A collagen extraction process state monitoring method and device

Abstract

This invention relates to the fields of artificial intelligence and process status monitoring technology, specifically to a method and apparatus for monitoring the status of collagen extraction processes. The method involves: deploying a sensor network on key equipment in the collagen extraction process to collect process parameters in real time, synchronizing these parameters with laboratory-tested collagen quality indicators, and labeling the status of the collected data; calculating a coupling coefficient function based on the collected process parameters and quality indicator data, and performing coupling normalization on the process parameters to obtain normalized attribute values; constructing and training a deep learning-based process status monitoring model to obtain a trained model; inputting newly collected collagen extraction process data into the trained model for real-time or offline status identification, and providing decision support based on the identified status. This invention enables highly reliable monitoring and intelligent decision support for complex process systems.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and process status monitoring technology, and in particular to a method and apparatus for monitoring the process status of collagen extraction. Background Technology

[0002] Collagen extraction is a crucial step in biomaterial processing, as its purity, molecular weight distribution, and yield directly determine its final application performance. This process involves multiple parameters, including temperature, pH, enzyme concentration, stirring speed, and ultrasonic power. These parameters exhibit complex nonlinear coupling relationships and significantly impact the final quality indicators. Therefore, it is necessary to improve the accuracy of state identification and the robustness of the model in the collagen extraction process through process status monitoring methods, enabling highly reliable monitoring and intelligent decision support for complex process systems.

[0003] Existing process status monitoring technologies commonly employ independent analysis or linear feature extraction methods for single parameters, such as independent normalization, Z-score standardization, or principal component analysis. These methods neglect the interaction effects between different parameters, making it difficult to accurately characterize the true dynamic features of the process. When minor perturbations or anomalies occur during the extraction process, such as instantaneous temperature fluctuations, pH jumps, or changes in enzyme activity, traditional models often fail to identify them in a timely manner, leading to delayed or even misjudged process status assessments. Furthermore, while existing deep learning models demonstrate strong nonlinear modeling capabilities in pattern recognition, their input features and network structures often lack constraints on the rationality of the process. Weight initialization and activation function settings do not consider the coupling characteristics between parameters, easily resulting in problems such as unstable convergence and insufficient generalization. This is particularly true in cases of marginal or progressively anomalous samples, where recognition accuracy significantly decreases. Specifically, existing technologies have the following shortcomings in practical applications:

[0004] Existing technologies often employ independent normalization or Z-score normalization for process parameter preprocessing, neglecting the nonlinear coupling between multiple source parameters, making it difficult for the model to reflect the real process correlation. Conventional feature extraction methods such as PCA and ICA can only reduce dimensionality linearly and cannot express the complex coupling features between parameters such as temperature, pH, and enzyme concentration, resulting in insufficient feature representation. Conventional neural networks use fixed activation functions and random weight initialization, lacking a response to process rationality constraints, easily getting trapped in local optima and leading to low accuracy in identifying abnormal samples. The loss function of existing process status monitoring models only considers classification error and does not introduce a rationality penalty mechanism, resulting in insufficient detection capabilities for marginal and progressive anomalies.

[0005] Therefore, this invention proposes a method and apparatus for monitoring the process status of collagen extraction to solve the above problems. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention develops a method and device for monitoring the state of collagen extraction processes. This invention can improve the accuracy of state identification and the robustness of the model in the collagen extraction process, and achieve highly reliable monitoring and intelligent decision support for complex process systems.

[0007] On the one hand, the technical solution of this invention to solve the technical problem is a method for monitoring the state of collagen extraction process, including the following steps:

[0008] S1. Deploy a sensor network on key equipment in the collagen extraction process to collect process parameters in real time, synchronize them with the quality indicators of collagen tested in the laboratory, form sample data, and label the status of the sample data.

[0009] S2. Calculate the coupling coefficient function based on the collected process parameters and quality index data, perform coupling normalization processing on the sample data, and obtain the normalized attribute values.

[0010] S3. Construct a deep learning-based process condition monitoring model. Input the normalized attribute values ​​into the model, and then proceed through the following steps: extraction of process parameter rationality coupling features, initialization of deep neural network weights, forward propagation and coupled activation of deep neural network, process condition classification and prediction, loss function calculation, back propagation and parameter update, and training of the deep learning-based process condition monitoring model to finally obtain the trained model.

[0011] S4. Input the newly collected collagen extraction process data into the trained model to perform real-time or offline status recognition, and provide decision support based on the recognized status.

[0012] S1 is as follows:

[0013] Key equipment includes reaction vessels, centrifuges, and drying units;

[0014] Process parameters include extraction temperature, pH value, extraction time, enzyme concentration, stirring speed, material concentration, pressure, solvent ratio, ultrasonic power, and centrifugation speed;

[0015] Quality indicators include the purity, yield, molecular weight distribution, ash content, and moisture content of collagen.

[0016] Set the sampling frequency to continuously collect complete production batch data;

[0017] Data labeling is determined jointly by process experts based on quality index thresholds and process curve morphology.

[0018] S2 is as follows:

[0019] The coupling coefficient function is calculated using the Gaussian kernel function. The results of the coupling coefficient function calculation are then used to perform coupling normalization on the collected sample data to obtain the normalized attribute values.

[0020] The coupling coefficient function is specifically obtained by operating on the original attribute values, coupling strength coefficient, conditional mean, and conditional standard deviation of the collected process parameters. The original attribute values ​​are the values ​​of different attributes collected by different sensing devices in the collected process parameters or the quality index data synchronized with the process parameters. The coupling strength coefficient is the Pearson correlation coefficient between different attribute values. The conditional mean and conditional standard deviation are obtained by using the K-means clustering algorithm based on historical data.

[0021] Normalized attribute values ​​are calculated based on the original attribute values, the minimum and maximum values ​​of the attributes, and the coupling coefficient function. The data is then mapped to a unified interval to obtain the normalized attribute values.

[0022] The specific process of extracting coupled features based on the rationality of process parameters is as follows:

[0023] Feature extraction is performed based on the normalized attribute values, nonlinear coupling strength, and linear projection weights to obtain nonlinear coupling features that integrate multi-parameter interactions.

[0024] The specific steps for initializing the weights of a deep neural network are as follows:

[0025] Based on the statistical properties of coupling features and rationality constraints, the initial weights of the neural network are calculated. First, a rationality score is calculated using a rationality indicator function. Then, a coupling correction term matrix is ​​calculated based on the coupling features, the mean of coupling features of all samples, and the rationality score. Finally, based on the mean of coupling features of all samples and the coupling correction term matrix, the influence strength of the correction term is adjusted in conjunction with the rationality constraint coefficients, and the initial weight matrix is ​​calculated.

[0026] The rationality indicator function first calculates the distance between the coupling feature and the ideal process center, and then maps the distance to a rationality score through the Sigmoid function.

[0027] The distance is obtained by calculating the L2 norm of the coupling features and the features of the ideal process center. The features of the ideal process center are the mean of the coupling features of samples labeled as normal in the historical data.

[0028] The specific operations of forward propagation and coupled activation in deep neural networks are as follows:

[0029] In the hidden layer of the neural network, a coupled activation function is used. First, the coupled modulation factor is calculated. Based on the output features of the previous layer and the rationality weights, the modulation factor is generated by the hyperbolic tangent function. Then, the output features of the current layer are calculated based on the output of the baseline activation function and the coupled modulation factor. The output features are calculated layer by layer using the coupled activation function.

[0030] The baseline activation function is the Sigmoid activation function.

[0031] The specific process for classifying and predicting process status is as follows:

[0032] A rationality adjustment mechanism is adopted to dynamically adjust the rationality of the classification layer output. First, the rationality adjustment logits are calculated. Based on the output layer weights, the output features of the last layer, and the rationality indicator function, the rationality adjustment intensity coefficient is broadcast to all categories to increase the original output value of the rational samples, thereby obtaining the adjustment logits vector, which adjusts the model's prediction rationality.

[0033] Then, based on the adjusted logits vector, the predicted probability for each state category is generated by transforming it using the Softmax function.

[0034] The training process of the model is as follows:

[0035] The training process of the deep learning-based process condition monitoring model is an iterative optimization process. First, based on the predicted probability vector, the true label, and the rationality indicator function, the total loss function is calculated by combining the cross-entropy loss and the rationality penalty term.

[0036] Then, a coupled gradient correction term is used to adjust the gradient direction based on the rationality constraint. The coupled gradient is calculated based on the total loss gradient and the coupling modulation factor. Specifically, the regular gradient direction is adjusted through the gradient correction coefficient. The Adam optimizer is used to update the weights with the coupled gradient. Specifically, the weights are updated based on the current weight matrix, the coupled gradient, and the average rationality indicator function to obtain the weight matrix of each layer of the neural network after each iteration.

[0037] Finally, the model parameters are updated through multiple forward and backward propagations. In the forward propagation stage, the training data is normalized and coupled feature extraction is performed before being input into the model to generate prediction results. The total loss is calculated based on the prediction results and the true labels. In the backward propagation stage, the gradient direction is adjusted by using coupled gradient correction terms, and the weights and bias parameters are updated in combination with the Adam optimizer.

[0038] The training process is performed in batches, and the number of iterations is dynamically controlled based on the training loss and the performance of the dataset used for validation.

[0039] The iteration condition is then stopped until the condition is met, at which point the training ends and a trained model is obtained.

[0040] S4 is as follows:

[0041] The system identifies the state based on the predicted probability of each state category, then provides decision support based on the identified state to determine whether to trigger an alarm, and generates operational suggestions by combining historical data and process knowledge.

[0042] On the other hand, the present invention also provides a collagen extraction process status monitoring device for performing a collagen extraction process status monitoring method, comprising:

[0043] Data acquisition and annotation module: performs real-time acquisition, synchronous correlation, and precise annotation of process data;

[0044] Data preprocessing module: Set up a normalization processing unit. The collected data first goes through the normalization processing unit to obtain normalized attribute values ​​that integrate multi-parameter correlation information;

[0045] The model building and training module includes a coupled feature extraction submodule, a network weight initialization submodule, a forward propagation and coupled activation submodule, a state classification prediction submodule, and a model training optimization submodule. It receives the normalized attribute values ​​output by the data preprocessing module, passes them through each module in sequence, and outputs the trained model.

[0046] Condition monitoring and decision support module: Based on the trained model output by the model building and training module, it performs real-time or offline identification of process conditions and provides decision support suggestions.

[0047] The effects described in the invention are merely those of the embodiments, and not all the effects of the invention. The above technical solutions have the following advantages or beneficial effects:

[0048] This invention employs a Gaussian kernel coupled normalization mechanism to integrate the nonlinear correlation between process parameters and quality indicators into the normalization process, dynamically adjusting the weights between attributes. This solves the problem that conventional independent normalization cannot capture parameter interaction effects. Based on a coupled feature extraction method, this invention combines linear projection with nonlinear coupling strength to achieve the fusion of rationality features among multiple parameters, overcoming the limitation of conventional principal component analysis, which can only handle linear relationships. By using a rationality indicator function and coupled activation mechanism in the deep neural network, the model can dynamically adjust the activation strength and learning rate according to process rationality, thereby significantly improving the sensitivity and robustness of abnormal state identification. Furthermore, by embedding rationality constraints in the loss function and parameter updates, this invention achieves adaptive penalty and gradient correction for abnormal samples, enhancing the model's ability to identify and stabilize marginal process anomalies. Attached Figure Description

[0049] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0050] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0051] Figure 2 This is a comparison chart showing the impact of the normalization method of this invention on the accuracy of condition monitoring compared to existing normalization methods.

[0052] Figure 3 This is a scatter plot of the feature space distribution using conventional methods.

[0053] Figure 4 This is a scatter plot showing the spatial distribution of features of the method of the present invention.

[0054] Figure 5 This is a comparison chart showing the ability of the method of this invention and existing methods to detect abnormal states in collagen processing. Detailed Implementation

[0055] To clearly illustrate the technical features of this solution, the invention will be described in detail below through specific implementation methods and in conjunction with the accompanying drawings.

[0056] Example 1

[0057] A method for monitoring the processing status of collagen extraction includes the following steps:

[0058] S1. Deploy a sensor network on key equipment in the collagen extraction process to collect process parameters in real time, synchronize them with the quality indicators of collagen tested in the laboratory, form sample data, and label the status of the sample data.

[0059] S2. Calculate the coupling coefficient function based on the collected process parameters and quality index data, perform coupling normalization processing on the sample data, and obtain the normalized attribute values.

[0060] S3. Construct a deep learning-based process condition monitoring model. Input the normalized attribute values ​​into the model, and then proceed through the following steps: extraction of process parameter rationality coupling features, initialization of deep neural network weights, forward propagation and coupled activation of deep neural network, process condition classification and prediction, loss function calculation, back propagation and parameter update, and training of the deep learning-based process condition monitoring model to finally obtain the trained model.

[0061] S4. Input the newly collected collagen extraction process data into the trained model to perform real-time or offline status recognition, and provide decision support based on the recognized status.

[0062] In a specific implementation, step S1 is as follows:

[0063] By deploying a sensor network, process parameters of key equipment such as reactors, centrifuges, and drying units are collected in real time, including parameters such as extraction temperature, pH value, extraction time, enzyme concentration, stirring speed, material concentration, pressure, solvent ratio, ultrasonic power, and centrifugation speed. The sampling frequency is set to 1Hz to continuously collect complete production batch data.

[0064] Simultaneously, five quality indicators of collagen—purity, yield, molecular weight distribution, ash content, and moisture content—were tested in the laboratory to form a sample that was aligned with the process parameters in sequence.

[0065] Data labeling is jointly determined by three process experts based on quality index thresholds and process curve morphology: when all quality indicators meet GB / T standards and the parameter curve is smooth without abrupt changes, it is labeled as "normal"; when purity is below 90%, molecular weight distribution exceeds the 10-100kDa range, ash content exceeds the standard, or parameters change abruptly (such as temperature fluctuation ±5℃ / min, pH jump ±1.5), it is labeled as "abnormal".

[0066] In the specific implementation process, a dataset containing 500 production batches and 15,000 samples was constructed. Each sample contains 15-dimensional attributes (10-dimensional process parameters and 5-dimensional quality indicators) and corresponding "normal / abnormal" binary labels.

[0067] In a specific implementation, step S2 is as follows:

[0068] Due to the multi-source heterogeneity, large scale differences, and nonlinear coupling between parameters in collagen extraction process parameters and quality index data, conventional processing methods, which use independent normalization, easily overlook the coupling relationship between parameters. This leads to the model's inability to effectively capture the interaction effect between parameters, thus affecting the accuracy of state recognition. Therefore, this invention uses a coupling coefficient to dynamically adjust the normalization process to reflect the reasonable correlation between parameters. The specific steps are as follows:

[0069] 1) Calculation of coupling coefficient function

[0070] The coupling coefficient function is calculated using a Gaussian kernel function, specifically based on the original attribute values, coupling strength coefficient, conditional mean, and conditional standard deviation. This dynamically quantifies the reasonable correlation between attributes, and then dynamically adjusts the normalization process to ensure that the normalization result accurately reflects the interaction between parameters, thereby improving the model's ability to perceive the reasonableness of parameters. This is expressed as:

[0071]

[0072] In the formula, Indicates the first The coupling coefficient function of each attribute, in In the case of considering the inclusion of the first All other attributes, including the first attribute, affect the first... The influence of each attribute is used to dynamically adjust the normalization result, representing the interaction of parameter rationality;

[0073] Indicates the first The first sample Each original attribute value is the original process parameter or quality indicator data, which directly reflects the actual measured value of the attribute.

[0074] For sample index;

[0075] For attribute indexing;

[0076] To distinguish it from the attribute index of j;

[0077] Indicates the first In the sample except the first All attribute values ​​other than the specified attribute;

[0078] Indicates the first The first attribute and the first The coupling strength coefficient of the attribute is learned through historical data and represents the . The first attribute and the first The degree of correlation between the attributes;

[0079] Indicates the first The attribute in the first The conditional mean under the influence of each attribute is initialized using a clustering algorithm;

[0080] Indicates the first The attribute in the first The conditional standard deviation under the influence of each attribute is initialized using a clustering algorithm;

[0081] This represents the natural exponential function.

[0082] In practical implementation, the coupling strength coefficient is learned based on a historical dataset containing process parameters and quality indicators. First, the Pearson correlation coefficient between the j-th and k-th attributes is calculated as an initial value. Then, during the neural network training phase... As trainable parameters, they are optimized using a gradient descent optimization algorithm based on backpropagation, with the goal of minimizing the state classification loss function. During training, Update based on loss gradient to adaptively capture nonlinear correlations between attributes and ensure that coupling strength dynamically reflects the actual process interaction mode;

[0083] Conditional mean The historical data was grouped using the K-means clustering algorithm. The number of clusters was determined using the elbow method. First, all historical samples were clustered into K clusters based on all attributes. Then, for each cluster, the mean of the j-th attribute within that cluster was calculated, defining... For the first The attribute in the first The conditional mean of each attribute is used to assign samples to the nearest cluster based on the value of the k-th attribute. The nearest cluster is calculated based on the Euclidean distance between the k-th attribute and the cluster center, and the conditional mean of this cluster is used as the basis for the assignment. Similarly, conditional standard deviation After K-means clustering, the standard deviation of the j-th attribute is calculated for each cluster, and the definition is... For the first The attribute in the first The conditional standard deviation under the influence of the first attribute, based on the first attribute The values ​​of each attribute are assigned to the nearest cluster, and the conditional standard deviation of that cluster is used as the basis for determining the value of the sample. .

[0084] 2) Coupled Normalization Calculation

[0085] Normalized attribute values ​​are calculated based on the original attribute values, minimum and maximum attribute values, and coupling coefficient functions. This maps the data to a unified interval, thereby eliminating dimensional differences between process parameters and quality indicators. Simultaneously, it strengthens the rational constraints between parameters, improving the robustness and accuracy of state identification. This is expressed as:

[0086]

[0087] In the formula, Indicates the first The first sample A normalized attribute value, representing standardized data after coupling adjustment, is used to eliminate dimensional differences and enhance parameter correlation;

[0088] Indicates the first The minimum value of each attribute across all samples;

[0089] Indicates the first The maximum value of each attribute across all samples;

[0090] Represents the first of all samples A set of attribute values.

[0091] It should be noted that the coupling coefficient function Through Gaussian kernel Dynamically quantify the reasonable correlation between attributes, such as the reasonable range of pH under abnormal temperature conditions, and amplify this coupling during normalization, not only mapping the data to... Eliminating dimensional differences within intervals can also strengthen the rationality constraints between parameters. In highly coupled scenarios such as sudden temperature changes, the model can more sensitively identify edge anomalies such as protein denaturation initiation points, improving the robustness of state recognition; learning from historical data... and cluster initialization , It can ensure the generalization of the model, transforming conventional normalization from static scaling to dynamic coupled perception, so that subsequent feature extraction can more accurately encode the rationality of the process. This is something that conventional independent normalization cannot achieve, because conventional methods ignore the nonlinear nature of the synergistic effect between parameters.

[0092] In a specific implementation, step S3 is as follows:

[0093] S3.1 Extraction of Coupled Features for Rationality of Process Parameters:

[0094] There are strong coupling relationships among the parameters of collagen extraction process. For example, temperature and pH jointly affect the reaction rate. Conventional feature extraction methods, such as principal component analysis, only perform linear dimensionality reduction and cannot capture nonlinear coupling features, resulting in insufficient feature expression and affecting the accuracy of state recognition. Therefore, this invention calculates coupling features based on normalized attribute values, nonlinear coupling strength, and linear projection weights to characterize the nonlinear features of multi-parameter interactions. The calculation formula is as follows:

[0095]

[0096] In the formula, Indicates the first The first sample A coupling feature characterizes the nonlinear characteristics of multi-parameter interaction;

[0097] Indicates the first The first sample A normalized attribute value;

[0098] Indicates the total number of attributes, such as The total dimension characterizing process parameters and quality indicators;

[0099] The parameter represents the non-linear coupling strength between the j-th attribute and the k-th attribute. It is a trainable parameter and is initialized based on the attribute correlation matrix.

[0100] The linear projection weights of the j-th and k-th attributes are trainable parameters that are randomly initialized and updated during training. They are used to linearly project the normalized attribute values ​​to extract potential coupling features.

[0101] This represents the hyperbolic tangent function, used to implement nonlinear transformations and enhance the ability to express features.

[0102] In the specific implementation process, there are a total of 15 attributes for process parameters and quality indicators, including: extraction temperature (in °C), pH value (dimensionless), extraction time (in min), enzyme concentration (in g / L), stirring speed (in rpm), material concentration (in %), pressure (in MPa), solvent ratio (in v / v), ultrasonic power (in W), centrifugation speed (in rpm), collagen purity (in %), yield (in g / L), molecular weight distribution (in kDa), ash content (in %), and moisture content (in %).

[0103] pass Dynamic fusion of multiple parameters and interaction Learn the coupling strength between attributes. Combining linear projection with hyperbolic tangent nonlinear activation generates higher-order features, making the features... Not only retain its own attribute information It also encodes the reasonable contributions of other attributes, such as the correction of temperature characteristics when pH is abnormal, and the dynamic characteristics of the reasonableness of process parameters, which can enhance the model's ability to perceive complex coupling modes.

[0104] S3.2, Initialization of Deep Neural Network Weights:

[0105] Faced with high-dimensional coupling features, conventional deep neural networks use random initialization of weights, which can easily lead to unstable gradients or slow convergence, making it difficult to adapt to the complex patterns of reasonable collagen extraction process parameters. Therefore, this invention calculates the initial weights of the neural network based on the statistical characteristics and reasonable constraints of the coupling features to improve training efficiency and robustness. The specific steps are as follows:

[0106] 1) Calculation of the rationality indicator function

[0107] Based on the coupled feature vector and the feature vector of the ideal process center, the distance is mapped to a reasonableness score through the Sigmoid function, quantifying the reasonableness of the sample parameters and providing a reasonable constraint basis for weight initialization, expressed as:

[0108]

[0109] In the formula, Indicates the first The rationality indicator function of the nth sample, characterizing the... The reasonableness of the parameters for each sample, with a range of values. A larger value indicates a higher degree of rationality;

[0110] Indicates the first The coupled feature vector of each sample is composed of... All coupling features, including , dimension ;

[0111] Indicates the first The first sample One coupling feature, Indicates the first The first coupling feature of each sample, Indicates the first The second coupling feature of each sample Indicates the first The Mth coupling feature of a sample;

[0112] Indicates the first The distance between the coupled feature vector of each sample and the ideal process center is calculated using the L2 norm, which characterizes the degree to which the sample features deviate from the ideal state.

[0113] In specific implementation, the first Distance between the coupled feature vector of each sample and the ideal process center It is obtained by calculating the L2 norm, specifically by calculating the first... Coupled feature vectors of each sample eigenvectors of the ideal process center The L2 norm realization, the eigenvector of the ideal process center The mean of the coupled features of samples with normal labels in historical data, and the distance It quantifies the degree to which a sample deviates from the ideal state; the larger the value, the lower the rationality.

[0114] 2) Calculation of Coupling Correction Term

[0115] The coupling correction term matrix is ​​calculated based on the coupled feature vector, the mean vector of coupled features of all samples, and the rationality indicator function. While weighting feature bias, unreasonable samples are penalized, thereby strengthening the constraint on abnormal parameter combinations during weight initialization, as expressed below:

[0116]

[0117] In the formula, This represents the coupling correction term matrix, with the same dimensions as the neural network weights, used to generate reasonable constraints during weight initialization;

[0118] This represents the mean vector of coupled features for all samples, calculated as follows: The overall distribution center characterizing coupling features;

[0119] This represents the total number of training samples.

[0120] 3) Neural network weight initialization

[0121] Based on the mean vector of all sample coupled features and the coupled correction term matrix, the influence strength of the correction term is adjusted by combining the rationality constraint coefficient. Then, the initial weight matrix is ​​calculated through these key terms to ensure that the weight distribution is biased towards the rational feature pattern, thereby improving the model's convergence speed and robustness. This is expressed as:

[0122]

[0123] In the formula, This represents the initial weight matrix of the first layer of the neural network, with dimension 1. By combining the distribution characteristics of coupling features, we can ensure that the weights adapt to the reasonableness of the parameters.

[0124] The input feature dimension of the neural network;

[0125] This represents the number of hidden units in the first layer of the neural network.

[0126] This represents the rationality constraint coefficient, which is preferably set to 0.2 to control the influence intensity of the coupling correction term.

[0127] Reasonableness indicator function Distance is calculated using the Sigmoid function. Mapped to Interval rationality scoring, quantifying sample rationality, and coupling correction term matrix. Then utilize Weighted eigenvalue bias Punish unreasonable samples, high The sample contribution is greater. By incorporating the rationality of the process into the weight initialization process, the initial weights are biased towards the rational feature distribution. At the same time, the model focuses on the rationality pattern in the early stage of training, which accelerates convergence and improves robustness.

[0128] S3.3, Forward Propagation and Coupled Activation in Deep Neural Networks:

[0129] Classifying the state of collagen extraction processes requires capturing deep nonlinear relationships. Conventional activation functions such as ReLU cannot adapt to dynamic changes in parameter coupling, resulting in insufficient model response to the rationality of process parameters. This invention employs coupled activation functions in the hidden layers of neural networks, which can dynamically adjust the activation threshold based on parameter rationality. The specific steps are as follows:

[0130] 1) Calculation of Coupling Modulation Factor

[0131] Based on the output features and rationality weights of the previous layer, a modulation factor is generated through the hyperbolic tangent function, and then the rationality adjustment amount is dynamically calculated to enhance the activation function's ability to respond to the dynamic coupling of process parameters, expressed as:

[0132]

[0133] In the formula, Indicates the neural network's first... The coupling modulation factor of the layer is a scalar value, based on the first layer. The feature calculation of the layer is rationally adjusted to dynamically adjust the output strength of the activation function;

[0134] This represents the output vector of the (l-1)th layer of the neural network, when hour, The corresponding i-th sample is equivalent to the i-th sample. Coupled feature vectors of each sample surface;

[0135] This represents the q-th element of the output vector of the (l-1)-th layer of the neural network;

[0136] For the layer index of the neural network;

[0137] For feature index;

[0138] This represents the modulation intensity hyperparameter, which is preferably set to 0.2 to control the degree of influence of coupling modulation.

[0139] express The dimension, i.e., the dimensionality of the neural network. The number of hidden units in the layer;

[0140] Indicates the first Each rationality weight is a trainable parameter, initialized based on process knowledge, used to weight the contribution of different features to rationality adjustment.

[0141] 2) Coupled activation calculation

[0142] The output features of the current layer are calculated based on the output of the baseline activation function and the coupling modulation factor, realizing dynamic activation based on parameter rationality. This suppresses the propagation of anomalous features during the reaction process, thereby improving the model's accuracy in identifying edge anomalous samples, as expressed below:

[0143]

[0144] In the formula, This represents the output vector of the l-th layer of the neural network, which characterizes the feature representation after coupled activation and serves as the input or final output of the next layer.

[0145] This represents the Sigmoid activation function, which serves as a baseline activation function for general nonlinear transformations, mapping the input to a specific range.

[0146] Indicates the neural network's first... The weight matrix of a layer is a trainable parameter used to linearly transform the output of the previous layer;

[0147] Indicates the neural network's first... The layer's bias vector is a trainable parameter used to adjust the threshold of the activation function.

[0148] It should be noted that the coupling modulation factor Using the hyperbolic tangent function and rational weights The system dynamically calculates the appropriate adjustment amount and modulates the output of the baseline activation function to account for the time-varying characteristics of process rationality, such as the synergistic effect of parameters during the reaction process. Based on the features of the previous layer Generate scalar modulation factor, when When indicating an unreasonable state, such as In this case, the negative value of the hyperbolic tangent function output decreases. This suppresses activation intensity, thereby making the model more sensitive to anomalous coupling patterns.

[0149] S3.4, Process Status Classification Prediction:

[0150] This invention employs a rationality adjustment mechanism to dynamically adjust the output of the classification layer based on the rationality of the parameters. The specific steps are as follows:

[0151] 1) Adjust the logits calculation reasonably

[0152] Based on the output layer weights, the last layer output features, and the rationality indicator function, and combined with the rationality adjustment strength coefficient broadcast to all categories, the original output values ​​of rational samples are increased, making the model more inclined to predict rational states, thereby strengthening the parameter rationality constraints in classification decisions, as expressed as:

[0153]

[0154] In the formula, Indicates the first The adjusted logits vector for each sample, with dimension [ ]. ;

[0155] For the number of categories, such as normal, abnormal;

[0156] The weight matrix of the output layer represents trainable parameters with dimension 1. ;

[0157] The dimension of the last hidden layer in the neural network;

[0158] This represents the output vector of the last layer of the neural network, with dimension 1. ;

[0159] The bias vector of the output layer is a trainable parameter with dimension . ;

[0160] The hyperparameter representing the rationality adjustment intensity is preferably set to 0.1, and is used to control the impact of the rationality indicator function on logits.

[0161] Indicates the first The rationality indicator function for each sample is based on the coupled feature vector. The calculation takes values ​​in the range (0,1), with larger values ​​indicating higher reasonableness.

[0162] This represents a vector consisting entirely of 1s, with dimension . This is used to broadcast the reasonableness score to all categories.

[0163] It should be noted that the logits vector, also known as the log-odds vector, is used in classification models. The model's raw output can be interpreted as the strength of "evidence" for each class. Calculating the exponent of this evidence and then normalizing it yields the probability; therefore, the raw output is the log-odds of that probability. By employing... When the sample is highly reasonable, the value of all logits is increased, thereby indirectly affecting the probability distribution and making the model more inclined to predict reasonable states, thus strengthening the role of parameter reasonableness constraints in classification decision-making.

[0164] 2) Calculation of prediction probability

[0165] The operation is based on adjusting the logits vector, and the predicted probability of each class is generated through the Softmax function to ensure that the probability distribution is normalized. Reasonable adjustments are then made to improve the classification robustness of marginal samples, as shown below:

[0166]

[0167] In the formula, Indicates the neural network for the first... The predicted probability vector of each sample has a dimension of . Each element represents the probability of the corresponding category;

[0168] For the Softmax function;

[0169] Indicates the first The adjusted logits vector for each sample The One element;

[0170] For category indexing.

[0171] It should be noted that the Softmax function ensures that the sum of the elements of the predicted probability vector is 1, and that each element is between 0 and 1, thereby completing the classification prediction. The reasonable adjustment mechanism makes the model more robust on edge samples with abnormal parameter coupling but not completely invalid, which can improve the accuracy of state recognition.

[0172] S3.5 Loss Function Calculation:

[0173] The classification of collagen extraction process states needs to distinguish between normal and abnormal states, such as protein denaturation or excessive impurities. Conventional cross-entropy loss functions ignore parameter rationality constraints, easily leading to inaccurate classification of marginal samples and failing to effectively utilize the coupling relationship between process parameters to improve model robustness. This invention calculates the total loss function based on the predicted probability vector, true label, and rationality indicator function, combined with cross-entropy loss and a rationality penalty term. This enhances the penalty for unreasonable samples, increases the loss value when parameter coupling is abnormal, and thus improves the model's classification accuracy for marginal samples. The calculation formula is as follows:

[0174]

[0175] In the formula, This represents the total loss function, used to optimize neural network parameters, taking into account classification error and reasonableness constraints.

[0176] Cross-entropy loss function measures the difference between the predicted probability and the true label;

[0177] For the first The true labels of each sample are in one-hot encoded format;

[0178] For the neural network to the first The predicted probability vector of each sample;

[0179] The optimal setting for the weighting hyperparameter of the reasonableness loss is 0.2, which controls the strength of the reasonableness penalty term.

[0180] The square of the L2 norm is used as an error measure in the rationality penalty term.

[0181] S3.6 Backpropagation and parameter update:

[0182] This invention employs a coupled gradient correction term to adjust the gradient direction based on rationality constraints, and uses the Adam optimizer to update the weights using coupled gradients. The specific steps are as follows:

[0183] 1) Coupled gradient calculation

[0184] Based on the total loss gradient and coupling modulation factor, the conventional gradient direction is adjusted by the gradient correction coefficient, and then reasonable constraints are superimposed on the gradient to make the update direction more consistent with the coupling characteristics of process parameters, thus avoiding the model from getting trapped in local optima. This is expressed as follows:

[0185]

[0186] In the formula, Indicates the neural network's first... The coupled gradient of the layer weight matrix is ​​used as the input gradient for weight update, combined with the regular gradient and rationality constraints;

[0187] Represents the total loss function For the neural network Layer weight matrix The gradient;

[0188] This represents the gradient correction coefficient, which is preferably set to 0.1, and controls the weight of the gradient term of the coupling modulation factor.

[0189] The symbol represents the partial derivative.

[0190] 2) Neural network parameter update

[0191] Based on the current weight matrix, coupled gradient, and average rationality indicator function, the weights are updated using the Adam optimizer. This accelerates convergence when rationality is high and refines the update when rationality is low, making parameter optimization more consistent with the process rationality pattern, thus improving training stability and state classification accuracy. This can be expressed as:

[0192]

[0193] In the formula, Indicates the first The neural network in the nth iteration Layer weight matrix;

[0194] Indicates the first The neural network in the nth iteration Layer weight matrix;

[0195] This represents the base learning rate, which is preferably set to 0.001.

[0196] This represents the reasonableness learning rate adjustment factor, preferably set to 0.05, used to dynamically adjust the learning rate based on the reasonableness score;

[0197] The average reasonableness indicator function for the current batch of samples is calculated as follows: ;

[0198] This is the number of training samples in a batch; the default value is 16.

[0199] This indicates that the Adam optimizer is based on the neural network's first... Coupled gradient of layer weight matrix The deviation correction value of the first moment estimate is calculated in accordance with the Adam standard procedure;

[0200] This indicates that the Adam optimizer is based on the neural network's first... Coupled gradient of layer weight matrix The deviation correction value of the calculated second-order moment estimate is calculated in accordance with the Adam standard procedure;

[0201] It is a small constant used for numerical stability, and is preferably set to... .

[0202] It should be noted that when the average of a batch of samples has high reasonableness, Increasing the learning rate accelerates model convergence within a reasonable region. When the reasonableness is low, decreasing the learning rate refines the updates to avoid oscillations, making parameter updates more consistent with the reasonableness pattern of the process, thereby improving training stability and state classification accuracy.

[0203] S3.7 Training of a Deep Learning-Based Process Condition Monitoring Model:

[0204] The training process of the deep learning-based process status monitoring model is an iterative optimization process. The model parameters are updated through multiple forward and backward propagations, enabling the model to accurately identify the collagen extraction process status.

[0205] At the start of training, the constructed training dataset is used as input. First, coupling normalization is performed to eliminate dimensional differences and enhance the correlation between parameters. Then, nonlinear coupling features are generated through the process parameter rationality coupling feature extraction module, which serves as the input to the neural network.

[0206] During training, the model performs forward propagation, sequentially going through deep neural network weight initialization, coupled activation function calculation, and process state classification prediction to generate a predicted probability vector for each sample.

[0207] Subsequently, the total loss function is calculated based on the prediction results and the true labels. This function combines cross-entropy loss and a reasonableness penalty term to strengthen the constraint on unreasonable samples.

[0208] During the backpropagation phase, a coupled gradient correction term is used to adjust the gradient direction, and the Adam optimizer is used to update the weights and bias parameters. The learning rate is dynamically adjusted according to the average reasonableness indicator function of the batch samples to accelerate the convergence of reasonable regions and refine the update of abnormal regions.

[0209] The training process is performed in batches, with a default batch size of 16. The number of iterations is dynamically controlled based on the training loss and validation set performance.

[0210] The criteria for stopping iterations include two main cases: First, both the training loss and the validation loss converge, meaning that the change in loss is less than a preset threshold within several consecutive iteration cycles, indicating that the model has reached a stable state; second, the maximum number of iterations is reached to prevent overfitting.

[0211] Meanwhile, the accuracy of the validation set is monitored during training, and the training is stopped early when the accuracy no longer improves, in order to ensure the model's generalization ability.

[0212] In a specific implementation, step S4 is as follows:

[0213] After the model training is completed, it is applied to newly collected collagen extraction process data to achieve real-time or offline status recognition and decision support;

[0214] For new data, the same preprocessing as in the training phase is performed first, including data cleaning and coupling normalization. The coupling strength coefficient learned from historical data and the conditional mean and conditional standard deviation of cluster initialization are used to map the new process parameters and quality index data to a unified interval and strengthen the reasonable correlation between parameters.

[0215] The preprocessed data is input into a trained deep learning-based process condition monitoring model. The model calculates coupled features through forward propagation and undergoes multi-layer processing by a deep neural network, including dynamic adjustment of feature outputs by coupled activation functions. Finally, a state classification prediction probability is generated through a reasonable adjustment mechanism.

[0216] The status monitoring results are output in the form of probabilities, such as the probability of normal or abnormal status. Operators can determine the current process status based on the probability threshold (such as setting 0.5 as the classification boundary).

[0217] The decision support component is based on the status monitoring results. When an abnormal state is identified, the system automatically triggers an alarm and generates suggestions by combining historical data and process knowledge. For example, it may adjust the temperature or pH value to restore reasonable parameter coupling, or recommend a shutdown for inspection to avoid quality loss. For normal states, it outputs optimization suggestions to improve efficiency, such as fine-tuning the stirring speed or extraction time.

[0218] Example 2

[0219] A collagen extraction process status monitoring device is used to perform a collagen extraction process status monitoring method, comprising:

[0220] Data acquisition and annotation module: performs real-time acquisition, synchronous correlation, and precise annotation of process data;

[0221] Data preprocessing module: Set up a normalization processing unit. The collected data first goes through the normalization processing unit to obtain normalized attribute values ​​that integrate multi-parameter correlation information;

[0222] The model building and training module includes a coupled feature extraction submodule, a network weight initialization submodule, a forward propagation and coupled activation submodule, a state classification prediction submodule, and a model training optimization submodule. It receives the normalized attribute values ​​output by the data preprocessing module, passes them through each module in sequence, and outputs the trained model.

[0223] Condition monitoring and decision support module: Based on the trained model output by the model building and training module, it performs real-time or offline identification of process conditions and provides decision support suggestions.

[0224] Example 3

[0225] like Figure 2 As shown, the impact of different normalization methods on the accuracy of collagen extraction process status monitoring was evaluated. The experiment compared the proposed coupled normalization method with existing normalization methods, namely independent normalization, Z-score normalization, and Min-Max normalization. In the experimental configuration, the number of training batches was gradually increased from 50 to 500. The horizontal axis represents the number of training batches, and the vertical axis represents the status monitoring accuracy. Higher accuracy indicates better model performance in identifying process status. Figure 2 The experimental results show that the accuracy of all methods increases with the number of training batches, consistent with the basic rules of machine learning model training. However, the coupled normalization method proposed in this invention significantly outperforms the other three conventional normalization methods at all training stages. In the early stages of training, the accuracy of the coupled normalization method is significantly higher than other methods, and this advantage further expands with the increase in training batches. Particularly noteworthy is that the accuracy improvement of the coupled normalization method is more significant when the number of training batches reaches a high level, indicating that this method can better utilize large amounts of training data to improve model performance. Among the conventional normalization methods, Min-Max normalization performs relatively well, followed by independent normalization and Z-score normalization, indicating that conventional normalization methods only process each process parameter independently, ignoring the complex coupling relationships between parameters. In contrast, this invention dynamically adjusts the normalization process through coupling coefficients, fully considering the interactions between parameters such as temperature and pH, enzyme concentration and extraction time, making the normalized data more reflective of the true process state characteristics, thereby significantly improving the accuracy of state monitoring.

[0226] Example 4

[0227] like Figure 3 and Figure 4 As shown, the differences in feature learning capabilities among different methods are illustrated by visualizing the feature space distribution. The experimental results are presented in the form of a scatter plot. Figure 3 , Figure 4 The x and y axes represent the two main feature dimensions after feature extraction, with green dots indicating samples under normal process conditions and red dots indicating samples under abnormal process conditions. The experiment compared the feature distribution of conventional methods and the method of this invention on the same dataset. Figure 3In the feature space distribution map of the conventional method shown, a significant overlap between normal and abnormal state samples is clearly observed. This feature aliasing makes it difficult for the classifier to establish clear decision boundaries, thus affecting the accuracy of state monitoring. Particularly in the central region of the feature space, normal and abnormal samples are almost completely intertwined. In contrast, Figure 4 The feature space distribution of the method of the present invention exhibits a distinctly different pattern. Normal state samples cluster in specific regions of the feature space, forming relatively compact distribution clusters, while abnormal state samples deviate significantly from the main clustering areas and are distributed at the edges of the feature space. The separation between the two categories is significantly improved, and the overlapping area is greatly reduced. Experimental results show that the coupled feature extraction mechanism in this invention can fully exploit the nonlinear coupling relationship between process parameters and generate more discriminative feature representations. Furthermore, the concentration of feature distribution in the figure also reflects the feature learning effect of different methods. The tight clustering of normal samples in the method of this invention indicates that the method can effectively learn the essential features of normal process states, while the dispersed distribution of abnormal samples demonstrates its ability to sensitively capture various abnormal patterns.

[0228] Example 5

[0229] like Figure 5 As shown, the abnormal state detection capabilities are compared, and the sensitivity of different methods to progressive process abnormalities is tested. The detection rate change trends of the present invention and two conventional methods are compared under nine abnormality levels (10%-90%). Abnormalities are simulated by parameter perturbation (temperature ±5℃ / min jump), and nine test sets (1000 samples per set) are generated according to the proportion of abnormal samples. Figure 5 In the figure, the horizontal axis represents the degree of anomaly (0.1-0.9) in decimal percentage form, and the vertical axis represents the anomaly detection rate (0.5-1.05) in decimal percentage form. The blue dotted line represents the method of this invention, the orange square line represents the method of independent normalization combined with convolutional neural network, and the red diamond line represents the method of conventional principal component analysis combined with support vector machine. The experimental results show that the detection rate of this invention is significantly higher than the other two methods, proving that the coupling modulation factor is sensitive to edge anomalies and showing that the parameter coupling sensing mechanism has the most significant advantage in the early stage of anomalies.

[0230] Although the specific embodiments of the invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the invention. Based on the technical solutions of the invention, various modifications or variations that can be made by those skilled in the art without creative effort are still within the scope of protection of the invention.

Claims

1. A method for monitoring the processing status of collagen extraction, characterized in that, Includes the following steps: S1. Deploy a sensor network on key equipment in the collagen extraction process to collect process parameters in real time, synchronize them with the quality indicators of collagen tested in the laboratory, form sample data, and label the status of the sample data. S2. Calculate the coupling coefficient function based on the collected process parameters and quality index data, perform coupling normalization processing on the sample data, and obtain the normalized attribute values. S3. Construct a deep learning-based process condition monitoring model. Input the normalized attribute values ​​into the model, and then proceed through the following steps: extraction of process parameter rationality coupling features, initialization of deep neural network weights, forward propagation and coupled activation of deep neural network, process condition classification and prediction, loss function calculation, back propagation and parameter update, and training of the deep learning-based process condition monitoring model to finally obtain the trained model. The specific steps for initializing the weights of a deep neural network are as follows: Based on the statistical properties of coupling features and rationality constraints, the initial weights of the neural network are calculated. First, a rationality score is calculated using a rationality indicator function. Then, a coupling correction term matrix is ​​calculated based on the coupling features, the mean of coupling features of all samples, and the rationality score. Finally, based on the mean of coupling features of all samples and the coupling correction term matrix, the influence strength of the correction term is adjusted in conjunction with the rationality constraint coefficients, and the initial weight matrix is ​​calculated. The rationality indicator function first calculates the distance between the coupling feature and the ideal process center, and then maps the distance to a rationality score through the Sigmoid function. The distance is obtained by calculating the L2 norm of the coupling features and the features of the ideal process center. The features of the ideal process center are the mean of the coupling features of samples with normal labels in the historical data. The specific operations of forward propagation and coupled activation in deep neural networks are as follows: In the hidden layer of the neural network, a coupled activation function is used. First, the coupled modulation factor is calculated. Based on the output features of the previous layer and the rationality weights, the modulation factor is generated by the hyperbolic tangent function. Then, the output features of the current layer are calculated based on the output of the baseline activation function and the coupled modulation factor. The output features are calculated layer by layer using the coupled activation function. The baseline activation function is the Sigmoid activation function. The specific process for classifying and predicting process status is as follows: A rationality adjustment mechanism is adopted to dynamically adjust the rationality of the classification layer output. First, the rationality adjustment logits are calculated. Based on the output layer weights, the output features of the last layer, and the rationality indicator function, the rationality adjustment intensity coefficient is broadcast to all categories to increase the original output value of the rational samples, thereby obtaining the adjustment logits vector, which adjusts the model's prediction rationality. Then, based on the adjusted logits vector, the predicted probability for each state category is generated by transforming it using the Softmax function; The training process of the model is as follows: The training process of the deep learning-based process condition monitoring model is an iterative optimization process. First, based on the predicted probability vector, the true label, and the rationality indicator function, the total loss function is calculated by combining the cross-entropy loss and the rationality penalty term. Then, a coupled gradient correction term is used to adjust the gradient direction based on the rationality constraint. The coupled gradient is calculated based on the total loss gradient and the coupling modulation factor. Specifically, the regular gradient direction is adjusted through the gradient correction coefficient. The Adam optimizer is used to update the weights with the coupled gradient. Specifically, the weights are updated based on the current weight matrix, the coupled gradient, and the average rationality indicator function to obtain the weight matrix of each layer of the neural network after each iteration. Finally, the model parameters are updated through multiple forward and backward propagations. In the forward propagation stage, the training data is normalized and coupled feature extraction is performed before being input into the model to generate prediction results. The total loss is calculated based on the prediction results and the true labels. In the backward propagation stage, the gradient direction is adjusted by using coupled gradient correction terms, and the weights and bias parameters are updated in combination with the Adam optimizer. The training process is performed in batches, and the number of iterations is dynamically controlled based on the training loss and the performance of the dataset used for validation. The iteration condition is then stopped until the condition is met, at which point the training ends and the trained model is obtained. S4. Input the newly collected collagen extraction process data into the trained model to perform real-time or offline status recognition, and provide decision support based on the recognized status.

2. The method for monitoring the state of collagen extraction process according to claim 1, characterized in that, S2 is as follows: The coupling coefficient function is calculated using the Gaussian kernel function. The results of the coupling coefficient function calculation are then used to perform coupling normalization on the collected sample data to obtain the normalized attribute values. The coupling coefficient function is specifically obtained by operating on the original attribute values, coupling strength coefficient, conditional mean, and conditional standard deviation of the collected process parameters. The original attribute values ​​are the values ​​of different attributes collected by different sensing devices in the collected process parameters or the quality index data synchronized with the process parameters. The coupling strength coefficient is the Pearson correlation coefficient between different attribute values. The conditional mean and conditional standard deviation are obtained by using the K-means clustering algorithm based on historical data. Normalized attribute values ​​are calculated based on the original attribute values, the minimum and maximum values ​​of the attributes, and the coupling coefficient function. The data is then mapped to a unified interval to obtain the normalized attribute values.

3. The method for monitoring the state of collagen extraction process according to claim 1, characterized in that, The specific process of extracting coupled features based on the rationality of process parameters is as follows: Feature extraction is performed based on the normalized attribute values, nonlinear coupling strength, and linear projection weights to obtain nonlinear coupling features that integrate multi-parameter interactions.

4. The method for monitoring the state of collagen extraction process according to claim 1, characterized in that, S4 is as follows: The system identifies the state based on the predicted probability of each state category, then provides decision support based on the identified state to determine whether to trigger an alarm, and generates operational suggestions by combining historical data and process knowledge.

5. The method for monitoring the state of collagen extraction process according to claim 1, characterized in that, S1 is as follows: Key equipment includes reaction vessels, centrifuges, and drying units; Process parameters include extraction temperature, pH value, extraction time, enzyme concentration, stirring speed, material concentration, pressure, solvent ratio, ultrasonic power, and centrifugation speed; Quality indicators include the purity, yield, molecular weight distribution, ash content, and moisture content of collagen. Set the sampling frequency to continuously collect complete production batch data; Data labeling is determined jointly by process experts based on quality index thresholds and process curve morphology.

6. A collagen extraction process status monitoring device, used to execute the collagen extraction process status monitoring method according to any one of claims 1-5, characterized in that, include: Data acquisition and annotation module: performs real-time acquisition, synchronous correlation, and precise annotation of process data; Data preprocessing module: Set up a normalization processing unit. The collected data first goes through the normalization processing unit to obtain normalized attribute values ​​that integrate multi-parameter correlation information; The model building and training module includes a coupled feature extraction submodule, a network weight initialization submodule, a forward propagation and coupled activation submodule, a state classification prediction submodule, and a model training optimization submodule. It receives the normalized attribute values ​​output by the data preprocessing module, passes them through each module in sequence, and outputs the trained model. Condition monitoring and decision support module: Based on the trained model output by the model building and training module, it performs real-time or offline identification of process conditions and provides decision support suggestions.

Images