Food fermentation environment regulation and control method and system based on Internet of Things

Abstract

The invention provides a food fermentation environment regulation and control method and system based on the Internet of Things, and the system employs a local and cloud cooperation architecture, and comprises a sensing module which is responsible for collecting fermentation environment data and various biochemical data in real time, and carrying out the preprocessing; a local unit in the decision-making module judges a current fermentation stage by using a stage recognition model constructed by a deep neural network, and matches a regulation and control strategy; the cloud unit analyzes historical high-quality fermentation data through a machine learning algorithm, and continuously optimizes and issues regulation and control strategies of all stages; and the execution module finally drives a temperature and humidity execution mechanism and the like to accurately regulate and control the fermentation environment. According to the invention, multi-modal sensing, intelligent decision making and precise control of the food fermentation process are realized, and the product quality, the production efficiency and the stability are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of IoT intelligent control and food industry automation, specifically to a method and system for regulating the food fermentation environment based on IoT. Background Technology

[0002] Traditional food fermentation processes rely heavily on human experience for environmental control, which generally suffers from problems such as lagging control, poor consistency, and insufficient stability. Existing technologies typically use single environmental parameters (such as temperature and humidity) for localized control, lacking real-time, multi-dimensional perception of the biochemical state of the fermenting material itself. This makes it impossible to accurately identify the fermentation stage, resulting in a mismatch between control strategies and the actual metabolic needs of microorganisms, affecting product quality and production efficiency.

[0003] In addition, existing control systems are mostly limited to local closed-loop control, lacking optimization mechanisms based on historical data and continuous learning capabilities. They are difficult to achieve adaptive adjustment under different batches and raw material conditions, and cannot integrate multiple objectives such as quality, energy consumption and efficiency for global optimization. Therefore, there is an urgent need for a fermentation environment control solution that can achieve real-time perception, intelligent judgment, precise control and has continuous evolution capabilities. Summary of the Invention

[0004] In order to solve the technical problems mentioned in the background art, the purpose of this invention is to provide a method and system for controlling the food fermentation environment based on the Internet of Things.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: An Internet of Things (IoT)-based food fermentation environment control system includes: Sensing module: Real-time acquisition of a first dataset including fermentation environmental data and biochemical data, wherein the biochemical data includes at least infrared spectral data, electronic nose data, and pH data; preprocessing the first dataset to generate a second dataset; Decision-making module: includes local decision-making units and cloud-based learning units; The local decision-making unit identifies the fermentation stage of the fermenting product based on the second dataset and a preset stage identification model, which is constructed based on a deep neural network; and matches a control strategy from a preset control strategy library based on the fermentation stage. The cloud-based learning unit collects historical high-quality fermentation process data and final quality evaluations, sets an optimization objective function, and generates control strategies for each fermentation stage by solving the optimization objective function. The optimization objective function is a multi-objective weighted comprehensive function used to quantify quality, efficiency, and production costs. The final quality evaluation includes objective indicators, subjective evaluations, and energy consumption data, and the control strategies are distributed to the control strategy library. Execution module: Based on the control strategy, the preset execution mechanism is driven to control the fermentation environment. The execution mechanism includes at least a heating device, a cooling device, a humidifying device, and a dehumidifying device.

[0006] Furthermore, the infrared spectral data are processed using standard normal variable transformation and Savitzky-Golay filter to obtain spectral features; The electronic nose data was normalized and its features were extracted.

[0007] Furthermore, the stage recognition model includes an input layer, an encoding layer, a fusion layer, an abstraction layer, and an output layer.

[0008] Furthermore, the input layer receives a fixed-length time-series window of data, which is constructed based on the second dataset; The encoding layer performs feature extraction and abstraction on the temporal window data, specifically as follows: The first abstract feature is obtained by performing linear transformation and nonlinear activation on the environmental data and pH data through a preset fully connected layer; The second abstract feature is obtained by capturing local patterns of the spectral features through a preset one-dimensional convolutional layer, as shown in the following formula:

[0009] in, The second abstract feature is the output of the one-dimensional convolutional layer. For spectral characteristics, For convolution operations, and For learnable convolutional kernel parameters and bias terms, It is a non-linear activation function; The electronic nose features are abstracted using a fully connected layer to obtain a third abstract feature; The fusion layer concatenates the first abstract feature, the second abstract feature, and the third abstract feature into a unified temporal feature, and performs temporal modeling on the temporal feature to obtain a behavioral feature vector; The abstraction layer consists of one or more fully connected layers, which further integrate the behavioral feature vectors and learn higher-order nonlinear combinations between features to generate more discriminative classification vectors for the final classification task. The calculation formula is as follows:

[0010] in, For behavioral feature vectors, and are the weights and biases of the fully connected layer, respectively, and are the trainable parameters of the fully connected layer; The output layer consists of a Softmax classifier, which maps the classification vector to the probability distribution of each fermentation stage. The calculation formula is as follows:

[0011] in, For the first The probability of each fermentation stage. The preset total number of stages, and These are trainable parameters; Ultimately, the stage with the highest probability was selected as the fermentation stage.

[0012] Furthermore, the cross-entropy loss function is used as the loss function for the stage recognition model, and the specific formula is as follows:

[0013] in, It is the number of training samples. It refers to the number of stages. For the first In the training samples, the th Are these fermentation stages truly fermentation stages? For the first In the training samples, the th The probability of each fermentation stage.

[0014] Furthermore, the steps for generating the control strategy are as follows: Construct a data package of the entire fermentation process for a high-quality fermentation case, including the second dataset and the final quality evaluation. A retrospective analysis was performed on the first dataset of the entire fermentation process of the high-quality fermentation case, and the start and end timestamps of each fermentation stage were marked. Subsequently, all timestamps corresponding to the fermentation stages were extracted from the complete second dataset. Data fragments constitute the fermentation stage dataset. ; The formula for the optimization objective function is as follows:

[0015] in, For regulation strategy Under the guidance of the fermentation stage The expected overall score, For regulation strategy Expected score in product quality For regulation strategy Expected score in terms of production efficiency For regulation strategy Expected score in production costs , and These are respectively quality weight, efficiency weight, and cost weight; The machine learning algorithm is used to solve the objective function to obtain the control strategy, and the control strategy is then distributed to the control strategy library.

[0016] Furthermore, the control strategy is expressed as the target value and deviation range of environmental parameters.

[0017] A method for controlling the food fermentation environment based on the Internet of Things, comprising: Step 1: Collect a first dataset in real time, including environmental data and biochemical data related to fermentation. The biochemical data includes at least infrared spectral data, electronic nose data, and pH data. Preprocess the first dataset to generate a second dataset. Step 2: Includes local decision-making units and cloud-based learning units; The local decision-making unit identifies the fermentation stage of the fermenting product based on the second dataset and a preset stage identification model, which is constructed based on a deep neural network; and matches a control strategy from a preset control strategy library based on the fermentation stage. The cloud-based learning unit collects historical high-quality fermentation process data and final quality evaluations, sets an optimization objective function, and generates control strategies for each fermentation stage by solving the optimization objective function. The optimization objective function is a multi-objective weighted comprehensive function used to quantify quality, efficiency, and production costs. The final quality evaluation includes objective indicators, subjective evaluations, and energy consumption data, and the control strategies are distributed to the control strategy library. Step 3: Based on the control strategy, drive the preset actuator to control the fermentation environment. The actuator includes at least a heating device, a cooling device, a humidifying device, and a dehumidifying device.

[0018] Compared with the prior art, the advantages of the present invention are as follows: 1. This invention overcomes the limitations of traditional methods that rely on a single environmental parameter and human experience by integrating environmental data and various biochemical data and using a deep neural network model for comprehensive analysis. It can more accurately and automatically identify the fermentation stage in real time, laying the foundation for precise control. 2. By adopting an architecture that combines local decision-making with cloud-based machine learning, this invention not only ensures the real-time nature of the control response, but also continuously iterates and optimizes the control strategies at each stage by analyzing high-quality historical data, thereby achieving the self-evolution of the system's decision-making capabilities and significantly improving the consistency and stability of the product. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of the system workflow of the present invention; Figure 2 This is a schematic diagram of the execution flow of the stage identification model of the present invention; Figure 3 This is a schematic diagram of the structure of the biochemical data of the present invention. Detailed Implementation

[0021] To achieve the above objectives, the present invention provides a food fermentation environment control system based on the Internet of Things (IoT). Please refer to the provided technical solution. Figure 1 The system includes: Sensing module: Real-time acquisition of a first dataset including fermentation environmental data and biochemical data, wherein the biochemical data includes at least infrared spectral data, electronic nose data, and pH data; preprocessing the first dataset to generate a second dataset; Environmental data is continuously and stably monitored by temperature and humidity sensors deployed in the fermentation environment. These sensors are accurate and have long-term stability, providing reliable direct parameters for environmental control. Biochemical data, including at least infrared spectroscopy data, electronic nose data, and pH data, are used to directly or indirectly monitor the bulk state and metabolic activities of the fermentation product through sensing technology. Specific data collection steps are as follows; please refer to [link / reference needed]. Figure 3 : Non-contact near-infrared (NIR) spectral sensor: This sensor is aimed at fermentation materials (such as mash or solid koji in a fermentation tank) in a non-contact manner. By emitting near-infrared light and receiving the diffuse reflectance spectrum after the material interacts with it, it indirectly monitors key biochemical parameters in real time. Through the established spectral analysis model, it can simultaneously track changes in sugar content, concentration of key metabolites, and the growth and decline trends of microbial biomass. This completely changes the lagging mode of traditional offline testing that relies on manual sampling, and realizes the essential, continuous, in-situ perception of the fermentation process. Electronic nose gas sensor array: This array consists of a group of gas sensors that are sensitive to and cross-responsive to specific volatile organic compounds (VOCs). It is placed in the headspace or exhaust port of the fermentation vessel to monitor the composition and concentration change patterns of the rich volatile flavor substances produced during fermentation, such as alcohols, esters, and acids. Its signal patterns can not only be used to determine the formation stage of flavor substances, but also, through pattern recognition technology, can keenly capture the off-odor characteristics produced by abnormal fermentation such as contamination by miscellaneous bacteria, thereby achieving early warning. pH sensor: For liquid or semi-liquid fermentation of soy sauce, yogurt, beer, etc., an in-situ washable (CIP) embedded pH sensor can be used to directly immerse the fermentation material and provide real-time and continuous measurement data of hydrogen ion concentration index as a key direct indicator reflecting the fermentation acidity process. All sensor nodes collect raw data via wireless communication protocols or wired connections, which are then aggregated to the edge computing node to form the initial first dataset. The specific steps of data preprocessing are as follows: For temperature, humidity and pH data, the preprocessing follows the standard signal processing procedure, which mainly includes outlier removal and smoothing filtering and data alignment. Outlier removal and smoothing filtering employ threshold-based and statistical methods (such as sliding window Z-score) to remove obvious outliers, and apply Kalman filtering or low-pass digital filtering algorithms to suppress high-frequency noise, ensuring the smoothness and continuity of the data. Data alignment unifies all sensor data onto the same timestamp sequence. Typically, the sensor with the highest data acquisition frequency is used as the benchmark, and other data are timestamp aligned and resampled to ensure the accuracy of subsequent multimodal data fusion. The raw near-infrared spectral signals of infrared spectral data are easily affected by ambient temperature, light path scattering, and sample physical state (such as particle size and density). Therefore, fine spectral preprocessing is required to extract effective chemical information. The specific operation is as follows: The Standard Normal Variable Transform (SNV) is used to eliminate baseline drift and multiplicative scattering effects caused by sample scattering and optical path variation. The calculation formula is as follows:

[0022] in, For the first k Absorbance values ​​after standard normal transformation at each wavenumber. For the first k The original absorbance values ​​at each wavenumber. and These are the mean and standard deviation of a single spectrum at all wavenumber points, respectively. After SNV processing, the mean of each spectral data becomes 0 and the standard deviation becomes 1; A Savitzky-Golay filter is used to perform polynomial least squares fitting within a preset moving window, while smoothing and differentiation operations are performed to ensure the original shape of the spectral signal. After the above preprocessing, the infrared spectral data has improved signal-to-noise ratio and effectiveness, and its spectral characteristics have been obtained. The raw signal from the electronic nose sensor is susceptible to fluctuations in ambient temperature and humidity, baseline drift of the sensor itself, and interference from non-target gases. Gas information is contained in the dynamic response pattern and cross-response spectrum of the electronic nose sensor array to gases. Therefore, preprocessing is necessary to eliminate interference and extract stable and highly discriminative features, thereby transforming the raw signal into a signal that effectively characterizes the volatile organic compound (VOC) components. The specific steps are as follows: Within each sampling period, the response value of each sensor is first normalized relative to its stable initial baseline value; Feature values ​​that represent both the response intensity and the dynamic process of the response are extracted from each response curve to form a feature vector, which serves as the electronic nose feature. Commonly used features include: Maximum response value :

[0023] in, This is the change in resistance, representing the difference between the sensor's resistance value and the baseline resistance value at a given point in time. This is the baseline resistance value. This is a function to find the maximum value. Average differential coefficient: The average value of the differential values ​​calculated during the rise phase of the response; Integral value (total response): For the entire response process... Integrate points; To eliminate the common impact of changes in ambient temperature and humidity on all sensors and to highlight the pattern differences brought about by different gas components, the feature vectors extracted by the entire sensor array at the same time will be normalized or transformed into a standard normal, which enhances the system's ability to recognize patterns of specific flavor stages or abnormal odors.

[0024] Finally, all the preprocessed sensor data (environmental data, preprocessed spectral feature values, biochemical parameter inversion results, and electronic nose feature vectors) are integrated and aligned according to a unified and precise timestamp to generate time series data points that constitute the second dataset.

[0025] The decision-making module includes a local decision-making unit and a cloud-based learning unit. The local decision-making unit identifies the fermentation stage of the fermenting product based on the second dataset and a preset stage identification model, which is constructed based on a deep neural network. Based on the fermentation stage, it matches control strategies from a preset control strategy library. The cloud-based learning unit collects historical high-quality fermentation process data and final quality evaluations, sets an optimization objective function, and generates control strategies for each fermentation stage by solving the optimization objective function. The optimization objective function is a multi-objective weighted comprehensive function used to quantify quality, efficiency, and production costs. The final quality evaluation includes objective indicators, subjective evaluations, and energy consumption data. The control strategies are then distributed to the control strategy library. The phase identification model is deployed locally; The core of the stage recognition model is a hierarchical fusion structure specifically designed for multimodal time series classification, including an input layer, encoding layer, fusion layer, abstraction layer, and output layer. (See also...) Figure 2 ; The input layer receives a time window of a second dataset of fixed length (time step T). The input is a two-dimensional data matrix, which is mathematically expressed as:

[0026] Each row represents a point in time (from...) arrive Each column represents data from a second dataset; The coding layer performs preliminary and targeted feature extraction and abstraction on input features with different physical meanings. The specific operations are as follows; For time-series data of environmental data and pH values, linear transformation and nonlinear activation are performed through a preset fully connected layer, which projects the data onto a higher-dimensional feature space to obtain the first abstract feature. For spectral features, a pre-defined one-dimensional convolutional layer is used to capture local patterns of the spectral shape, as shown in the following formula:

[0027] in, The second abstract feature is the output of the one-dimensional convolutional layer. For spectral characteristics, For convolution operations, and For learnable convolutional kernel parameters and bias terms, It is a non-linear activation function; For electronic nose features, feature abstraction is performed through a fully connected layer to obtain a third abstract feature that enhances the pattern discriminativeness of the electronic nose features. The fusion layer combines the first, second, and third abstract features into a unified temporal feature. and temporal characteristics Temporal modeling is performed to obtain a behavioral feature vector representing the dynamic behavior of the entire window; In this embodiment, a long short-term memory network is used to process temporal features. For time series modeling, the calculation can be simplified to:

[0028] in, Let be the temporal characteristics at time t. and These represent the hidden state and cell state at time t, respectively, carrying the information from the beginning of the temporal features to... t Contextual information at any given moment; The final output is the final hidden state of the entire input window. It is a behavioral feature vector that represents the dynamic behavior of the entire window; The abstraction layer consists of one or more fully connected layers, which further integrate behavioral feature vectors and learn higher-order nonlinear combinations between features to form more discriminative classification vectors for the final classification task. The calculation formula is as follows:

[0029] in, For behavioral feature vectors, and Here, represents the weights and biases of the fully connected layer, and represents the trainable parameters of the fully connected layer. It is a non-linear activation function.

[0030] The output layer consists of a Softmax classifier, which classifies the vectors. The mapping is represented by the probability distribution over each fermentation stage, and the calculation formula is as follows:

[0031] in, For the first The probability of each fermentation stage. The preset total number of stages, and These are trainable parameters; Ultimately, the stage with the highest probability was selected as the fermentation stage.

[0032] The construction of the stage identification model relies on massive amounts of high-quality labeled data. The specific steps are as follows: Based on historical data of the entire fermentation process, experts in the field precisely mark the start and end times of each fermentation stage on the timeline, including inflection points of key biochemical indicators, trends in environmental parameters, and the quality results of the final product. For example, experts will define the start of the main fermentation period based on the moment of a sharp drop in pH, the peak of CO2 release rate, or the appearance of specific flavor substances. For each piece of historical fermentation data, a sliding window is used to extract countless segments of length [missing information]. T The training sample set is constructed from time-series segments, and the label of each segment is the fermentation stage corresponding to the last moment of that segment, which is annotated by experts. Cross-entropy loss function The loss function for the stage identification model is defined by the following formula:

[0033] in, It is the number of training samples. It refers to the number of stages. For the first In the training samples, the th The value is set to 1 if the fermentation stage is a true fermentation stage, and 0 otherwise. For the first In the training samples, the th The probability of each fermentation stage; The training process of the stage recognition model uses the backpropagation algorithm and an optimizer (such as Adam) to iteratively adjust all parameters in the model, making the loss function... L It continues to decrease.

[0034] The regulation strategy library is a local, dynamic knowledge base whose initial content and subsequent updates rely on the powerful support of a cloud platform.

[0035] The steps for generating the control strategy are as follows: Construct a complete fermentation process data package for high-quality fermentation cases, including a second dataset and final quality evaluation; The final quality evaluation is a structured evaluation vector, consisting of objective indicators, subjective evaluation, and energy consumption data; Objective indicators refer to the concentration of key components of the product, yield, fermentation cycle time, etc. The concentrations of key components include, but are not limited to: For condiments such as soy sauce and fermented black beans: amino acid nitrogen, total nitrogen, reducing sugar, salt concentration, and total acid; For dairy products such as yogurt and cheese: live lactic acid bacteria count, lactic acid content, protein content, viscosity, and concentration of key flavor compounds such as acetaldehyde or diacetyl; For beer: original wort concentration, alcohol content, bitterness value, diacetyl content, and color; For yellow wine and rice wine: alcohol content, sugar content, total acidity, and amino acids.

[0036] These objective indicator data will be normalized to a uniform scoring range, such as 0-100 points; To compensate for the inability of objective indicators to fully cover sensory experiences such as flavor and texture, this invention introduces a structured subjective evaluation system and transforms it into calculable numerical values, as follows: Evaluation criteria: An evaluation panel of at least three trained senior tasters conducts blind evaluations of the products in a standardized environment, based on a pre-defined sensory evaluation form. The evaluation form typically includes the following dimensions and weights (taking soy sauce as an example): Color (weight 20%): dark brown, gloss, clarity; Aroma (weight 30%): intensity of soy sauce aroma, richness, and absence of any off-odors; Flavor (weight 50%): umami intensity, salty-sweet balance, lingering aftertaste, body; Each dimension is scored using a 5-point or 10-point Likert scale; Calculate the weighted score for each taster across all dimensions; Calculate the total score for each taster; Finally, the average of all the tasters' total scores is taken as the subjective evaluation score for that batch of products; Energy consumption data refers to the total energy consumption of the entire fermentation process or the energy consumption per unit of output; For each high-quality fermentation case, a retrospective analysis is performed on the first dataset of the entire fermentation process, accurately marking the start and end timestamps of each fermentation stage. Subsequently, from the complete second dataset, all data segments corresponding to fermentation stage i are extracted to constitute the fermentation stage. i Dataset ; The formula for optimizing the objective function is as follows:

[0037] in, For regulation strategy Under the guidance of the fermentation stage The expected overall score, For regulation strategy The expected score in product quality, through the fermentation stage i The correlation between certain key biochemical characteristics and the final quality assessment was obtained. For regulation strategy The expected score in production efficiency, through the fermentation stagei The shorter the fermentation time, the higher the expected score. For regulation strategy The expected score regarding production costs is obtained through energy consumption during the fermentation stage; the lower the energy consumption, the higher the expected score. , and These are quality weights, efficiency weights, and cost weights, which are pre-set by domain experts or production managers based on business objectives and product positioning, and serve as hyperparameters for optimizing the objective function. In this embodiment, algorithms such as inverse optimization or imitation learning are used to solve the objective function. The solution algorithm is applied to a large amount of high-quality fermentation stage data. The search aims to find a fixed control strategy. This allows the regulatory strategy to be applied to the fermentation stage. At that time, the goal is to achieve the highest overall score. The formula is as follows:

[0038] The final control strategy obtained will be distributed to the control strategy library.

[0039] Execution and feedback phase: Based on the control strategy, the preset actuators are driven to control the fermentation environment. The actuators include at least a heating device, a cooling device, a humidifying device, and a dehumidifying device. Control strategies are typically expressed as a series of target points for environmental parameters and their allowable fluctuation tolerances. Their general form can be expressed as:

[0040] in, and The control strategy is the fermentation stage. The set target temperature and target humidity, and This is the allowable deviation range from the target value; The system compares the control strategy with the current environmental data. Based on this difference, it generates specific control commands using preset proportional, integral, and derivative control algorithms, such as: For temperature control: if the current temperature is lower than the target temperature and not within the deviation range, a heating command is generated to drive the heating device (such as an infrared heating tube or a hot air blower) to work.

[0041] Example 1: In order to fully disclose the technical solution of the present invention and verify its beneficial effects, the implementation method of the present invention will be described in detail using the high-salt dilute fermentation production process of soy sauce as an example.

[0042] In this embodiment, the high-salt dilute-state fermentation process of soy sauce is divided into three distinct stages. The specific fermentation stages and the criteria for judgment are as follows: During the hydrolysis and acidification period, the protease and amylase secreted by Aspergillus oryzae violently hydrolyze the protein and starch in the raw materials into amino acids and sugars, while lactic acid bacteria begin to multiply and produce acid. The classification is based on the following criteria: The pH value dropped rapidly from the initial 6.5-7.0 to 5.2-5.5, the ammonia nitrogen concentration increased by more than 0.05% per day, and the electronic nose signal was mainly raw material and enzymatic odor; During the primary fermentation and esterification period, salt-tolerant yeasts and lactic acid bacteria become the dominant microbial population, carrying out alcoholic fermentation and lactic acid fermentation, and initiating esterification reactions to generate aroma substances. The classification is based on the following criteria: Due to the heat released during fermentation, the temperature of the tank spontaneously rises by 2-4°C, requiring external cooling to maintain the set temperature. The electronic nose response pattern undergoes a significant shift, with the characteristic signal intensities of alcohols and acids reaching peak values; Sugar content continued to decrease, while ethanol concentration steadily increased; During the post-ripening and flavor fusion period, the intense biochemical reactions slow down, and the long process of flavor substance synthesis and maturation begins, such as the Maillard reaction.

[0043] The classification is based on the following criteria: The changes in environmental parameters tend to stabilize; The electronic nose detected a significant increase in the signal intensity of complex aroma substances such as esters, which then tended to stabilize. The spectral change rate of each key biochemical parameter was less than 0.5% per day, indicating that the reaction had entered a plateau phase.

[0044] In this embodiment, the system is deployed in a 10-ton soy sauce fermentation tank.

[0045] The tank is equipped with a PT100 temperature sensor and a capacitive humidity sensor; A clean-in-situ (CIP) antimony electrode pH sensor was immersed in the mash. An online near-infrared spectroscopy probe with a wavelength range of 900-1700 nm was installed on the side wall of the tank and scanned every 4 hours. An electronic nose system containing 16 metal oxide sensors was deployed at the top exhaust port of the tank to sample headspace gas every 8 hours. The phase identification model is a pre-trained deep neural network whose input is a time-series window consisting of all sensor data from the past 48 hours.

[0046] When the stage identification model transitions to a new stage, the corresponding control strategy is immediately called from the local strategy library. For example, after entering the main fermentation and esterification period, the control strategy adjusts the temperature target from 28℃ to 30℃ and allows a deviation of ±0.5℃. Based on this strategy, the execution module drives the chiller or heating element to operate in order to precisely maintain the target temperature.

[0047] A six-month comparative experiment was conducted using the system of this invention and manual control to quantify the technical effects of this invention. The experimental group consisted of three fermentation tanks that were fully automated using the system of this invention, while the control group also consisted of three fermentation tanks that were manually controlled by senior engineers based on traditional experience. All batches used soy sauce koji from the same source. Key data comparisons are as follows: Table 1: Comparison and Analysis of the Effects of the Invention System and Traditional Manual Control Evaluation Dimensions Performance indicators Traditional manual control (control group) This invention system control (experimental group) Improvement Effects and Analysis Product Quality Amino acid nitrogen (g / 100mL) 0.83 ±0.12 0.95 ± 0.03 The core indicators improved by 14.5%, and batch stability (standard deviation) improved by 4 times, solving the core pain point of large quality fluctuations. Product Quality Sensory blind rating score (out of 100) 80 ± 6 92 ± 2 The system achieved higher overall scores for flavor, aroma, and color, with excellent consistency, demonstrating the effectiveness of its flavor optimization. Production efficiency Complete the total fermentation cycle (days) 125 ± 10 108 ± 3 The fermentation cycle is shortened by an average of 13.6%, and the cycle is more controllable, which is conducive to production planning. Economic benefits Energy consumption per unit output (kWh / kg) 1.55 ±0.25 1.18 ± 0.08 Energy consumption was reduced by 23.9%, achieving energy conservation and consumption reduction. Economic benefits Product quality rate 82% 99% The rate of high-quality products has increased significantly, directly reducing the loss of defective products and resulting in significant economic benefits. A method for controlling the food fermentation environment based on the Internet of Things, comprising: Step 1: Collect a first dataset in real time, including environmental data from fermentation and biochemical data for indirectly monitoring microbial community activity or metabolite concentration. The biochemical data includes at least infrared spectral data, electronic nose data, and pH data. Preprocess the first dataset to generate a second dataset. Step 2: Includes local decision-making units and cloud-based learning units; The local decision-making unit identifies the fermentation stage of the fermenting product based on the second dataset and a preset stage identification model, which is constructed based on a deep neural network; and matches a control strategy from a preset control strategy library based on the fermentation stage. The cloud-based learning unit collects historical high-quality fermentation process data and final quality evaluations, sets an optimization objective function, and generates control strategies for each fermentation stage by solving the optimization objective function. The optimization objective function is a multi-objective weighted comprehensive function used to quantify quality, efficiency, and production costs. The final quality evaluation includes objective indicators, subjective evaluations, and energy consumption data, and the control strategies are distributed to the control strategy library. Step 3: Based on the control strategy, drive the preset actuator to control the fermentation environment. The actuator includes at least a heating device, a cooling device, a humidifying device, and a dehumidifying device.

[0048] In summary, this invention comprehensively considers multiple key indicators such as final product quality, production efficiency, and energy consumption costs when formulating control strategies. By generating strategies through a multi-objective weighted optimization algorithm, it overcomes the shortcomings of traditional methods that only focus on a single objective, and achieves a balance between economic benefits and product quality.

[0049] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A food fermentation environment control system based on the Internet of Things, the system being divided into local and cloud components, characterized in that, include: Sensing module: Real-time acquisition of a first dataset including fermentation environmental data and biochemical data, wherein the biochemical data includes at least infrared spectral data, electronic nose data and pH data; The first dataset is preprocessed to generate the second dataset; Decision-making module: includes local decision-making units and cloud-based learning units; The local decision-making unit identifies the fermentation stage of the fermenting product based on the second dataset and a preset stage identification model, wherein the stage identification model is constructed based on a deep neural network; Based on the fermentation stage, a control strategy is matched from a preset control strategy library; The cloud-based learning unit collects historical high-quality fermentation process data and final quality evaluations, sets an optimization objective function, and generates control strategies for each fermentation stage by solving the optimization objective function. The optimization objective function is a multi-objective weighted comprehensive function used to quantify quality, efficiency, and production costs. The final quality evaluation includes objective indicators, subjective evaluations, and energy consumption data, and the control strategies are distributed to the control strategy library. Execution module: Based on the control strategy, the preset execution mechanism is driven to control the fermentation environment. The execution mechanism includes at least a heating device, a cooling device, a humidifying device, and a dehumidifying device.

2. The system according to claim 1, characterized in that, The infrared spectral data were processed using standard normal variable transformation and Savitzky-Golay filter to obtain spectral features; The electronic nose data was normalized and its features were extracted.

3. The system according to claim 2, characterized in that, The stage recognition model includes an input layer, an encoding layer, a fusion layer, an abstraction layer, and an output layer.

4. The system according to claim 3, characterized in that, The input layer receives a fixed-length time-series window of data, which is constructed based on the second dataset. The encoding layer performs feature extraction and abstraction on the temporal window data, specifically as follows: The first abstract feature is obtained by performing linear transformation and nonlinear activation on the environmental data and pH data through a preset fully connected layer; The second abstract feature is obtained by capturing local patterns of the spectral features through a preset one-dimensional convolutional layer, as shown in the following formula: ； in, The second abstract feature is the output of the one-dimensional convolutional layer. For spectral characteristics, For convolution operations, and For learnable convolutional kernel parameters and bias terms, It is a non-linear activation function; The electronic nose features are abstracted using a fully connected layer to obtain a third abstract feature; The fusion layer concatenates the first abstract feature, the second abstract feature, and the third abstract feature into a unified temporal feature, and performs temporal modeling on the temporal feature to obtain a behavioral feature vector; The abstraction layer consists of one or more fully connected layers, which further integrate the behavioral feature vectors and learn higher-order nonlinear combinations between features to generate more discriminative classification vectors for the final classification task. The calculation formula is as follows: ； in, For behavioral feature vectors, and are the weights and biases of the fully connected layer, respectively, and are the trainable parameters of the fully connected layer; The output layer consists of a Softmax classifier, which maps the classification vector to the probability distribution of each fermentation stage. The calculation formula is as follows: ； in, For the first The probability of each fermentation stage. The preset total number of stages, and These are trainable parameters; Ultimately, the stage with the highest probability was selected as the fermentation stage.

5. The system according to claim 4, characterized in that, Cross-entropy loss function The loss function for the stage identification model is defined by the following formula: ； in, It is the number of training samples. It refers to the number of stages. For the first In the training samples, the th Are these fermentation stages truly fermentation stages? For the first In the training samples, the th The probability of each fermentation stage.

6. The system according to claim 5, characterized in that, The steps for generating the control strategy are as follows: Construct a data package of the entire fermentation process for a high-quality fermentation case, including the second dataset and the final quality evaluation. A retrospective analysis was performed on the first dataset of the entire fermentation process of the high-quality fermentation case, and the start and end timestamps of each fermentation stage were marked. Subsequently, all data fragments corresponding to fermentation stage i were extracted from the complete second dataset to form the fermentation stage dataset. ; The formula for the optimization objective function is as follows: ； in, For regulation strategy Under the guidance of the fermentation stage The expected overall score, For regulation strategy Expected score in product quality For regulation strategy Expected score in terms of production efficiency For regulation strategy Expected score in terms of production costs , and These are respectively quality weight, efficiency weight, and cost weight; The optimization objective function is solved using a machine learning algorithm to obtain the regulation strategy, and the regulation strategy is then distributed to the regulation strategy library.

7. The system according to claim 1, characterized in that, The control strategy is expressed as the target value and deviation range of environmental parameters.

8. A method for controlling the food fermentation environment based on the Internet of Things, implemented by the system described in any one of claims 1-7, characterized in that, include: Step 1: Collect the first dataset in real time, including environmental data and biochemical data of fermentation. The biochemical data includes at least infrared spectral data, electronic nose data, and pH data. The first dataset is preprocessed to generate the second dataset; Step 2: Includes local decision-making units and cloud-based learning units; The local decision-making unit identifies the fermentation stage of the fermenting product based on the second dataset and a preset stage identification model, wherein the stage identification model is constructed based on a deep neural network; Based on the fermentation stage, a control strategy is matched from a preset control strategy library; The cloud-based learning unit collects historical high-quality fermentation process data and final quality evaluations, sets an optimization objective function, and generates control strategies for each fermentation stage by solving the optimization objective function. The optimization objective function is a multi-objective weighted comprehensive function used to quantify quality, efficiency, and production costs. The final quality evaluation includes objective indicators, subjective evaluations, and energy consumption data, and the control strategies are distributed to the control strategy library. Step 3: Based on the control strategy, drive the preset actuator to control the fermentation environment. The actuator includes at least a heating device, a cooling device, a humidifying device, and a dehumidifying device.

Images