A method and system for temperature and humidity coordinated control of a fermentation process

Abstract

The present application belongs to the field of advanced process control, and particularly relates to a temperature and humidity coordinated control method and system for fermentation process, which comprises the following steps: obtaining fermentation historical data for preprocessing, calculating sample weight based on time change rate, constructing a regression model to solve regression coefficient; constructing weighted Euclidean distance by using the calculated sensitivity coefficient, dividing stable working condition cluster by clustering algorithm, and constructing knowledge base; estimating temporary regression coefficient based on real-time data in online stage, calculating membership probability and weighted fusion model output, correcting rolling optimization objective function based on risk coefficient, and outputting control input. The present application solves the problem of model mismatch caused by time-varying working condition, improves the dynamic tracking and feedback regulation ability of temperature and humidity, and ensures the fermentation quality.

Description

Technical Field

[0001] This invention relates to the field of advanced process control technology. More specifically, this invention relates to a method and system for coordinated temperature and humidity control in a fermentation process. Background Technology

[0002] Fermentation is a crucial production process in fields such as biomedicine, food processing, and fine chemicals. Temperature and humidity are key environmental parameters affecting the stability of the fermentation process. They are coupled and work together to influence microbial growth and metabolism, enzyme activity changes, heat and mass transfer processes, and the efficiency of target product formation. As the fermentation cycle progresses, the moisture content, pH, metabolic intensity, and heat and mass transfer states of the material continuously change, causing the fermentation system to exhibit significant stage-specific, time-varying, and nonlinear characteristics. Traditional temperature and humidity control methods typically rely on independent single-loop regulation or fixed global mathematical models, which are ill-suited to adapting to changes in system parameters and multivariate coupling interference at different fermentation stages. This can easily lead to temperature or humidity deviations from the optimal process range, causing fluctuations in fermentation status, differences in metabolites, and batch-to-batch quality instability.

[0003] Currently, the introduction of multi-model predictive control (MMC) methods to control the fermentation process based on its nonlinear and time-varying characteristics is a current research direction. Chinese patent application CN114417720A discloses a MMC method for Pichia pastoris fermentation. This method utilizes fuzzy C-means clustering to divide prior data into multiple training sample clusters and employs support vector machines to establish corresponding local sub-predictive models. Based on these techniques, corresponding predictive controllers can be designed for different subspaces. By performing relative error weighted summation on the outputs of multiple sub-predictive controllers, a fused control input is obtained, improving the transient performance and output tracking effect of the multivariable strongly coupled system.

[0004] However, the aforementioned technical solutions can overcome the limitations of a single global model to some extent and improve the system's adaptability to nonlinear dynamic changes. However, existing technologies, when processing historical operating condition features extraction and model matching in fermentation processes, often employ conventional data clustering and global error assessment methods. This tends to weaken local stage characteristics and fails to reflect the differences in the impact of control inputs on temperature and humidity responses and quality indicators such as total acid content at different fermentation stages. During online control, due to the lack of in-depth extraction of historical typical operating condition features and a sensitive variable weighting mechanism, the control system struggles to identify the correlation between the current fermentation state and historical typical operating conditions in a timely and accurate manner. This leads to the interference of marginal variable fluctuations in operating condition clustering and state space localization. When the fermentation process faces operating condition switching or external multivariate coupling disturbances, the control model is prone to state matching lag and mismatch, thereby reducing the stability of temperature and humidity coordinated regulation and failing to effectively guarantee the batch quality consistency of the final fermentation product. Summary of the Invention

[0005] To address the technical problems of model mismatch, lag in temperature and humidity regulation, and unstable quality caused by time-varying operating conditions, this invention provides solutions in the following aspects.

[0006] In a first aspect, the present invention provides a method for coordinated temperature and humidity control in a fermentation process, comprising: S1, acquiring and preprocessing historical data including temperature, humidity, control input, key fermentation state parameters, and fermentation quality indicators; calculating sample weights based on the time change rate of the key fermentation state parameters, constructing a multi-output weighted elastic network regression model using a local sliding time window, and solving for the set of regression coefficient matrices corresponding to historical sampling times; S2, constructing a sensitivity analysis model of control input to temperature and humidity, calculating sensitivity coefficients, using the sensitivity coefficients to weight the regression coefficient components and constructing a weighted Euclidean distance; and using an affine propagation clustering algorithm based on the weighted Euclidean distance to partition the set of regression coefficient matrices to obtain a stable process. S3. In the online control stage, the temporary regression coefficient matrix is ​​estimated based on real-time data. The weighted Euclidean distance between the temporary regression coefficient matrix and the cluster core regression coefficient matrix contained in the knowledge base is calculated. The membership probability is calculated using the Softmax function. The membership probability is used as a weight to perform weighted fusion of the temperature and humidity prediction outputs of the models corresponding to each cluster core regression coefficient matrix to obtain the temperature and humidity prediction value sequence in the prediction time domain. The risk coefficient corresponding to the current working condition is calculated based on the fusion of cluster quality evaluation parameters according to the membership probability. A rolling optimization objective function is constructed, and the control input of the fermentation equipment corresponding to the next time step is obtained by solving the rolling optimization objective function.

[0007] This invention solves the problem of single-model mismatch caused by time-varying operating conditions by constructing a knowledge base and integrating online models; by extracting the sample weights of the state change rate and constructing a weighted elastic network, combined with rolling optimization control with risk coefficient correction, it improves the dynamic tracking and feedback adjustment capability of temperature and humidity, thereby ensuring the stability of fermentation quality.

[0008] Preferably, the step of acquiring and preprocessing historical data including temperature, humidity, control input, key fermentation state parameters, and fermentation quality indicators includes: using a cubic spline interpolation algorithm to align the historical data corresponding to different sampling frequencies in time sequence, so that the data of each batch are unified to the set standard time interval; extracting the global maximum and minimum values ​​of each parameter dimension to calculate the range, and normalizing all process data and fermentation quality indicators to the range of 0 to 1 by subtracting the minimum value from the current parameter value and dividing by the range, thereby generating a set of standardized feature matrices.

[0009] This invention uses a cubic spline interpolation algorithm for time-series alignment, eliminating timestamp misalignment caused by asynchronous data from multiple sources. It also uses range for normalization, eliminating the dimensional influence between parameters of different dimensions and improving the quality of historical data preprocessing.

[0010] Preferably, the process of solving for the set of regression coefficient matrices corresponding to historical sampling times includes: calculating the absolute difference between key fermentation state parameters at adjacent sampling times as the rate of change over time; smoothing the rate of change over time; normalizing the smoothed rate of change over time and mapping it to sample weights using a forward mapping function with an upper limit truncation; based on a local sliding time window, using temperature and humidity as joint prediction targets, and using control inputs and historical temperature and humidity states as regression variables, constructing a local weighted sum of squared errors loss function in combination with sample weights; using L1 regularization terms and L2 regularization terms in the local weighted sum of squared errors loss function to construct an elastic network objective function, minimizing the objective function using the coordinate descent method, and solving for the set of regression coefficient matrices.

[0011] This invention calculates sample weights using a positive mapping function with an upper limit truncation, which makes the model pay more attention to the effective dynamic change stage of the fermentation state during training, reducing the interference of abnormal peak samples on the training of multi-output weighted elastic network regression models.

[0012] Preferably, the step of constructing a sensitivity analysis model for temperature and humidity based on control input quantities and calculating sensitivity coefficients includes: constructing a feedforward neural network prediction model with the control input quantity sequence as the input node and temperature and humidity as the output nodes; training the network weights using historical data through a backpropagation algorithm until the network weights converge; calculating the partial derivatives of temperature and humidity with respect to each control input quantity based on the trained feedforward neural network prediction model; calculating the average value of the absolute values ​​of the partial derivatives of each control input quantity in the historical dataset and the time dimension of the input sequence, and normalizing the average value corresponding to each control input quantity as the sensitivity coefficient of the corresponding control input quantity.

[0013] This invention utilizes a constructed feedforward neural network prediction model to calculate partial derivatives to obtain sensitivity coefficients, accurately and quantitatively reflecting the degree of heterogeneity in the effect of different control inputs on temperature and humidity responses.

[0014] Preferably, the step of constructing a weighted Euclidean distance by weighting the regression coefficient components using sensitivity coefficients includes: for any two regression coefficient matrices, extracting the regression coefficient components at corresponding positions and calculating the square of the difference between the regression coefficient components; determining the weights according to the type of regression independent variable corresponding to the regression coefficient components, using the sensitivity coefficient associated with the corresponding control input as the weight for the regression coefficient components corresponding to the control input, and using preset state weights as the weights for the regression coefficient components corresponding to historical temperature and historical humidity states; multiplying the determined weights by the square of the corresponding regression coefficient difference; summing all the weighted squared terms and taking the square root of the summation result to obtain the weighted Euclidean distance between the two regression coefficient matrices.

[0015] Preferably, the step of using a weighted Euclidean distance-based affine propagation clustering algorithm to divide the regression coefficient matrix set to obtain stable operating condition clusters and cluster center regression coefficient matrices, and generating cluster quality evaluation parameters based on fermentation quality indicators to construct a knowledge base, includes: using the negative value of the weighted Euclidean distance as the similarity measurement matrix of the affine propagation clustering algorithm, setting the initial reference value as the median of the similarity measurement matrix; iteratively transmitting attraction and affiliation information between data points in the regression coefficient matrix, updating the probability of a data point becoming a cluster center and its affiliation to the cluster center; using the cluster center nodes obtained from clustering as the cluster center regression coefficient matrix of the stable operating condition clusters, statistically analyzing the quality mean, quality fluctuation range, and quality compliance rate corresponding to historical samples within the stable operating condition clusters, and storing the statistically analyzed parameters as cluster quality evaluation parameters associated with the cluster center regression coefficient matrix in the knowledge base.

[0016] Preferably, the calculation of the weighted Euclidean distance between the temporary regression coefficient matrix and the cluster center regression coefficient matrix, using the Softmax function to calculate the membership probability, includes: calculating the weighted Euclidean distance between the currently estimated temporary regression coefficient matrix and all cluster center regression coefficient matrices in the knowledge base; selecting the N cluster center regression coefficient matrices with the smallest distances, and calculating the reciprocal of the weighted Euclidean distance between the temporary regression coefficient matrix and the selected N cluster center regression coefficient matrices respectively; multiplying each reciprocal distance by a set smoothing constant as an exponent, and substituting it into the natural exponential function to calculate the exponent value corresponding to each cluster center; dividing the exponent value corresponding to a single cluster center by the sum of the exponent values ​​corresponding to the N cluster centers to obtain the membership probability of the current working condition to the adjacent N working condition clusters.

[0017] This invention utilizes the inverse of distance combined with the Softmax function to calculate membership probability, effectively adjusting the distribution steepness during multi-model switching, ensuring a smooth model switching process, and reducing the occurrence of abrupt changes in control commands.

[0018] Preferably, the step of constructing the rolling optimization objective function and solving the rolling optimization objective function to obtain the control input of the fermentation equipment at the next time step includes: within a set prediction time domain, calculating the sum of squares of the deviations between the predicted temperature and humidity values ​​obtained by weighted fusion of the outputs of each cluster core model and the expected temperature and humidity setpoints; calculating the sum of squares of the penalty terms corresponding to the rate of change of control input within the prediction time domain; adjusting the weights of the temperature and humidity tracking error term and the penalty term for the change of control input according to the risk coefficient corresponding to the current operating condition; weightedly adding the sum of squares of the deviations after weight adjustment to the sum of squares of the penalty terms to construct a rolling optimization objective function containing upper and lower limits of hardware execution constraints; using a mixed integer quadratic programming algorithm to solve for the optimal control input sequence, and extracting the first element in the optimal control input sequence as the temperature and humidity coordinated control input applied to the fermentation equipment at the next time step.

[0019] Preferably, the risk coefficient is obtained by weighting one or more of the following: quality deviation, quality fluctuation, and quality non-compliance rate corresponding to the fermentation quality indicators.

[0020] Secondly, the present invention provides a temperature and humidity coordinated control system for a fermentation process, including a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned temperature and humidity coordinated control method for a fermentation process is implemented.

[0021] By adopting the above technical solution, a computer program is generated for the temperature and humidity coordinated control method of the fermentation process, and stored in a memory so that it can be loaded and executed by a processor. A terminal device can then be made based on the memory and the processor for convenient use.

[0022] The beneficial effects of this invention are as follows: This invention introduces sample weights for the time-varying rates of change corresponding to key fermentation state parameters, enabling the modeling process to focus more on the effective change stages corresponding to the fermentation state and improving the expressive power of the dynamic response relationship between temperature and humidity. By constructing a weighted Euclidean distance using the sensitivity coefficients of control inputs to temperature and humidity responses, the working condition clustering and online matching processes can highlight the influence of key control components and reduce the interference of fluctuations in weakly correlated variables on the working condition identification results. By constructing a knowledge base of typical stable working conditions and combining it with total acid content to form cluster quality evaluation parameters, the online control process can simultaneously refer to historical working condition characteristics and fermentation quality status.

[0023] Furthermore, by fusing the cluster core model output through membership probability and correcting the weights of the rolling optimization objective function based on the risk coefficient of total acid content, the control system reduces the deviations corresponding to temperature and humidity under operating condition switching and external disturbances, and adjusts the change amplitude of the control input within the equipment constraints, thereby improving the stability of temperature and humidity coordinated control in the fermentation process and improving the batch consistency of total acid content in fermentation products. Attached Figure Description

[0024] Figure 1 This is a flowchart of a method for coordinated temperature and humidity control in a fermentation process according to the present invention; Figure 2 This is a schematic diagram of cluster analysis in this invention; Figure 3 This is a schematic diagram comparing the variance of the valve action change rate in this invention. Detailed Implementation

[0025] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0026] This invention discloses a method for coordinated temperature and humidity control in a fermentation process, referring to... Figure 1 This includes steps S1-S3: S1. Obtain historical data and solve the regression model.

[0027] In an optional embodiment, historical fermentation data including temperature, humidity, control inputs, key fermentation state parameters, and fermentation quality indicators are acquired and preprocessed; sample weights are calculated based on the time change rate of key fermentation state parameters, and a multi-output weighted elastic network regression model is constructed using a local sliding time window to solve for the set of regression coefficient matrices corresponding to each historical sampling time.

[0028] The system acquires historical fermentation data by reading the status of temperature and humidity sensors and actuators using a programmable logic controller (PLC). It also obtains pH values ​​as key fermentation parameters using biosensors and records total acid content offline as a fermentation quality indicator. The collected data undergoes missing value imputation, outlier removal, and normalization to the 0-1 range to obtain preprocessed historical fermentation data. For the preprocessed pH value time series, the absolute difference between pH values ​​at adjacent sampling times is calculated as the rate of change over time. This rate of change is then smoothed and normalized. The processed values ​​are then substituted into a forward mapping function with an upper limit truncation to obtain the sample weights for each sampling time. This ensures that the sample weights for effective dynamic changes are higher than those for stable periods, and reduces the impact of abnormal spikes on model training.

[0029] A fixed-length sliding time window is defined, and data within each window is extracted using time series slicing. The temperature and humidity at the next sampling time are used as the output matrix, while the current valve opening and humidifier setting, along with the temperature and humidity at the previous time, are used as the input matrix. A multi-output elastic network regression model is constructed, and the calculated sample weights are input into the model fitting function for weighted solving. By extracting the regression coefficient attributes, the regression coefficient matrix corresponding to each historical sampling time is obtained. For the model intercept vector, the corresponding model intercept vector and regression coefficient matrix are associated and stored, or a constant term column is added to the input matrix to incorporate the corresponding intercept term into the regression coefficient matrix. After iterating through all sliding time windows, a set of regression coefficient matrices is generated.

[0030] In an optional embodiment, historical fermentation data, including temperature, humidity, control inputs, key fermentation state parameters, and fermentation quality indicators, is acquired and preprocessed. Specifically, a cubic spline interpolation algorithm is used to time-align historical fermentation process data corresponding to different sampling frequencies, unifying the data of each batch to a set standard time interval. For each parameter dimension, the maximum and minimum values ​​corresponding to the historical global data are extracted to calculate the range. By subtracting the corresponding minimum value from the current parameter value and dividing by the calculated range, all process data and fermentation quality indicators are normalized to the 0-1 interval, generating a set of standardized feature matrices.

[0031] In the time-series alignment, multiple batches of historical fermentation data were collected. The temperature sensor was sampled every 30 seconds, while the humidity sensor and key fermentation parameters such as pH were sampled every 2 minutes. To reduce timestamp misalignment caused by asynchronous multi-source data, a cubic spline interpolation algorithm was used to construct a piecewise smooth polynomial function. A standard time interval of 1 minute was set, with specific intervals ranging from 30 seconds to 5 minutes. All batches of data were resampled and aligned to the hourly time at each 1-minute interval. For example, if humidity data for minute 11 was missing between minutes 10 and 12, the estimated humidity value for minute 11 was calculated using the continuity constraint of the derivatives of the current and adjacent intervals. In the data normalization stage, all aligned parameter dimensions were traversed to construct a historical global dataset. Taking temperature as an example, if the historical highest temperature was 45°C and the lowest temperature was 20°C, the calculated range was 25°C; if the temperature at the current sampling point was 30°C, the normalized value was 0.4. Similarly, all humidity parameters ranging from 40% to 95%, the operating ranges corresponding to the opening of the heating valve and the humidifier setting in the control input, and the total acid content of the fermentation quality index are mapped dimensionlessly according to the range calculation method to generate a standardized feature matrix set.

[0032] In an optional embodiment, sample weights are calculated based on the time-varying rate of change of key fermentation state parameters. A multi-output weighted elastic network regression model is constructed using a local sliding time window to solve for the set of regression coefficient matrices corresponding to each historical sampling moment. Specifically, the absolute difference between the key fermentation state parameters corresponding to adjacent sampling moments is calculated as the time-varying rate. The extracted time-varying rate is smoothed, and after normalization, it is mapped to the sample weights corresponding to each historical sample point through a forward mapping function with an upper limit truncation. Based on a local sliding time window at each historical sampling moment, temperature and humidity are used as joint prediction targets, and control inputs, historical temperature and humidity states are used as regression independent variables. A local weighted sum of squared errors loss function is constructed by combining the obtained sample weights. In the local weighted sum of squared errors loss function, L1 regularization and L2 regularization terms are used to construct a local elastic network objective function. The corresponding objective function is minimized by the coordinate descent method to obtain the set of regression coefficient matrices corresponding to each historical sampling moment.

[0033] When calculating sample weights, normalized key fermentation state parameters, such as pH, are extracted from adjacent sampling times. The absolute value of the difference is calculated as the rate of change over time, and a moving average filter with a window size of 3 is used to smooth the extracted rate of change. The smoothed rate of change is denoted as... The minimum and maximum values ​​from historical samples were used to measure the rate of change over time. Normalization is performed to obtain the normalized rate of change over time. Set the exception truncation threshold to Its preferred value is 0.85; when the normalized time rate of change Greater than the abnormal cutoff threshold When, normalize the rate of change over time. Truncating to an abnormal truncation threshold To ensure that the sample weights corresponding to the effective dynamic change phase are higher than those corresponding to the stationary phase, a new sample weight is theoretically derived by combining the ratio of the rate of change over time and the cutoff threshold with the gain coefficient.

[0034] in, For sample weights; This is the gain coefficient, which ranges from 0.5 to 5, with a preferred value of 2. This represents the normalized rate of change over time. This is the threshold for abnormal truncation.

[0035] The resulting sample weights The numerical range is from 1 to 3.

[0036] When constructing the local model, a local sliding time window with a length of 60 sampling periods is set, specifically ranging from 30 to 120 sampling periods. For samples within the time window, the temperature and humidity of the next step are used as the joint prediction target, and the current heating valve opening, humidifier setting, and historical temperature and humidity status are used as regression independent variables. A local weighted sum of squared errors loss function is constructed by combining the obtained sample weights. L1 and L2 regularization terms are added to the constructed loss function to form the elastic network objective function, where the L1 regularization coefficient is set to 0.05 (range 0.01 to 0.1), and the L2 regularization coefficient is set to 0.5 (range 0.1 to 1). The coordinate descent method is used to solve the established elastic network objective function. Iteration stops when the norm of the coefficient difference between two consecutive iterations is less than 0.0001, and the regression coefficient matrix set is output by sliding along the time axis window by window. The columns of the generated regression coefficient matrix correspond to the dimensions of the regression independent variables. It includes columns of control input coefficients corresponding to the opening degree of the heating valve and the setting of the humidifier, as well as columns of state coefficients corresponding to the historical temperature and humidity states. The model intercept vector is output as an independent parameter along with the regression coefficient matrix, or it can be incorporated into the corresponding regression coefficient matrix by adding a constant term column.

[0037] S2. Calculate distance clustering and build a knowledge base.

[0038] In an optional embodiment, a sensitivity analysis model of the control input to temperature and humidity is constructed to calculate the sensitivity coefficients. The calculated sensitivity coefficients are used to weight the regression coefficient components to construct a weighted Euclidean distance. An affine propagation clustering algorithm based on the weighted Euclidean distance is used to divide the regression coefficient matrix set to obtain the stable operating condition cluster and the cluster core regression coefficient matrix. Cluster quality evaluation parameters are generated based on the fermentation quality index total acid content, thus completing the construction of the knowledge base.

[0039] A multi-output prediction model based on gradient boosting trees is constructed by extracting control inputs from historical data as independent variables and temperature and humidity as dependent variables. After model training, the TreeExplainer interpreter from the SHAP library is used to perform feature contribution analysis on the constructed multi-output prediction model, calculating the SHAP contribution value of each control input feature to the temperature and humidity prediction outputs. The absolute value of the SHAP contribution value corresponding to the same control input is taken, and the average is calculated in both the sample and output dimensions. The average value corresponding to each control input is then normalized to obtain the corresponding sensitivity coefficient.

[0040] A weighted Euclidean distance calculation function is established between regression coefficient matrices, and the weights are determined based on the type of regression independent variable corresponding to the regression coefficient components. Regression coefficient components corresponding to control inputs are weighted using the sensitivity coefficient of the corresponding control input; regression coefficient components corresponding to historical temperature and humidity states are weighted using preset state weights. Following the custom weighted Euclidean distance calculation rules, the distance value is calculated for each pair of regression coefficient matrices in the set, and all pairwise distance values ​​are arranged in matrix form to generate a distance matrix.

[0041] The negative of the generated distance matrix yields the similarity matrix, and the median of the similarity matrix is ​​extracted as the reference parameter for the affine propagation clustering algorithm. The regression coefficient matrix set is clustered to obtain the cluster labels for each regression coefficient matrix and the cluster center index for each stable operating condition cluster. Regression coefficient matrices with the same cluster labels are grouped into the same stable operating condition cluster, and the total acid content corresponding to historical samples within each stable operating condition cluster is extracted. The mean total acid content, the range of total acid content fluctuation, and the compliance rate of total acid content are calculated as cluster quality evaluation parameters. Finally, the cluster center regression coefficient matrix, intercept vector or constant term parameter, sensitivity coefficient, preset state weights, and cluster quality evaluation parameters are associated and stored in a Redis database, completing the knowledge base construction.

[0042] In an optional embodiment, a sensitivity analysis model of the control input quantities to temperature and humidity is constructed to calculate the sensitivity coefficients. Specifically, a feedforward neural network prediction model is constructed with the control input quantity sequence as the input node and temperature and humidity as the output nodes. The network weights are trained using historical fermentation data through a backpropagation algorithm until the network weights converge. Based on the trained feedforward neural network prediction model, the partial derivatives of temperature and humidity with respect to each control input quantity are calculated. The average value of the absolute values ​​of the partial derivatives of each control input quantity over the entire historical dataset and the time dimension of the input sequence is calculated, and the average value corresponding to each control input quantity is normalized to obtain the sensitivity coefficient of the corresponding control input quantity.

[0043] When constructing the feedforward neural network prediction model, the network structure includes an input layer, two hidden layers, and an output layer. The input layer is used to receive the sequence of control input quantities within 10 consecutive sampling periods. The corresponding control input quantities include two features: the opening degree of the heating valve and the humidifier setting. The input data is represented as a 10×2 two-dimensional matrix and flattened into a 20-dimensional vector before being input to the model.

[0044] The first hidden layer has 64 neurons to extract high-dimensional response features from the control input sequence; the second hidden layer has 32 neurons to further compress and combine the extracted high-dimensional response features. Both hidden layers use the ReLU activation function for non-linear mapping. The output layer has two output nodes, corresponding to the temperature prediction and humidity prediction values ​​respectively, and the output is represented as a two-dimensional vector composed of the temperature and humidity prediction values. The model layers are linearly transformed through weight matrices and bias vectors, and then passed layer by layer through the hidden layer activation functions to finally obtain the temperature and humidity prediction outputs corresponding to the control input sequence.

[0045] The network was trained using backpropagation on a historical normalized fermentation dataset with a learning rate of 0.005. The Adam optimizer was used to update the network weight parameters. Convergence was defined as the loss function decreasing by less than one ten-thousandth within 20 consecutive training epochs, at which point training was stopped and the model structure was fixed. When obtaining the sensitivity coefficients, a feedforward neural network prediction model with fixed weights was used, and the chain rule was applied to calculate the partial derivatives of specific control input variables with respect to temperature and humidity. For each sample batch in the historical dataset and each time step in the input sequence, the partial derivatives of temperature and humidity with respect to the current control input were calculated. The partial derivatives of each control input were extracted across all sample batches (1000 batches) and the sequence time dimension (10 steps). For each control input, the absolute values ​​of the partial derivatives with respect to temperature and humidity were averaged to obtain the basic sensitivity index for that control input. For example, if the average partial derivative corresponding to the heating valve opening is 0.8 and the average partial derivative corresponding to the humidifier setting is 0.2, the basic indicators are normalized, and 0.8 and 0.2 are used as the sensitivity coefficients of the heating valve opening and the humidifier setting, respectively, to represent the degree of influence of different operating variables on the temperature and humidity response.

[0046] In an optional embodiment, a weighted Euclidean distance is constructed by weighting the regression coefficient components using a sensitivity coefficient. Specifically, for any two regression coefficient matrices, the regression coefficient components at corresponding positions are extracted, and the square of the difference between the extracted regression coefficient components is calculated. The weights are determined according to the type of regression independent variable corresponding to the regression coefficient components. The regression coefficient components corresponding to the control input are weighted using the sensitivity coefficient associated with the corresponding control input, while the regression coefficient components corresponding to historical temperature and humidity states are weighted using preset state weights. The determined weights are multiplied by the square of the corresponding regression coefficient difference. All weighted squared terms are summed, and the square root of the summation is taken to obtain the weighted Euclidean distance between the two regression coefficient matrices.

[0047] When calculating the spatial similarity between regression coefficient matrices of the same dimension generated at two different times, a double loop is used to traverse all rows and columns of the matrix and extract the regression coefficient components corresponding to the positions. The element differences corresponding to the same predicted output (i.e., temperature or humidity) under the same regression independent variable dimension are calculated sequentially, and the calculated differences are squared to obtain the basic squared distance term reflecting the differences in the underlying mapping of each single dimension.

[0048] After obtaining the squared terms for all dimensions, the weights are determined based on the type of the regression independent variable corresponding to each regression coefficient component. For regression coefficient components corresponding to control inputs, the sensitivity coefficients associated with the corresponding control inputs are indexed and multiplied term by term. For regression coefficient components corresponding to historical temperature and humidity states, preset state weights are used and multiplied term by term. The preset state weights are either 1 or the average of the sensitivity coefficients of each control input. For example, if the first dimension represents the opening of a heating valve, with a corresponding sensitivity coefficient of 0.8 and a squared difference of 0.04, then the weighted squared term at the corresponding position is 0.032. This mechanism increases the weight of key control inputs that play an important role in the dynamic changes of temperature and humidity in the distance metric. All the obtained weighted squared terms are globally summed and accumulated, and the square root of the sum is taken. The resulting weighted Euclidean distance reduces the interference of non-core variable fluctuations on the distance metric.

[0049] In an optional embodiment, an affine propagation clustering algorithm based on weighted Euclidean distance is used to partition the set of regression coefficient matrices, obtaining stable operating condition clusters and cluster core regression coefficient matrices. Cluster quality evaluation parameters are then generated based on the fermentation quality indicator of total acid content, thereby constructing a knowledge base. Specifically, the negative value of the obtained weighted Euclidean distance is used as the similarity metric matrix of the affine propagation clustering algorithm, and the initial reference value is set as the median of the similarity metric matrix. Attraction and affiliation information are iteratively passed between data points in each regression coefficient matrix, updating the probability of each data point becoming a cluster core and the affiliation relationship of other data points to the cluster core. The central nodes of each cluster are used as the cluster core regression coefficient matrix of the stable operating condition clusters, and the mean total acid content, total acid content fluctuation range, and total acid content compliance rate of historical samples within each stable operating condition cluster are statistically analyzed. These statistical parameters are then used as cluster quality evaluation parameters and associated with the corresponding cluster core regression coefficient matrices and stored in the knowledge base.

[0050] When constructing the clustering similarity metric matrix, the weighted Euclidean distance between each pair of data points in the regression coefficient matrix of the historical dataset is calculated. The negative value of the calculated distance is used as the input parameter for the affine propagation clustering algorithm, so that the similarity of data points that are spatially closer is closer to 0. To help the algorithm determine the number of operating condition clusters, the global median of the entire similarity matrix, for example, a median value of -3.5, is set as the initial reference value, and the extracted median is filled into the diagonal of the similarity matrix. Using the median reduces the generation of redundant small clusters and also reduces the probability of merging different heating or insulation operating conditions.

[0051] During the iterative clustering calculation, two key information variables continuously interact among the data points. Attraction information indicates the suitability of a candidate data point as a cluster center, while membership information indicates the suitability of a data point in selecting a candidate node as a cluster center. To suppress numerical oscillations during information transmission in high-dimensional space, a damping coefficient is input, with an optimal range of 0.5 to 0.9 (e.g., set to 0.7). In each information iteration, the latest calculated value and the previous round's value are proportionally weighted and smoothed. The iterative process continuously monitors the composite probability of each data point becoming a cluster center. The process terminates when the cluster center set remains static for 50 consecutive iterations, or when the total number of iterations reaches the maximum constraint limit of 1000. Nodes with a probability greater than 0 are identified as cluster centers representing different typical fermentation stages, and their corresponding regression coefficient matrices are labeled with an index and stored in the knowledge base.

[0052] Reference Figure 2 This paper demonstrates the results of clustering the regression coefficient matrix set using the affine propagation clustering algorithm. The two-dimensional distribution space formed by weighted feature 1 and weighted feature 2 reveals the cluster center locations of clusters 1, 2, and 3. By employing a clustering method based on weighted Euclidean distance, the discretely distributed regression coefficient matrix data points exhibit a clear cluster separation, proving that the weighted Euclidean distance feature effectively highlights the dynamic characteristics and biases of the multidimensional coefficient space, enabling the resulting clusters to accurately represent different stable fermentation process stages.

[0053] S3. Calculate the control input and optimize the control.

[0054] In an optional embodiment, during the online control phase, a temporary regression coefficient matrix is ​​estimated based on real-time data within a set sliding time window. The weighted Euclidean distance between the temporary regression coefficient matrix and the cluster core regression coefficient matrix contained in the knowledge base is calculated, and the membership probability is calculated using the Softmax function. The calculated membership probability is used as a weight to perform weighted fusion of the temperature and humidity prediction outputs of the models corresponding to each cluster core regression coefficient matrix, resulting in a sequence of predicted temperature and humidity values ​​in the prediction time domain. Based on the calculated membership probability and the cluster quality evaluation parameters in the knowledge base, the total acid content risk coefficient corresponding to the current operating condition is calculated. A rolling optimization objective function is constructed, and the constructed rolling optimization objective function is solved to obtain the control input of the fermentation equipment corresponding to the next time step.

[0055] During the online fermentation process, temperature, humidity, and control input data are collected in real time within a set sliding time window. The collected real-time data are then normalized using the maximum and minimum values ​​saved from the historical preprocessing stage. Based on the normalized window data, a local regression sample is constructed with temperature and humidity as outputs and control inputs and historical temperature and humidity states as inputs. The temporary regression coefficient matrix corresponding to the current window is then obtained. The intercept term in the temporary regression coefficient matrix is ​​extracted separately using the same method as in the offline modeling stage, or it is incorporated into the temporary regression coefficient matrix by adding a constant term column.

[0056] Retrieve the cluster center regression coefficient matrix, intercept vector or constant term parameter, sensitivity coefficient, preset state weight, and cluster quality evaluation parameters corresponding to each stable operating condition cluster from the knowledge base. Calculate the distance between the temporary regression coefficient matrix and each cluster center regression coefficient matrix according to the weighted Euclidean distance calculation rule, convert the obtained distance values ​​into similarity values, and then use Softmax normalization to obtain the membership probability of the current operating condition to each candidate stable operating condition cluster.

[0057] Within the prediction time domain, each membership probability is used as a weight for the output of the corresponding cluster core model. For each prediction step, the candidate control input corresponding to the current prediction step and the temperature and humidity prediction states corresponding to the previous prediction step are input into each cluster core model to obtain the temperature and humidity prediction values ​​corresponding to each cluster core model. Then, the outputs of each cluster core model are weighted and summed according to the membership probability to obtain the fused temperature and humidity prediction value corresponding to the current prediction step. By iteratively advancing along the prediction time domain, a sequence of temperature and humidity prediction values ​​within the prediction time domain is obtained.

[0058] Based on the membership probability corresponding to the current operating condition, the mean total acid content, the fluctuation range of total acid content, and the compliance rate of total acid content for the corresponding stable operating condition cluster in the knowledge base are weighted and fused to obtain the fused mean total acid content, the fluctuation range of total acid content, and the compliance rate of total acid content, respectively. The deviation of total acid content is calculated based on the difference between the fused mean total acid content and the target total acid content; the fluctuation of total acid content is calculated based on the fluctuation range of fused total acid content; and the non-compliance rate of total acid content is calculated based on the value obtained by subtracting the compliance rate of fused total acid content from 1. One or more of the calculated deviation of total acid content, fluctuation of total acid content, and non-compliance rate of total acid content are weighted and summed according to preset weights to obtain the total acid content risk coefficient.

[0059] The temperature and humidity tracking error weights and control input change penalty weights in the rolling optimization objective function are adjusted based on the calculated total acid content risk coefficient. Specifically, when the total acid content risk coefficient increases, the temperature and humidity tracking error weights are increased, while the control input change penalty weights are appropriately reduced according to a preset ratio. This improves the temperature and humidity correction capability within the equipment constraints, while limiting excessive control actions through control change rate constraints.

[0060] A rolling optimization objective function is constructed based on the fused temperature and humidity prediction sequence in the prediction time domain, the corrected temperature and humidity tracking error weights, and the corrected control input change penalty weights. The constructed rolling optimization objective function includes temperature and humidity tracking error terms and control input change penalty terms, and is solved in conjunction with upper and lower limits of equipment hardware execution constraints. When the control input is treated as a continuous variable and both the prediction equation and constraints are linear, a quadratic programming algorithm is used for solving; when the humidifier setting is used as a discrete control input parameter for optimization, a mixed-integer quadratic programming algorithm is used. The first element of the obtained control input sequence is extracted as the control command for the heating valve opening and the humidifier setting at the next time step, and the control commands are sent to the underlying hardware for execution.

[0061] In an optional embodiment, the weighted Euclidean distance between the temporary regression coefficient matrix and the cluster center regression coefficient matrix in the knowledge base is calculated, and the membership probability is calculated using the Softmax function. Specifically, the weighted Euclidean distance between the currently estimated temporary regression coefficient matrix and all cluster center regression coefficient matrices in the knowledge base is calculated; the N cluster center regression coefficient matrices with the smallest distances are selected, and the reciprocal of the weighted Euclidean distance between the temporary regression coefficient matrix and the extracted N cluster center regression coefficient matrices is calculated for each; each reciprocal distance is multiplied by a set smoothing constant and used as an exponent term, which is then substituted into the natural exponential function to calculate the exponent value corresponding to each cluster center; the exponent value corresponding to a single cluster center is divided by the sum of the exponent values ​​corresponding to the N cluster centers to obtain the membership probability of the current working condition to the N adjacent working condition clusters.

[0062] During the online control phase, a temporary regression coefficient matrix reflecting the current transient characteristics is estimated based on a real-time sliding time window containing sampling points from the past 30 minutes. A knowledge base is searched, and a weighted Euclidean distance is calculated between the temporary regression coefficient matrix and all pre-stored regression coefficient matrices for cluster centers in 15 typical stages, according to a predetermined weighting mechanism, to pinpoint the current state's spatial location. All distance measures are sorted in ascending order, and the three candidate cluster centers with the smallest distance values ​​(preferably N values ​​of 3 to 5) are selected. The reciprocals of these three minimum distances are then calculated. This reciprocal transformation essentially converts the distance index into a positive similarity multiplier; candidate cluster centers with shorter distances receive a non-linearly amplified similarity metric.

[0063] When generating the membership probability distribution, a smoothing constant is preset, preferably ranging from 0.1 to 1. For example, if the smoothing constant is set to 0.5, the three extracted reciprocals of distance are multiplied by 0.5 and then used as independent variables in the natural exponential function to obtain their corresponding activation index values. The smoothing constant is used to adjust the distribution steepness during multi-model switching, reducing the magnitude of abrupt changes in control commands. A standard Softmax normalization operation is performed, dividing the activation index value generated by each candidate cluster center by the sum of the three calculated index values, outputting a probability distribution vector with a sum of 1 and each component between 0 and 1. The result provides the membership confidence of the current transient fermentation condition to the three most recent typical conditions and forms the mathematical basis for subsequent controller output fusion.

[0064] In an optional embodiment, a rolling optimization objective function is constructed, and the control input of the fermentation equipment at the next time step is obtained by solving the constructed rolling optimization objective function. Specifically, within the set prediction time domain, the sum of squares of the deviations between the predicted temperature and humidity values ​​obtained by weighted fusion of the outputs of each cluster core model and the expected temperature and humidity setpoints is calculated; the sum of squares of the penalty terms corresponding to the rate of change of control input within the prediction time domain is calculated; the weights of the temperature and humidity tracking error term and the penalty term for the change of control input are adjusted according to the total acid content risk coefficient corresponding to the current operating condition; wherein, the total acid content risk coefficient is obtained by weighting one or more of the total acid content deviation, total acid content fluctuation, and total acid content non-compliance rate; the weighted sum of squares of the deviation and the weighted sum of squares of the penalty terms are weighted and summed to construct a rolling optimization objective function containing upper and lower limits constraints for equipment hardware execution; a mixed integer quadratic programming algorithm is used to solve the problem to obtain the optimal control input sequence within the prediction time domain, and the first control input in the optimal control input sequence is extracted as the temperature and humidity collaborative control input applied to the fermentation equipment at the next time step.

[0065] When performing predictive control calculations, a future-oriented prediction time domain is set, preferably with a sampling step size of 5 to 20 steps, for example, 10 steps. Based on the fusion prediction model corresponding to the current operating conditions, the predicted temperature and humidity values ​​within the prediction time domain are derived, and the deviations between the predicted temperature and humidity values ​​and the process set temperature and humidity are calculated respectively.

[0066] The baseline temperature tracking error weight is set to 10, and the baseline humidity tracking error weight is set to 5. The total acid content risk coefficient corresponding to the current operating condition is normalized to the range of 0 to 1, denoted as... The temperature and humidity tracking error weights are adjusted based on the total acid content risk coefficient; that is, the temperature tracking error weight is equal to the base temperature tracking error weight multiplied by [the factor]. The humidity tracking error weight is equal to the humidity baseline tracking error weight multiplied by ;in, To track the weight adjustment coefficient, its preferred value is 1. When the total acid content risk coefficient increases, the weight of the temperature and humidity tracking error increases, thereby improving the tracking accuracy of the rolling optimization process for the set temperature and humidity.

[0067] Simultaneously, the change in control input between adjacent sampling periods is calculated within the control time domain, which is preferably performed in three steps. The basic penalty weight for the control action is set to 0.8, and the penalty weight for changes in control input is adjusted inversely based on the total acid content risk coefficient; that is, the penalty weight for changes in control input equals the basic penalty weight for the control action multiplied by... And set a lower limit for the penalty weight; among which, The penalty adjustment coefficient is set to 0.3, and the lower limit of the penalty weight is preferably set to 50% of the basic penalty weight of the control action. When the risk coefficient of total acid content increases, the penalty weight for changes in control input is appropriately reduced, so that the opening of the heating valve and the humidifier setting can be adjusted more fully within the allowable range of the equipment.

[0068] The weighted sum of the squared errors in temperature and humidity tracking and the squared penalty for changes in control input are used to form the rolling optimization objective function. Therefore, under conditions with high total acid content risk, the rolling optimization process improves the tracking accuracy of temperature and humidity setpoints and appropriately relaxes the constraints on changes in control input to reduce the risk of quality deviation.

[0069] In the constructed objective function, constraint boundary conditions for the equipment hardware are set. For example, the absolute opening degree of the heating valve is constrained to be between 0% and 100%, the humidifier control input is constrained to be within the continuous interval corresponding to the upper and lower limits of the preset gear, and the maximum rate of change of action in a single control cycle is limited to the range of -10% to +10%. For the discrete control input of the humidifier gear type, it is solved by candidate gear enumeration combined with a quadratic programming subproblem, or by a mixed integer quadratic programming algorithm, to obtain a sequence of control inputs that satisfy the discrete gear constraints, equipment upper and lower limit constraints, and control rate of change constraints. Based on the established constrained optimization model, the corresponding solver is called to solve for the minimum point of the objective function within the constrained feasible region, and a set of incremental control input sequences covering the next three steps is output. In accordance with the rolling optimization principle of model predictive control, only the first incremental instruction in the optimal sequence is extracted and attached to the current baseline action, and sent to the underlying PLC hardware as the control input that actually acts on the fermentation equipment at the next moment. After the next sampling to obtain new feedback, the entire feedback prediction calculation process is retried.

[0070] A 30-day, multi-batch controlled comparative experiment was conducted using industrial fermentation equipment in a fermentation workshop. Each experimental group comprised 20 independent fermentation batches, with initial fermentation conditions and raw material states kept largely consistent. The target process setting was a temperature of 35°C and a relative humidity target of 80%. The sampling cycle for multiple sensors and the issuance cycle of control commands were uniformly set to 1 minute. The experiment was divided into four groups: a baseline control group using a conventional model predictive control algorithm; ablation group 1 with sensitivity coefficient weighting removed; ablation group 2 with time-varying rate sample weighting removed; and the proposed scheme group using the complete algorithm logic. Evaluation metrics included mean absolute error of temperature, mean absolute error of humidity, variance of valve action rate of change, and the pass rate of batches meeting the total acid content standard.

[0071] Experimental results showed that the baseline control group had a mean absolute error of 1.5°C for temperature, a mean absolute error of 4.2% for humidity, a variance of 12.5 for valve action rate of change, and a total acid content pass rate of 82%. Ablation group 1 used unweighted Euclidean distance for condition clustering and matching, with mean absolute errors of 1.1°C and 3.5% for temperature and humidity, respectively, a variance of 9.2 for valve action rate of change, and a total acid content pass rate of 88%. Ablation group 2 used equally weighted samples for regression fitting, with mean absolute errors of 1.2°C and 3.8% for temperature and humidity, respectively, a variance of 10.1 for valve action rate of change, and a total acid content pass rate of 86%. The complete experimental scheme had a mean absolute error of 0.4°C for temperature, a mean absolute error of 1.5% for humidity, a variance of 3.4 for valve action rate of change, and a total acid content pass rate of 98%.

[0072] Reference Figure 3 The baseline control group had the largest variance in valve action change rate, while the ablation groups 1 and 2 had lower variances. The proposed scheme group had the smallest variance in valve action change rate. This demonstrates that constructing a weighted Euclidean distance using the sensitivity coefficients of control inputs to temperature and humidity responses reduces the interference of marginal variable fluctuations on condition clustering and state-space localization. Using positive sample weights with upper limit truncation improves the model's fit to the effective dynamic response phase and reduces the impact of abnormal peak samples on model training. Combining the probabilistic fusion prediction model, the total acid content risk coefficient correction mechanism, and the rolling optimization objective function reduces temperature and humidity tracking errors and control command fluctuations, thereby improving the batch compliance rate of total acid content in fermentation products.

[0073] This invention also discloses a temperature and humidity coordinated control system for a fermentation process, including a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement a temperature and humidity coordinated control method for a fermentation process according to the present invention.

[0074] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0075] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for coordinated temperature and humidity control in a fermentation process, characterized in that, include: S1. Obtain historical data including temperature, humidity, control input, key fermentation state parameters and fermentation quality indicators and perform preprocessing; calculate sample weights based on the time change rate of key fermentation state parameters, construct a multi-output weighted elastic network regression model using a local sliding time window, and solve for the set of regression coefficient matrices corresponding to the historical sampling time. S2. Construct a sensitivity analysis model of the control input to temperature and humidity, calculate the sensitivity coefficient, and use the sensitivity coefficient to weight the regression coefficient components to construct a weighted Euclidean distance. Use an affine propagation clustering algorithm based on weighted Euclidean distance to divide the regression coefficient matrix set, obtain the regression coefficient matrix of the stable operating condition cluster and the cluster center, and generate cluster quality evaluation parameters based on fermentation quality indicators to build a knowledge base. S3. In the online control phase, the temporary regression coefficient matrix is ​​estimated based on real-time data. The weighted Euclidean distance between the temporary regression coefficient matrix and the cluster core regression coefficient matrix contained in the knowledge base is calculated. The membership probability is calculated using the Softmax function. The membership probability is used as a weight to perform weighted fusion of the temperature and humidity prediction outputs of the models corresponding to each cluster core regression coefficient matrix to obtain the temperature and humidity prediction value sequence in the prediction time domain. The risk coefficient corresponding to the current working condition is calculated based on the membership probability and the cluster quality evaluation parameters. Construct a rolling optimization objective function and solve the rolling optimization objective function to obtain the control input of the fermentation equipment at the next time step.

2. The method for coordinated temperature and humidity control in a fermentation process according to claim 1, characterized in that, The process of acquiring and preprocessing historical data including temperature, humidity, control inputs, key fermentation state parameters, and fermentation quality indicators includes: using a cubic spline interpolation algorithm to align the historical data corresponding to different sampling frequencies in time sequence, so that the data of each batch are unified to the set standard time interval; extracting the global maximum and minimum values ​​of each parameter dimension to calculate the range, and normalizing all process data and fermentation quality indicators to the 0 to 1 range by subtracting the minimum value from the current parameter value and dividing by the range, thereby generating a set of standardized feature matrices.

3. The method for coordinated temperature and humidity control in a fermentation process according to claim 1, characterized in that, The process of solving for the set of regression coefficient matrices corresponding to historical sampling times includes: calculating the absolute difference between key fermentation state parameters at adjacent sampling times as the rate of change over time; smoothing the rate of change over time; normalizing the smoothed rate of change over time and mapping it to sample weights using a forward mapping function with an upper limit truncation; constructing a local weighted sum of squared errors loss function by combining sample weights, using temperature and humidity as joint prediction targets based on a local sliding time window, and using control inputs and historical temperature and humidity states as regression variables; constructing an elastic network objective function using L1 regularization and L2 regularization terms in the local weighted sum of squared errors loss function, minimizing the objective function using the coordinate descent method, and solving for the set of regression coefficient matrices.

4. The method for coordinated temperature and humidity control in a fermentation process according to claim 1, characterized in that, The process of constructing a sensitivity analysis model for temperature and humidity based on control input quantities and calculating sensitivity coefficients includes: constructing a feedforward neural network prediction model with the control input quantity sequence as the input node and temperature and humidity as the output nodes; training the network weights using historical data through a backpropagation algorithm until the network weights converge; calculating the partial derivatives of temperature and humidity with respect to each control input quantity based on the trained feedforward neural network prediction model; calculating the average value of the absolute values ​​of the partial derivatives of each control input quantity in the historical dataset and the time dimension of the input sequence; and normalizing the average value corresponding to each control input quantity to obtain the sensitivity coefficient for the corresponding control input quantity.

5. The method for coordinated temperature and humidity control in a fermentation process according to claim 3, characterized in that, The method of constructing a weighted Euclidean distance by weighting the regression coefficient components using sensitivity coefficients includes: for any two regression coefficient matrices, extracting the regression coefficient components at corresponding positions and calculating the square of the difference between the regression coefficient components; determining the weights according to the type of regression independent variable corresponding to the regression coefficient components, using the sensitivity coefficient associated with the corresponding control input as the weight for the regression coefficient components corresponding to the control input, and using the preset state weights as the weights for the regression coefficient components corresponding to historical temperature and historical humidity states; multiplying the determined weights by the square of the corresponding regression coefficient difference; summing all the weighted squared terms and taking the square root of the summation result to obtain the weighted Euclidean distance between the two regression coefficient matrices.

6. The method for coordinated temperature and humidity control in a fermentation process according to claim 1, characterized in that, The method employs an affine propagation clustering algorithm based on weighted Euclidean distance to partition the regression coefficient matrix set, obtaining stable operating condition clusters and cluster center regression coefficient matrices. A knowledge base is then constructed based on fermentation quality indicators to generate cluster quality evaluation parameters. This includes: using the negative value of the weighted Euclidean distance as the similarity metric matrix of the affine propagation clustering algorithm; setting the initial reference value as the median of the similarity metric matrix; iteratively transmitting attraction and affiliation information between data points in the regression coefficient matrix, updating the probability of a data point becoming a cluster center and its affiliation to the cluster center; using the cluster center nodes obtained from clustering as the cluster center regression coefficient matrix of the stable operating condition clusters; statistically analyzing the quality mean, quality fluctuation range, and quality compliance rate corresponding to historical samples within the stable operating condition clusters; and storing the statistically analyzed parameters as cluster quality evaluation parameters, associated with the cluster center regression coefficient matrix, in the knowledge base.

7. The method for coordinated temperature and humidity control in a fermentation process according to claim 1, characterized in that, The calculation of the weighted Euclidean distance between the temporary regression coefficient matrix and the cluster center regression coefficient matrix, using the Softmax function to calculate the membership probability, includes: calculating the weighted Euclidean distance between the currently estimated temporary regression coefficient matrix and all cluster center regression coefficient matrices in the knowledge base; selecting the N cluster center regression coefficient matrices with the smallest distances, and calculating the reciprocal of the weighted Euclidean distance between the temporary regression coefficient matrix and the selected N cluster center regression coefficient matrices respectively; multiplying each reciprocal distance by a set smoothing constant as an exponent, and substituting it into the natural exponential function to calculate the exponent value corresponding to each cluster center; dividing the exponent value corresponding to a single cluster center by the sum of the exponent values ​​corresponding to the N cluster centers to obtain the membership probability of the current working condition to the adjacent N working condition clusters.

8. The method for coordinated temperature and humidity control in a fermentation process according to claim 1, characterized in that, The construction of the rolling optimization objective function and the solution of the rolling optimization objective function to obtain the control input of the fermentation equipment at the next time step include: within the set prediction time domain, calculating the sum of squares of the deviations between the predicted temperature and humidity values ​​obtained by weighted fusion of the outputs of each cluster core model and the expected temperature and humidity setpoints; calculating the sum of squares of the penalty terms corresponding to the rate of change of control input within the prediction time domain; adjusting the weights of the temperature and humidity tracking error term and the penalty term for the change of control input according to the risk coefficient corresponding to the current operating condition; weightedly adding the sum of squares of the deviations after weight adjustment and the sum of squares of the penalty terms to construct a rolling optimization objective function containing upper and lower limits of hardware execution constraints; using a mixed integer quadratic programming algorithm to solve for the optimal control input sequence, and extracting the first element in the optimal control input sequence as the temperature and humidity coordinated control input applied to the fermentation equipment at the next time step.

9. The method for coordinated temperature and humidity control in a fermentation process according to claim 8, characterized in that, The risk coefficient is obtained by weighting one or more of the following: quality deviation, quality fluctuation, and quality non-compliance rate corresponding to the fermentation quality indicators.

10. A temperature and humidity coordinated control system for a fermentation process, characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement a temperature and humidity coordinated control method for a fermentation process according to any one of claims 1-9.

Images