaction rate. A standardized enzyme activity voltage signal is output via a current-to-voltage conversion module, with a sampling frequency set at 10Hz. A substrate concentration spectrometer integrates a near-infrared LED light source and a photodiode array. The light source wavelength range is set to 1650-1750nm. After penetrating the reaction solution, the intensity of the transmission spectrum is detected by the receiver. The transmission data is converted into substrate concentration values ​​according to Beer-Lambert's law, and the measurement results are updated every 200 milliseconds. A product formation rate meter includes a high-precision micro-flow meter and a density sensor. The flow meter detects the volume of product flowing into the collection container per unit time, while the density sensor simultaneously measures the change in fluid density. The data from both are fused and the product formation rate is calculated using the mass conservation equation, with a data output interval of 500 milliseconds.

[0022] The standardization and denoising process for dynamic data comprises three parallel channels. The enzyme activity voltage signal first undergoes baseline correction. After subtracting the sensor's inherent background current, it is standardized using a sliding window Z-score: the mean and standard deviation of 100 data points within the window are calculated, converting the original voltage values ​​into a standard normal distribution. Substrate concentration spectral data is processed using a polynomial fitting denoising method. A fifth-order polynomial trend line is constructed for the spectral data at each sampling point, and the high-frequency fluctuation components are retained after subtracting the trend line from the original values. Product formation rate data undergoes double smoothing: first, a median filter with a 1-second time window is applied to eliminate impulse noise; then, an exponentially weighted moving average with a 5-second window is used to reduce random fluctuations. The three types of preprocessed data streams are aligned using a unified timestamp. Each alignment point corresponds to the enzyme activity value, substrate concentration value, and product formation rate value, forming a coordinated change matrix with the timestamp as the row index and the three channel parameters as column vectors.

[0023] The collaborative change matrix is ​​input to a feature extraction network for processing. The first layer of the network adopts a bidirectional long short-term memory structure, containing 64 memory units, 32 in the forward direction and 32 in the reverse direction. This layer extracts the long-term dependence features of three types of parameters in the time dimension, including the phase delay between enzyme activity and substrate concentration, and the periodic pattern of product generation rate. The second layer, the attention mechanism layer, sets up 8 attention heads, each of which calculates the dynamic weight coefficients of enzyme activity and substrate concentration. In the specific calculation, the substrate concentration sequence is used as the query vector, and the enzyme activity sequence is used as the key vector. Dynamic association weights are generated through a dot product attention scoring mechanism. The third fully connected layer receives the fused features from the first two layers and contains 12 neuron nodes. This layer reduces the dimensionality of the multidimensional features to a 32-dimensional feature vector, where the first 16 dimensions represent the intensity features of the enzymatic reaction, including the slope of enzyme activity change and the acceleration of substrate consumption; the last 16 dimensions represent the stability features, including the variance of enzyme activity fluctuation and the correlation coefficient between substrate and product generation rate.

[0024] The enzymatic hydrolysis reaction pathway association model is constructed as a four-layer deep network. After receiving a 32-dimensional feature vector as input, the first hidden layer employs a Restricted Boltzmann Machine (RBM) structure, containing 50 visible units and 30 hidden units, with connection weights trained using a contrastive divergence algorithm. The second hidden layer is a Sigmoid confidence network layer with 20 neurons, calculating the nonlinear interactions between features. The output layer is designed as a radial basis function network, with enzyme activity, substrate concentration, and product formation rate as target variables. Network training uses a supervised learning model, with training samples containing 500 sets of labeled enzymatic hydrolysis reaction process data. The connection strength between neurons is updated in each iteration. After training, the model outputs three sets of association weight coefficients: weights influenced by enzyme activity, weights influenced by substrate concentration, and weights influenced by product formation. These weight coefficients are normalized, and the sum of the three is constant at 1.

[0025] The maturity value of the enzymatic hydrolysis reaction is calculated using a weighted fusion method. The 32 dimensions of the feature vector are multiplied by their corresponding dimension's association weight coefficient, and the products are summed. To improve computational efficiency, a parallel processing architecture is adopted: the 32-dimensional vector is divided into four groups of eight-dimensional sub-vectors, each group is assigned a dedicated multiplier-accumulator, and the results from each group are finally integrated by an addition tree. The maturity value is set to a continuous range between 0 and 1, and its changes reflect the overall progress of the enzymatic hydrolysis reaction. A state classifier monitors the maturity value in real time: when 10 consecutive sampled values ​​are below 0.3, an insufficiently reacted state flag is activated; when 20 consecutive sampled values ​​are in the range of 0.3 to 0.7, it is marked as the optimal reaction state; when the maturity value exceeds 0.7 and remains above 0.7 for more than 5 seconds, it is determined to be an over-reacted state. The classification results are transmitted to the parameter adjustment module in real time and simultaneously written to the state log buffer for historical backtracking.

[0026] To maintain model adaptability, a dynamic update mechanism for the associated model parameters is implemented. When the same batch of enzymatic hydrolysis reaction continues for more than 30 minutes, the system automatically initiates an online fine-tuning procedure. During fine-tuning, the feature extraction network adds the feature vector of the current reaction as a temporary input node, and updates the connection weights of the restricted Boltzmann machine layer through forward propagation. A feedback loop is added to the model output: when there is a deviation between the state classification result and the actual observation value, the deviation signal is backpropagated to the radial basis function network layer to adjust the width coefficients and center positions of the basis functions. After each fine-tuning, the system recalculates the associated weight coefficients, overwriting the original parameter storage area. This mechanism enables the model to continuously adapt to the changing characteristics of different raw material batches and environmental conditions.

[0027] Feature vector transmission employs a pipelined architecture to improve real-time performance. After the co-transformation matrix is ​​generated, it is divided into 64×64 sub-matrix blocks and input into the processing queue. The feature extraction network is designed with a double buffering mechanism: while the first sub-matrix is ​​being processed in the bidirectional long short-time memory layer, the second sub-matrix is ​​simultaneously loaded into the input register. The attention mechanism layer performs time-critical operations, using a hardware accelerator to implement parallel dot product operations. Matrix multiplication in the fully connected layers is performed using a systolic array, keeping the processing latency of a single 32-dimensional vector within 8 milliseconds. During the model prediction phase, a checksum is added to the output of each hidden layer of the deep belief network, and the receiver uses cyclic redundancy check (CRC) to detect the integrity of data transmission. When a checksum error is detected, the system automatically triggers a resampling and processing procedure for the data in the current time window.

[0028] The state classification module has a built-in anomaly detection function. When the maturity value changes by more than 0.15 within five consecutive sampling periods, a drift alarm signal is activated. After the alarm is triggered, the system automatically freezes the execution of parameter adjustment commands and initiates the sensor calibration program: the enzyme activity biosensor performs zero-point calibration, the substrate concentration spectrometer performs white background correction, and the product formation rate meter initiates the empty pipeline self-cleaning program. After calibration is completed, the system reinitializes the co-variance matrix and clears the classification state records during the alarm period. Once the maturity value returns to a stable change, the classification module continues to perform state monitoring tasks from the frozen point. This protection mechanism effectively avoids misclassification caused by sensor drift.

[0029] Example 2: See Figure 3 When the enzymatic hydrolysis reaction of the prepared nutritional supplement is in an incompletely reacted state, the enzyme activity biosensor continuously outputs a current signal, and the substrate concentration spectrometer acquires a 1750nm wavelength transmission spectrum. The data processing unit extracts the substrate concentration values ​​to form a time series, with a fixed sampling interval of 500 milliseconds. The substrate residue change curve is constructed using a sliding window mechanism: a data point is generated every 5 seconds, and the value of this point is the arithmetic mean of 10 samples within the window. The curve is stored in a circular buffer with a capacity of 120 data points, covering a 10-minute reaction time.

[0030] The discrete differential algorithm is implemented by calculating the difference values ​​of the sequence. The processor performs a subtraction operation between the current data point and the previous data point to generate a first-order difference sequence. The second derivative is calculated using the central difference method: three adjacent difference values ​​are taken, and the result of subtracting the previous value from the subsequent value is taken as the second derivative value at that time point. The system sets a floating judgment threshold: when the second derivative values ​​at three consecutive time points all exceed the baseline threshold of 0.05, the intermediate time point is marked as a slope abrupt change point. The baseline threshold is dynamically adjusted according to the initial substrate concentration; for every 10 g / L increase in the initial concentration, the threshold is increased by 0.01 accordingly. The detection results are stored in the event log in the form of timestamp-abrupt change intensity pairs.

[0031] The substrate residue fluctuation index is generated from event log data. The system counts the number of mutation points within a 20-second time window and calculates the average of the absolute values ​​of the second derivatives corresponding to each mutation point. The ratio of the window duration to the number of mutation points is used as a time factor, which is multiplied by the average mutation intensity to obtain the initial value of the fluctuation index. The initial value is then normalized to the 0-1 range after being compressed by an order of magnitude through a logarithmic transformation and finally written into a 32-bit floating-point variable. The fluctuation index is fed back to the enzymatic reaction pathway association model asynchronously: before each forward inference in the association model, the fluctuation index is concatenated with the original input feature vector to form a 36-dimensional extended vector.

[0032] Once the reaction reaches its optimal state, the product formation rate meter activates a high-speed sampling mode. The microflow meter's sampling frequency is increased to 20Hz, and the density sensor simultaneously intensifies its scanning to 5 measurements per second. The data processing unit constructs a dual-channel data stream: the flow rate data is smoothed using a Kalman filter, and the density data undergoes adaptive threshold filtering. After timestamp alignment, the dual-channel data undergoes mass conservation calculations to generate a product formation rate curve with a time resolution of 0.05 seconds.

[0033] Spectral analysis employs the sliding Fourier transform method. The system uses a 10-second basic analysis window, with 200 data points within the window preprocessed using a Hanning window function. The fast Fourier transform processor performs complex number operations, outputting the frequency domain amplitude spectrum. The dominant frequency identification algorithm traverses the 0.1Hz to 2Hz frequency band to locate the frequency value corresponding to the maximum amplitude peak. Amplitude calculation consists of two steps: extracting the maximum amplitude within a ±0.05Hz bandwidth near the dominant frequency in the frequency domain; and measuring the peak-to-valley difference within the corresponding period in the time domain. The final amplitude value is the geometric mean of the frequency domain amplitude and the time domain peak-to-valley difference.

[0034] The product generation rate stabilization threshold is dynamically set based on the dominant frequency amplitude. The control parameter library has a preset amplitude-threshold mapping table: amplitude values ​​between 0 and 0.05 correspond to a ±2% fluctuation range, between 0.05 and 0.1 correspond to a ±4% range, between 0.1 and 0.2 correspond to a ±8% range, and fluctuations exceeding 0.2 are allowed at ±12%. In actual control, the system updates the average amplitude value over the last 30 seconds every second. After obtaining the basic threshold from the table, it is corrected based on the current reaction temperature: for every 1°C increase in temperature, the threshold range expands by 0.3 percentage points.

[0035] Steady-state control commands are generated based on a real-time monitoring mechanism. A data comparator performs 200 checks per second: the difference between the current product generation rate and the target baseline is divided by the target value to generate a deviation percentage, which is compared to a stability threshold. When 10 consecutive checks show a deviation exceeding the upper threshold, a substrate replenishment command is activated; when 15 consecutive checks show a deviation below the lower limit, an enzyme flow acceleration command is generated. The command amplitude uses a proportional control strategy: the absolute value of the deviation is multiplied by a gain coefficient of 0.8 to obtain the control quantity, which is then converted into a peristaltic pump speed adjustment or valve opening increment.

[0036] The multi-instruction conflict resolution module is designed with a dual-priority structure. When substrate replenishment and enzyme preparation addition instructions are generated simultaneously, the following decision-making process is executed: The current value of the residual substrate is checked; if it is lower than 40% of the initial concentration, substrate replenishment is executed first; otherwise, the enzyme activity value is checked; if it is lower than 70% of the optimal reaction threshold, enzyme preparation flow acceleration is executed first. Conflicting instructions are temporarily stored in a delay queue and the execution conditions are re-evaluated after 50 milliseconds. All executed instructions are recorded in the operation log, including timestamps, instruction types, and execution intensity parameters.

[0037] The state transition protection mechanism is activated in the critical region of the reaction state. When the maturity value fluctuates between 0.28 and 0.32, a state lockout procedure is triggered: dynamic adjustment of parameter thresholds is paused, and the control logic of the previous state is maintained. During the lockout period, continuous monitoring is performed for 30 seconds. If the maturity value steadily rises above 0.33, the optimal reaction state is confirmed; if it remains below 0.29, it reverts to the insufficient state. The state transition event generates an alarm signal, notifying the system to refresh the control parameter set and update the attention weight configuration of the feature extraction network.

[0038] Multi-point monitoring data within the reaction vessel are fused and processed. Substrate residue data are derived from measurements taken by three spectral probes at different depths, and a weighted average algorithm is used for fusion: the bottom probe has a weight of 0.6, the middle layer 0.3, and the surface layer 0.1. An eddy current compensation coefficient is introduced into the product formation rate calculation—when the stirrer speed exceeds 600 rpm, the flow meter reading is corrected proportionally to the speed. The data correction module is activated after each control command is issued, updating the parameters of the substrate mass balance equation; the correction period is set to the third second after the command is executed.

[0039] The control strategy self-optimization module runs periodically. Every 5 minutes, it analyzes the control records in the operation log and calculates the command execution success rate: using the target fluctuation range as a benchmark, it calculates the deviation compliance rate within 20 seconds after the command is issued. Command types with a compliance rate below 80% trigger a parameter optimization program: adjusting the gain coefficient in ±0.1 steps, simulating the control effect under different coefficients, and selecting the minimum gain value that satisfies an 85% success rate. The optimization results are updated to the control strategy library, and the original parameter values ​​are recorded for anomaly rollback.

[0040] Sensor fault emergency logic is embedded in the state machine kernel. When any sensor misses 10 consecutive sampling values, it automatically switches to a redundant operating mode: when enzyme activity values ​​are missing, reaction temperature is used as a substitute; when substrate concentration values ​​are missing, estimation is performed based on the product formation rate; when product rate values ​​are missing, historical data from the same state is used for interpolation. After the fault lasts for 90 seconds, the system degrades to a fixed parameter control mode until sensor data returns to stable output. All abnormal events are recorded in the system diagnostic report, triggering a yellow alarm.

[0041] Example 3: See Figure 4 When the enzymatic reaction transitions to an overreaction state, the enzyme activity biosensor activates its decay monitoring mode. The sensor acquisition interval is shortened to 100 milliseconds, and the current signal is converted into a digital sequence after 16-bit analog-to-digital conversion. The decay trajectory data preprocessing consists of three stages: subtracting the zero-point drift baseline value from the original current value, performing a logarithmic transformation on the result to compress the dynamic range, and finally using sliding window normalization to map the data to the [0,1] interval. The normalization window width is set to 30 seconds, and the normalization parameter update is paused when the difference between the maximum and minimum values ​​within the window is less than 0.01. The preprocessed data stream is written to a circular buffer, designed to store 300 consecutive sampling points.

[0042] The convolutional autoencoder network architecture is designed as a symmetrical structure. The encoder part contains two convolutional layers: the first layer uses 64 5-point one-dimensional convolutional kernels with a stride of 2, and the output feature map is activated by the ReLU function; the second layer uses 32 3-point convolutional kernels with the stride maintained at 2, and the activation function is LeakyReLU. Each convolutional layer is followed by a max pooling operation with a fixed pooling window of 2 points. The decoder part mirrors the encoder structure: the first deconvolutional layer uses 32 3-point kernels, the second deconvolutional layer uses 64 5-point kernels, and the final output layer uses a linear activation function to reconstruct the input signal. The network is trained using the mean squared error loss function, the Adam algorithm is used as the optimizer, and the initial learning rate is set to 0.001.

[0043] The enzyme activity decay feature encoding generation process is as follows: A preprocessed 300-point time series sequence is input into the encoder network. After two convolution and pooling operations, an 8-dimensional feature vector is output. Each dimension of the feature vector represents: the first dimension corresponds to the overall decay slope; the second dimension reflects the nonlinearity of the decay curve; the third to fifth dimensions describe local fluctuation characteristics; and the sixth to eighth dimensions represent the similarity to the standard decay pattern. The feature encoding is stored as a single-precision floating-point array and updated once per second.

[0044] The substrate residue fluctuation index is obtained from the feedback channel of Example 2, and its value range is limited to the interval [0,1]. The coupling analysis module receives the feature code and the fluctuation index, and performs the following calculation process: Where φ represents the abnormal risk value of the enzymatic hydrolysis reaction, ξ is the first dimension value of the enzyme activity decay characteristic encoding, δ represents the substrate residue fluctuation index, and α and β are the decay weighting factor and fluctuation weighting factor, respectively, with default values ​​set to α=0.7 and β=0.3. When the calculated result φ exceeds 0.5, the decision-making process is triggered, and each increase of 0.1 in the risk value corresponds to an increase of one alarm level.

[0045] The decision tree classifier is constructed as a three-level structure. The first-level nodes determine if the risk value exceeds 0.7: if so, a process pause command is output directly; otherwise, the process proceeds to the second level to determine if the decay slope ξ is greater than 0.8. The second-level affirmative branch continues to detect the volatility index δ: if δ exceeds 0.6, a pause command is output; otherwise, a rate suppression command is generated. The third level handles the remaining cases: it checks if the fifth dimension of the feature encoding has increased three consecutive times; if the condition is met, a supplementary detection process is triggered. Each branch node of the decision tree has a timeout limit, with the total time for a single decision not exceeding 50 milliseconds.

[0046] The rate suppression instruction execution phase employs a gradient descent algorithm to optimize control parameters. The initial target range for substrate concentration is set to ±20% of the current value, and the loss function is defined as the squared difference between the risk value φ and the target threshold of 0.3. In each iteration, the partial derivative of the loss function with respect to the target range boundary is calculated, and the boundary values ​​are adjusted according to the sign of the partial derivative. The step size is set to 0.5% of the concentration range. The iteration terminates when any of the following conditions are met: the loss function value is below 0.01, the improvement amount is less than 0.001 for three consecutive iterations, or the total number of iterations reaches 50. The optimization results are written to the control parameter register, and the target range indicator bar on the display interface is updated simultaneously.

[0047] The weight parameters of the enzymatic hydrolysis reaction pathway association model are updated using an incremental learning approach. The optimized target range is converted into a 4-dimensional feature vector, including upper and lower boundary values, boundary differences, and historical trends. This vector is concatenated with the original model's input features, and the weight adjustment is calculated through a two-layer fully connected network. The first layer contains 16 neurons, using the tanh activation function; the second layer has the same number of output nodes as the original model's weight parameters and uses a linear output. The adjustment is multiplied by a learning rate coefficient of 0.1 and then added to the original weights. The entire process is completed within 200 milliseconds.

[0048] The anomaly handling subsystem is designed with a multi-level response mechanism. The primary response is activated when the risk value φ∈(0.5,0.7]: reducing the reaction system temperature by 2℃, decreasing the stirring speed by 15%, and increasing the substrate concentration monitoring frequency. The intermediate response corresponds to φ∈(0.7,0.9]: immediately stopping enzyme addition, initiating cooling water circulation, and adjusting the pH of the reaction solution to the neutral range. The emergency response is executed when φ>0.9: completely terminating the reaction process, opening the emergency discharge valve, and triggering audible and visual alarm signals. Each level of response operation is recorded in the event log, along with a timestamp and a risk value snapshot.

[0049] The real-time monitoring interface integrates 3D visualization functionality. Enzyme activity decay trajectories are displayed as curves, accompanied by moving averages and confidence intervals. Substrate residue fluctuation indices are represented by bar charts, with bar heights proportional to index values ​​and colors transitioning from green to red to reflect risk levels. Abnormal risk values ​​are displayed in a dashboard format, with pointer angles linearly corresponding to risk values, and background color divisions indicating safe, warning, and danger zones. The right panel of the interface dynamically updates the last 10 control commands and their execution effects.

[0050] The data backup module employs a dual storage strategy. Real-time running data is written to the memory cache every second and batch-transferred to the solid-state drive every 5 minutes. Key parameter change records are appended to a read-only log file, including intermediate results of risk value calculations, decision tree path selection, and weight adjustment details. The system automatically generates an integrity checksum every 24 hours, and checks for abnormal data changes by comparing the checksum. If data inconsistencies are detected, the parameter configuration is reloaded from the previous checksum point.

[0051] A hardware watchdog circuit ensures system reliability. The monitoring chip on the main control board operates independently, receiving a heartbeat signal from the software every 500 milliseconds. If the heartbeat is lost more than three times consecutively, the hardware circuit performs a tiered reset: first, it attempts to restart the application; if it fails to recover within 10 seconds, it restarts the operating system; the final measure is to cut off power and switch to the backup controller. The reset event is recorded in non-volatile memory, containing the last received sensor data packet and system status code.

[0052] The simulation testing interface supports offline verification. It simulates enzyme activity decay patterns using a virtual data generator: parameters such as initial slope, inflection point position, and terminal fluctuation amplitude are set to generate test sequences that conform to typical decay characteristics. In test mode, actual control command output is disabled; instead, theoretical control quantities are recorded and compared with expected results. A difference analysis report is automatically generated, annotating operation points with timing deviations exceeding 100 milliseconds and control items with parameter deviations exceeding 5%.

[0053] The version control subsystem manages algorithm updates. Each time the decision tree structure or neural network parameters are modified, a new version branch is created and a complete configuration snapshot is saved. Version information includes a timestamp, the modifier's code, and a change summary, supporting rollback to any historical version at any point in time. Abnormal events occurring during runtime are automatically associated with the current version number for statistical analysis of stability metrics for each version. The repository is periodically archived to a remote server, retaining the complete development history for the most recent 30 days.

[0054] Example 4: The historical database for the enzymatic hydrolysis reaction process of reconstituted nutritional products is stored using a distributed architecture, containing three main data tables: a reaction status record table storing enzyme activity values, substrate concentration values, and product formation rate values ​​at different time points; a control command log table recording all executed adjustment commands and their parameters; and an exception event table storing various alarm information detected by the system. The database uses a time-series storage engine, with each record accompanied by a timestamp accurate to milliseconds, supporting combined queries by batch number and reaction stage. Data tables are linked through transaction IDs to ensure complete traceability of operation records.

[0055] The core fields of the reaction status record table are designed as follows: Typical data recording snippets are shown below: When a batch of reactions enters its optimal state, the system records 6 complete sets of state parameters per minute. When fluctuations in the product generation rate exceed a threshold, the recording frequency is automatically increased to 2 sets per second. Each time a state transition occurs, the system additionally records 30 seconds of high-density sampled data before and after the transition, forming a state transition feature dataset.

[0056] Before executing the K-means clustering algorithm, historical data was preprocessed. Numerical fields were uniformly normalized using Min-Max, with enzyme activity values ​​mapped to the [0,1] range, substrate concentration values ​​converted to the [0,100] interval, and product generation rates adjusted to the [0,10] scale. Categorical fields, such as reaction stage, were converted to one-hot encoding, and temperature values ​​were binned in 5°C intervals. The preprocessed data matrix had a dimension of N×7, where N represents the total number of samples, and the 7 columns corresponded to the normalized feature dimensions.

[0057] The clustering process uses a modified K-means++ algorithm to initialize the centroids. First, the first cluster center is randomly selected. The probability of selecting each subsequent new centroid is proportional to the square of the minimum distance from the previously selected centroid. The algorithm is set to a maximum of 100 iterations, and the convergence threshold is set to a centroid movement distance of less than 0.001. When determining the optimal number of clusters using the elbow rule, the silhouette coefficient is calculated for K values ​​ranging from 2 to 10, and the K value with the largest rate of change is selected as the final number of clusters.

[0058] The construction process of the typical enzymatic hydrolysis reaction curve template library is as follows: Each cluster centroid corresponds to a standard reaction curve, which contains seven dimensions of feature values. The system extracts the 20 actual reaction samples closest to the centroid in each cluster, calculates the mean and standard deviation of their feature values ​​at each time point, and forms template curves with confidence intervals. The template metadata storage includes: curve ID, reaction stage, average duration, typical control command sequence, etc.

[0059] The matching of real-time data with the template library employs a dynamic time warping algorithm. The system transforms the current reaction data using the same preprocessing procedure and then compares it point-by-point with each template curve. The matching degree calculation considers three dimensions: the shape similarity of the enzyme activity curves, the consistency of substrate concentration change trends, and the phase difference of product formation rates. Each dimension is assigned a weighting coefficient, and the final matching score is the weighted average. When the scores of multiple templates differ by less than 5%, a secondary matching process is initiated: comparing the fit between the actual control command sequence and the typical template commands.

[0060] The maturity value calibration module's workflow comprises two stages: data alignment and bias compensation. The alignment stage linearly compresses or expands the timeline of the real-time curve to match the template length. The compensation stage calculates the eigenvalue deviations between each point on the real-time curve and the template curve, multiplying the mean deviation by the matching score to obtain the calibration factor. The calibrated maturity value is calculated as: original value plus the calibration factor multiplied by the template confidence weight. The system has a calibration cap mechanism, limiting the calibration amplitude to no more than 15% of the original value per calibration.

[0061] The template library maintenance subsystem implements an automated update mechanism. A template optimization process is executed weekly: new historical data is added to the clustering sample set, and the positions of the cluster centroids are recalculated. When the matching usage rate of a template is below 5% for two consecutive weeks, a template merging check is triggered: the center distance between this template and its adjacent templates is calculated; if the distance is less than 60% of the average spacing, the two templates are merged, and a new standard curve is generated. After each template update, the system automatically recalculates the matching records of all historical data and updates the template usage statistics report.

[0062] The anomaly template detection function monitors matching results in real time. When the deviation of a batch reaction curve from the best-matching template consistently exceeds three standard deviations, a potential anomaly event is generated. The system extracts detailed parameters for the anomaly period, including: the feature dimension with the largest deviation, the duration, and accompanying control instructions. This information is stored in a review queue, where quality engineers manually review and confirm it before deciding whether to include it in the template library as a new anomaly pattern reference.

[0063] The data visualization interface displays the template matching process. The main view overlays the real-time reaction curve with the best-matching template, using different colors to distinguish the three trajectories of enzyme activity, substrate concentration, and product formation rate. The sidebar panel lists the information of the top three matching templates, including template ID, cluster, and historical matching count. Clicking on any template allows you to view its detailed parameter distribution and typical control strategies. The status bar at the bottom of the interface displays auxiliary information such as maturity value calibration and current template confidence level in real time.

[0064] The version control system manages the evolution history of the template library. Each template update creates a new version, recording changes categorized as follows: templates with a center point shift exceeding 10% are marked as major updates; adding or deleting templates is considered a structural adjustment; and columns with changed parameter distribution but a stable center point are considered minor adjustments. The version rollback function supports restoring to any historical version. Rollback operations require dual authentication and record a complete audit log.

[0065] The table data export function meets the needs of quality analysis. Users can filter matching records for specific time periods or batches to export detailed comparison tables containing raw values, template values, and calibration values. Export file formats support both CSV and JSON standards, and field configuration allows for custom selection. The system automatically appends data description fields to the exported file, explaining the meaning of each field and unit conversion relationships.

[0066] The offline analysis toolkit provides batch processing capabilities. Quality analysts can upload multiple batches of historical data files, and the system background performs batch template matching calculations to generate a comprehensive report containing feature classifications and outlier detection results for each batch. The report is presented in a hierarchical structure: the summary page displays the overall matching distribution, while the details page allows drill-down to view parameter deviations at each outlier time point. Analysis tasks can be scheduled for execution, making full use of the system's idle computing resources.

[0067] The access control system manages access permissions to the template library. Basic operators can only view template matching results, process engineers have permission to provide template usage feedback, and algorithm administrators have permission to adjust the template library structure. Each template modification requires a reason for the modification, and the system automatically associates the operator's information and the modification time, forming a complete chain of responsibility. Sensitive operations such as template deletion or batch import require electronic signature confirmation from a second-level supervisor.

[0068] The data backup strategy employs a triple-layer protection mechanism. Real-time operational data is synchronized to a local redundant server every 15 minutes, a full backup is performed daily at midnight to enterprise-grade NAS storage, and incremental backups are uploaded to cloud object storage weekly. Backup data includes a complete template library, matching records, and calibration logs, with a retention period of 30 days locally and 180 days in the cloud. Backup and recovery tests are performed quarterly to verify data integrity and system restore capabilities.

[0069] Example 5: When constructing a feedback closed-loop mechanism for the enzymatic hydrolysis component detection system for instant nutritional supplements, the execution results of rate inhibition commands are used as new sample data. After each command execution, the system automatically records the abnormal risk value change trend, command type code, execution intensity parameters, and risk value response curve within 180 seconds after execution for the 60 seconds prior to execution. After desensitization processing, the new sample data removes batch information and timestamps, retaining only pure technical parameters to form standardized learning samples. The sample queue adopts a first-in, first-out (FIFO) management strategy, with a maximum capacity of 500 groups. When the queue is full, the earliest stored sample is automatically discarded.

[0070] The online incremental learning algorithm employs an elastic weight consolidation method to update network parameters. The system maintains a parameter importance matrix, recording the historical changes in the weights of each connection in the deep belief network. For a new input sample, the partial derivatives of its loss function with respect to each weight are first calculated. Then, the learning rate is dynamically adjusted based on the ratio of the absolute value of the partial derivative to the corresponding value in the importance matrix. The importance matrix update follows the recency reinforcement principle: the weight changes caused by new samples are added to the historical values ​​with a decay coefficient of 0.7. Adjustments to the parameters of the hidden layer nodes are restricted to orthogonal space to avoid disrupting the learned important feature representations.

[0071] The hidden layer structure adjustment of the deep belief network adopts a gradual strategy. The network monitoring module continuously tracks the classification error of the most recent 100 new samples. When the average error exceeds 15% for three consecutive statistical periods, a node expansion procedure is triggered. The number of new nodes is 10% of the current number of hidden layer nodes, with a minimum of three. The connection weights of new nodes are initialized to small perturbations of the weights of neighboring nodes, and the bias term is set to zero. Conversely, when the classification error is below 5% for 10 consecutive statistical periods, the system initiates node pruning: removing the node with the smallest absolute value of the output weight and distributing its input weights proportionally to neighboring nodes. During the structure adjustment process, the network maintains online service capability, and new nodes operate in shadow mode, only officially connecting to the main network after 50 forward propagation verifications.

[0072] The process of regenerating the association weight coefficients includes feature importance evaluation. The system performs backpropagation calculations for each new sample, recording the contribution of each input feature to the final output. The contribution calculation uses the integral gradient method: 20 intermediate points are uniformly sampled from the baseline input to the actual input, and the gradient values ​​at each point are accumulated. The feature importance score is the exponential moving average of the contributions of the most recent 50 samples, with a decay factor set to 0.9. The final association weight coefficients are obtained by multiplying the output weights of the base network by the feature importance score, and then undergoing Softmax normalization.

[0073] The model update validation mechanism employs a dual-validation process. The primary validation is performed immediately after the parameter update: forward propagation is conducted using a validation set sample to check if the output values ​​are within a reasonable range. The validation set contains 200 sets of historical data, covering various reaction states and abnormal situations. The secondary validation runs in the actual control scenario: control commands generated by the updated model are marked as experimental commands and sent to the actuator in parallel with commands from the stable model. The actuator compares the differences between the two commands; if the difference exceeds a safety threshold, the stable command is used, and the model update is rolled back. Experimental commands with differences within the allowable range are executed, and their performance data is added to the validation set for subsequent model optimization.

[0074] An anomaly handling module protects the learning process from interference. The system detects outliers in newly added samples: when a sample's feature value exceeds three standard deviations from the historical database, an anomaly review is triggered. The review process calculates the Mahalanobis distance between the sample and the 50 most recent similar samples; if the distance exceeds a threshold, the sample is temporarily stored in the confirmation zone. Quality engineers periodically check the samples in the confirmation zone; valid samples are manually added to the learning queue, while invalid samples are marked for exclusion. The learning algorithm uses a halved learning rate for confirmed anomaly samples to reduce their impact on the model.

[0075] The real-time performance monitoring panel displays the incremental learning effect. The main view curve shows the 30-day moving average of the model's classification accuracy, while the auxiliary view presents the changing trend of the feature distribution of newly added samples. The console outputs detailed logs, recording information such as the time of each parameter update, the adjustment magnitude, and the hidden layer node numbers involved. The monitoring system has two levels of alerts: a yellow alert is triggered when the validation set error increases by more than 5% after three consecutive updates; a red alert is immediately triggered and the learning process is paused when the model output shows numerical overflow or NaN values.

[0076] The version rollback function ensures system stability. Before each major network structure adjustment, the system automatically generates a complete snapshot, including core data such as network topology, connection weights, and importance matrices. Snapshot storage uses differential compression technology, recording only the changes relative to the previous version. Rollback operations can be completed within 30 seconds. After rollback, the model automatically enters stable mode and can only resume incremental learning after at least a 24-hour observation period. Detailed reports are generated for rollback events, analyzing potential update steps and sample characteristics that may have caused problems.

[0077] The sample feature drift detection module continuously tracks changes in data distribution. The system maintains a feature space reference distribution, composed of the top 100 principal components of historical data. After receiving 50 new samples, the KL divergence between their projections onto these principal components and the reference distribution is calculated. When the divergence value exceeds a warning threshold, feature recalibration is initiated: the principal component orientation is reselected, and the feature scaling parameters of the network input layer are adjusted. The recalibration process is performed gradually to avoid sudden changes that could cause drastic fluctuations in control commands.

[0078] The adaptive learning rate adjustment algorithm balances convergence speed and stability. The base learning rate is set according to the network layer depth: 0.001 for the input layer, decreasing by 20% for each hidden layer, and fixed at 0.0005 for the output layer. The dynamic adjustment factor considers three indicators: the average gradient magnitude of recent samples, the sparsity of the parameter importance matrix, and the trend of validation set error changes. These three indicators are used to calculate the final adjustment coefficients through a three-layer perceptron network, with the coefficients limited to the range [0.5, 2.0]. The learning rate is recalculated every 24 hours, with adjustments not exceeding 30% of the current value.

[0079] Distributed storage of network parameters enhances system robustness. The weight matrix of the deep belief network is divided into multiple shards and stored on different physical nodes. Each shard retains three replicas, deployed on servers in independent fault domains. Parameter updates employ a two-phase commit protocol: new parameters are pre-written to all replicas, and only after a majority of nodes confirm success are the updates committed and take effect. Read operations prioritize accessing local replicas, and an automatic repair process is initiated when data inconsistency is detected.

[0080] A smooth transition mechanism for control commands avoids oscillations caused by parameter updates. When a model update is detected that results in three consecutive control commands moving in opposite directions, a command filtering program is initiated. The filtering algorithm calculates the weighted average of the previous five commands, with more recent commands having higher weights. Simultaneously, the response speed of the control system is temporarily reduced, and the action interval of the actuators is extended by 50%. Filtering continues until the command sequence returns to a stable trend, with a maximum duration of 30 minutes. Anomalies collected during this period are stored separately for subsequent model defect analysis.

[0081] Incremental learning and periodic full training work in tandem. The system initiates a full training task weekly: using all valid data from the historical database, it reinitializes and trains the deep belief network. The full training results are compared with the incremental learning model, and the version that performs better on the validation set is selected as the new baseline model. Baseline model switching uses a blue-green deployment approach: the new model runs in shadow mode for 24 hours, and after verification, it replaces the old model in the production environment. Parameter update records from both training methods are archived uniformly, forming a complete model evolution map.

[0082] The user feedback interface collects operators' experience-based judgments. A dedicated button on the console allows engineers to manually rate the control commands generated by the model. Ratings are categorized into three levels: "Meets expectations," "Acceptable," and "Needs improvement," each accompanied by a brief description. These subjective evaluations are converted into structured data: positive ratings reinforce the learning weights, while negative ratings trigger specialized analysis processes. The analysis results are used to adjust feature extraction strategies or correct network structural defects, forming a continuous improvement cycle through human-machine collaboration.

[0083] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0084] 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. A method for detecting enzymatic hydrolysis components in instant nutritional supplements, characterized in that, The method includes: Dynamic data during the enzymatic hydrolysis reaction process are collected in real time using multi-source sensors. The dynamic data includes enzyme activity value, substrate concentration value, and product formation rate value. The dynamic data is subjected to multimodal feature fusion processing to generate an enzymatic reaction feature vector; The enzyme hydrolysis reaction state categories are divided based on the feature vector of the enzyme hydrolysis reaction, and the detection parameter thresholds are dynamically adjusted according to different state categories. Based on the deviation between the real-time enzymatic hydrolysis reaction status and the parameter threshold, a nonlinear optimization algorithm is used to correct the enzymatic hydrolysis reaction process control command.

2. The method for detecting enzymatic hydrolysis components in reconstituted nutritional supplements according to claim 1, characterized in that, The method includes: The multi-source sensor includes an enzyme activity biosensor, a substrate concentration spectrometer, and a product formation rate meter. The dynamic data is standardized and denoised to construct a time-series-based matrix of the coordinated changes in enzyme activity, substrate concentration, and product generation rate. The collaborative change matrix is ​​input into the feature extraction network, which outputs a multidimensional feature vector containing the intensity and stability of the enzymatic reaction.

3. The method for detecting enzymatic hydrolysis components in reconstituted nutritional supplements according to claim 2, characterized in that, The method includes: A model for the association of enzymatic hydrolysis reaction pathways is constructed, wherein the model uses multidimensional feature vectors as the input layer. The correlation weight coefficients of each dimension of the feature vector of the enzymatic hydrolysis reaction are calculated using a deep belief network. The enzyme hydrolysis reaction maturity value is generated based on the correlation weight coefficient, and the enzyme hydrolysis reaction state is divided into insufficient reaction state, optimal reaction state and over-reaction state according to the maturity value.

4. The method for detecting enzymatic hydrolysis components in reconstituted nutritional supplements according to claim 3, characterized in that, The method includes: Real-time tracking of substrate residue changes under incomplete reaction conditions; The abrupt change point of the slope of the substrate residue change curve was calculated using a discrete differential algorithm. The substrate residue fluctuation index is generated based on the slope mutation point, and the substrate residue fluctuation index is fed back to the input layer of the enzymatic reaction pathway association model.

5. The method for detecting enzymatic hydrolysis components in reconstituted nutritional supplements according to claim 4, characterized in that, The method includes: When the enzymatic hydrolysis reaction is in its optimal state, periodic fluctuation data of the product generation rate are collected. The dominant frequency amplitude value of periodic fluctuation data is extracted using spectral analysis. The product generation rate stability threshold is set based on the main frequency amplitude value, and the adjustment amplitude of the enzymatic hydrolysis reaction process control command is dynamically matched.

6. The method for detecting enzymatic hydrolysis components in reconstituted nutritional supplements according to claim 5, characterized in that, The method includes: If the enzymatic reaction state switches to an over-reaction state, the enzyme activity decay trajectory data is acquired in real time. By compressing the dimensions of enzyme activity decay trajectory data using a convolutional autoencoder, an enzyme activity decay feature code is generated. By coupling the enzyme activity decay characteristic encoding with the substrate residue fluctuation index, an abnormal risk value for enzymatic hydrolysis reaction is generated.

7. The method for detecting enzymatic hydrolysis components in reconstituted nutritional supplements according to claim 6, characterized in that, The method includes: Input the abnormal risk value of the enzymatic hydrolysis reaction into the decision tree classifier, and output the enzymatic hydrolysis reaction process pause instruction or rate inhibition instruction; When the output rate suppression command is executed, the gradient descent algorithm is used to iteratively optimize the target control range of the substrate concentration value. The optimized target control range is synchronously updated to the weight parameters of the enzymatic reaction pathway association model.

8. The method for detecting enzymatic hydrolysis components in reconstituted nutritional supplements according to claim 7, characterized in that, The method includes: A historical database of the enzymatic hydrolysis reaction process is established, which stores the mapping relationship between enzyme activity values, substrate concentration values ​​and product formation rate values ​​under different reaction states; The K-means clustering algorithm was used to classify the mapping relationships in the historical database into patterns, generating a template library of typical enzymatic hydrolysis reaction curves; The matching results between real-time dynamic data and a library of typical enzymatic hydrolysis reaction curve templates are used to calibrate the maturity value of the enzymatic hydrolysis reaction.

9. The method for detecting enzymatic hydrolysis components in reconstituted nutritional supplements according to claim 8, characterized in that, The method includes: A feedback closed-loop mechanism for detecting components in an enzymatic hydrolysis reaction is constructed, wherein the execution result of a rate inhibition command is used as a new sample. The hidden layer node parameters of the deep belief network are updated using an online incremental learning algorithm; The association weight coefficients of the enzymatic hydrolysis reaction feature vector are regenerated based on the updated deep belief network.

10. A detection system for enzymatic hydrolysis components of reconstituted nutritional supplements, used to implement the detection method for enzymatic hydrolysis components of reconstituted nutritional supplements according to any one of claims 1-9, characterized in that, The system includes: The dynamic data acquisition module is configured to acquire enzyme activity, substrate concentration, and product generation rate values ​​through multi-source sensors. The feature extraction and fusion module is configured to perform multimodal feature fusion on dynamic data and generate enzymatic hydrolysis reaction feature vectors; The reaction state classification module is configured to identify incompletely reacted states, optimal reaction states, and over-reacted states based on the feature vectors of enzymatic hydrolysis reactions. The parameter dynamic adjustment module is configured to adjust the detection parameter thresholds according to different reaction states and generate enzymatic hydrolysis reaction process control instructions. The feedback optimization module is configured to update the parameters of the enzymatic hydrolysis reaction path association model based on the execution results of control commands.

Images