Artificial intelligence-based deer wine production process optimization control method

Abstract

The invention relates to the technical field of brewing process intelligent control, and discloses an artificial intelligence-based deer wine production process optimization control method. The method comprises the following steps: configuring a multi-source sensing unit array integrated with a thermosensitive sensor, a humidity-sensitive sensor and an image collector, and synchronously collecting temperature, humidity and appearance data in a fermentation process; carrying out multi-modal feature fusion through a machine learning fusion engine, and generating a production state evaluation vector; constructing a virtual simulation platform for operating a physical mechanism model, and performing simulation deduction based on the vector to generate a simulation control signal; deploying an edge computing control grid, and generating a control strategy in real time by each node according to local data; and integrating the information through the data exchange middleware, and analyzing and optimizing the control instruction stream. According to the method, interactive optimization of digital simulation and a physical process and cooperative control of center planning and edge execution are realized, and the accuracy, the adaptability and the product stability of production process control are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for brewing processes, specifically to an artificial intelligence-based optimization control method for deer-based wine production processes. Background Technology

[0002] In traditional deer tuna brewing, the control of the fermentation process relies primarily on manual experience or basic automated instruments. Operators manually adjust fermentation parameters by periodically reading scattered thermometer and humidity readings, combined with subjective observations of the appearance and odor of the fermenting materials. This method is highly dependent on the operator's personal experience, resulting in inconsistent judgment standards and delayed responses, making it difficult to achieve precise and standardized production.

[0003] Existing automated monitoring systems typically perform closed-loop control on a single variable, lacking the collaborative perception and comprehensive analysis of multi-dimensional information such as humidity and changes in the macroscopic morphology of materials. These conventional control methods, based on preset fixed logic, cannot adapt to the nonlinear, time-varying characteristics and batch-to-batch material differences inherent in the complex biochemical reaction of fermentation. When production conditions are disturbed or raw material properties change, the fixed-program control strategy struggles to make proactive optimization adjustments, potentially causing the fermentation process to deviate from its optimal trajectory and affecting the stability and quality of the final product.

[0004] With the development of Industrial Internet of Things (IIoT) technology, solutions have emerged that upload sensor data to the cloud for centralized processing and monitoring. However, this centralized architecture presents risks such as data transmission latency, high network dependence, and untimely response to local emergencies, which are unsuitable for real-time control of fermentation processes. While edge computing has been introduced into industrial control, in fermentation optimization scenarios, it is typically used only for preliminary data filtering and uploading, and has not formed a distributed control mesh capable of autonomous strategy derivation and collaboration with centralized optimization decisions. The lack of an effective coordination mechanism between the centralized computing model and the rapid response capabilities of the edge leads to a disconnect between global optimization goals and local real-time control actions, making it difficult to balance the immediacy of control with overall optimality. Summary of the Invention

[0005] The purpose of this invention is to provide an artificial intelligence-based optimization and control method for the deer-based winemaking process, in order to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides an artificial intelligence-based optimization and control method for deer-based winemaking production processes, the method comprising: A multi-source sensing unit array is configured, which integrates a thermal sensor, a humidity sensor and an image acquisition device to simultaneously capture temperature readings, humidity readings and appearance change data during the deer wine fermentation process. A machine learning fusion engine is applied to perform multimodal alignment and feature-level fusion on the temperature readings, humidity readings, and appearance change data, and outputs a production status assessment vector. A virtual production simulation platform is constructed. After loading the production status evaluation vector, the virtual production simulation platform runs a physical mechanism model to generate a simulated control signal sequence. Design an edge computing control grid, which consists of multiple edge nodes, each edge node derives a set of control strategies in real time based on local data streams; A data exchange middleware is established, which integrates the production status evaluation vector and the control strategy set to parse out the optimized control command flow.

[0007] Preferably, the application of the machine learning fusion engine performs multimodal alignment and feature-level fusion on the temperature readings, humidity readings, and appearance change data, outputting a production status evaluation vector, including: The system receives raw data streams from the multi-source sensing unit array and performs sliding window segmentation on the raw data streams to form time-series data blocks. Wavelet transform algorithm is used to estimate the noise power spectrum of each time series data block to determine the adaptive filtering threshold; A digital filter bank is designed based on the adaptive filtering threshold, and frequency domain filtering is performed on the time series data block to obtain a clean data block. The time-domain and frequency-domain features of the clean data block are extracted and fused by a convolutional neural network to generate the production status evaluation vector.

[0008] Preferably, the step of extracting time-domain and frequency-domain features from the clean data block, and then performing weighted feature fusion using a convolutional neural network to generate the production status assessment vector includes: The mean, variance, and peak values ​​of each clean data block are calculated as a time-domain feature set; The clean data block is converted into a frequency domain representation by applying Fast Fourier Transform, and the dominant frequency and energy distribution are extracted as a frequency domain feature set. Construct a multi-branch convolutional neural network structure, where the first branch processes the time-domain feature set and the second branch processes the frequency-domain feature set; The attention mechanism is used to dynamically adjust the feature weights of the first and second branches to perform cross-modal feature concatenation. The concatenated features are mapped to a fixed-dimensional production status evaluation vector through a fully connected layer.

[0009] Preferably, the step of dynamically adjusting the feature weights of the first and second branches using an attention mechanism to perform cross-modal feature concatenation includes: Calculate the channel attention map on the output feature map of the first branch to generate temporal feature weight coefficients; Calculate the spatial attention map on the output feature map of the second branch to generate frequency domain feature weight coefficients; The time-domain feature weight coefficients and frequency-domain feature weight coefficients are normalized to obtain the fused weight matrix; The feature maps of the first and second branches are weighted and summed according to the fusion weight matrix to complete the cross-modal feature splicing.

[0010] Preferably, the construction of the virtual production simulation platform includes: Collect equipment and process parameters of the deer wine production line and establish a three-dimensional geometric model; An actuator interface module is embedded in the three-dimensional geometric model to connect to the actual actuator and inject control signals; A multi-channel signal synthesizer is configured, which is capable of generating analog temperature curves, analog humidity curves, and analog image sequences; Install a remote control agent to receive external control commands and forward them to the simulation platform. A unified simulation environment is formed by integrating a three-dimensional geometric model, an actuator interface module, a multi-channel signal synthesizer, and a remote control agent, serving as the virtual production simulation platform.

[0011] Preferably, after loading the production status evaluation vector, the virtual production simulation platform runs a physical mechanism model to generate a simulated control signal sequence, including: The production status assessment vector is mapped to the initial condition parameters of the physical mechanism model; Solve the partial differential equations in the physical mechanism model to obtain the temperature and concentration field distributions inside the fermenter; The control deviation index is calculated based on the temperature field and concentration field distribution, and the preset control rule library is queried according to the control deviation index. The matching control action sequence is retrieved from the control rule base and encoded as the analog control signal sequence.

[0012] Preferably, the method for constructing the physical mechanism model includes: Geometric structural parameters and physical property parameters of deer wine fermentation tanks are collected. The geometric structural parameters include tank volume and internal component layout, and the physical property parameters include material thermal conductivity and specific heat capacity. Based on the laws of conservation of energy and mass, a set of partial differential equations is established to describe the temperature distribution and reactant concentration distribution inside the fermenter. The partial differential equation system is spatially discretized, and the computational domain is divided into uniform grid nodes using the finite difference method. The initial state parameters and boundary condition parameters of the fermentation process are set. The initial state parameters are based on the measured temperature value at the time of feeding, and the boundary condition parameters include the heat convection coefficient of the tank wall and the ambient temperature value. Based on the initial state parameters and boundary condition parameters, the parameters in the partial differential equation system are iteratively corrected using historical fermentation operation data.

[0013] Preferably, the design edge computing control grid includes: Measure the spatial dimensions and equipment layout of the deer-based winemaking production area to determine the coverage radius and deployment density of edge nodes; Calculate the optimal set of node locations based on coverage radius and deployment density, and install edge computing devices at each location; The edge computing control mesh is constructed by configuring network addresses and communication protocol stacks for each edge computing device and building a mesh topology.

[0014] Preferably, each edge node derives a set of control strategies in real time based on its local data stream, including: Set up a data buffer at each edge node to continuously collect local sensor data streams; Use time series forecasting algorithms to predict data trends over a future period and generate forecast data sequences; The predicted data sequence is input into the policy optimization algorithm, which evaluates the utility value of different control actions based on the reward function. The control strategy set is composed of the control actions with the highest utility values.

[0015] Preferably, the establishment of the data exchange middleware, through which the production status evaluation vector and the control strategy set are integrated, and the optimized control command flow is parsed, includes: Deploy a message queue service to receive production status evaluation vectors from the machine learning fusion engine and a set of control policies from the edge computing control mesh; Multi-objective optimization calculations are performed on the production status assessment vector and the control strategy set to generate a candidate instruction set; Verify the feasibility and consistency of the candidate instruction set, eliminate conflicting instructions, and package them into the optimized control instruction stream.

[0016] Compared with the prior art, the beneficial effects of the present invention are: By employing a machine learning fusion engine to perform feature-level fusion of multi-source heterogeneous sensor data, a feature vector accurately characterizing the current fermentation state is generated and loaded into a physical mechanism-based virtual production simulation platform. Leveraging this understanding of the inherent biochemical laws of the fermentation process, various control strategies are simulated and their effects predicted in the digital space, resulting in near-optimal simulated control signals. This deep integration of data-driven front-end sensing and model-driven back-end simulation ensures that control commands no longer rely solely on learning from historical data but originate from adherence to and advanced calculations of the process's inherent mechanisms. It effectively overcomes the decision-making failures that may occur with single data-driven models under process variations or unseen operating conditions, and compensates for control biases caused by parameter mismatches in pure mechanistic models, thereby improving the predictability, adaptability, and accuracy of complex fermentation process control.

[0017] The design employs a control mesh composed of multiple edge nodes. Each node generates a set of control strategies in real time based on the local data stream of its assigned region, and integrates these strategies with optimization instructions from the central virtual simulation platform via data exchange middleware. This architecture rationally decomposes and coordinates control tasks. Edge nodes focus on handling high real-time, low-latency local control tasks, ensuring rapid response to unexpected situations on the production floor and basic autonomy in abnormal situations such as network outages. The central platform performs slower but more refined simulations and planning from a global and full-cycle production optimization perspective. The middleware analyzes and fuses information from both ends, achieving dynamic unification between local real-time control actions and global optimal goals. This distributed control model ensures system robustness and real-time performance while providing a structural foundation for dynamic optimization throughout the entire process, avoiding the response bottlenecks and single-point-of-failure risks that centralized processing may bring. Attached Figure Description

[0018] Figure 1 This is a schematic diagram illustrating the working principle of the artificial intelligence-based deer winemaking process optimization and control method described in this invention. Figure 2 A flowchart for data processing in a machine learning fusion engine; Figure 3 This is a flowchart of feature extraction and fusion. Figure 4 Timing diagram of monitoring and control signals during the fermentation process of deer winemaking; Figure 5 This is a simulation diagram of the temperature and concentration field distribution inside the fermenter. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1 This invention provides an artificial intelligence-based optimization control method for the deer winemaking process. The overall implementation scheme is as follows: A multi-source sensing unit array is configured for data acquisition during the deer winemaking fermentation process. This array integrates a thermal sensor, a humidity sensor, and an image acquisition device, capable of simultaneously capturing temperature readings, humidity readings, and appearance change data. A machine learning fusion engine receives the raw data from the multi-source sensing unit array, performs multimodal alignment and feature-level fusion processing, and outputs a production status evaluation vector. A virtual production simulation platform loads the production status evaluation vector and runs a physical mechanism model to generate a simulated control signal sequence. An edge computing control grid consists of multiple edge nodes, each of which derives a set of control strategies in real time based on its local data stream. A data exchange middleware integrates the production status evaluation vector and the control strategy set, parsing out the optimized control command stream, thereby achieving closed-loop optimized control of the deer winemaking process.

[0021] Example 1: See Figure 2In specific implementation, the machine learning fusion engine receives raw data streams from a multi-source sensing unit array. These raw data streams include temperature readings captured by a thermal sensor, humidity readings captured by a humidity sensor, and appearance change data captured by an image acquisition unit. The raw data stream is then segmented using a sliding window. The size of the sliding window is set according to the typical cycle of the deer wine fermentation process, and the window width covers the duration of a complete fermentation sub-stage. The window sliding step size is set to one-quarter of the window width, forming time-series data blocks. The length of each time-series data block is uniformly set to a fixed number of sampling points to ensure temporal alignment in subsequent processing. In some embodiments, the sliding window segmentation process considers the continuity of the data stream and employs an overlapping window strategy to increase the sample density of the data blocks. The size of the overlapping region is adjusted according to the dynamic characteristics of the fermentation process. The time-series data blocks are stored in a circular buffer for sequential processing by subsequent modules. In the specific implementation, a wavelet transform algorithm is used to estimate the noise power spectrum of each time series data block. The Db4 wavelet is selected as the basis function, and the number of decomposition levels is set to 5. By calculating the energy distribution of the detail coefficients of each level after wavelet decomposition, the power of noise in different frequency bands is estimated, and an adaptive filtering threshold is determined. The adaptive filtering threshold is dynamically adjusted according to the peak value of the noise power spectrum. The threshold calculation function adopts a soft thresholding rule to reduce signal distortion. It can be understood that the implementation of the wavelet transform algorithm involves discrete wavelet transform calculation, using the Mallat algorithm for fast decomposition, and the noise power spectrum estimation results are used to guide the setting of filtering parameters. In practical implementation, a digital filter bank is designed based on an adaptive filtering threshold. This digital filter bank includes a low-pass filter and a high-pass filter. The cutoff frequency of the low-pass filter is determined based on the distribution of low-frequency noise in the noise power spectrum, while the cutoff frequency of the high-pass filter is set to address high-frequency interference. The filter type is an FIR filter, and its order is calculated based on passband ripple and stopband attenuation requirements. Frequency domain filtering is performed on the time-series data blocks. Frequency domain filtering transforms the data blocks to the frequency domain using a Fast Fourier Transform (FFT), and then, after applying the filter transfer function, it is inversely transformed back to the time domain to obtain clean data blocks. Optionally, the digital filter bank design considers phase linearity requirements and employs zero-phase filtering technology to avoid group delay. This significantly improves the signal-to-noise ratio of the clean data blocks, laying the foundation for feature extraction.

[0022] In specific implementation, time-domain and frequency-domain features are extracted from the clean data block. Time-domain features include mean, variance, and kurtosis. The mean is calculated using the arithmetic mean method, the variance uses an unbiased estimation formula, and the kurtosis is obtained by identifying local extrema within the data block and calculating their deviation from the mean. Frequency-domain features are transformed into a frequency-domain representation using a Fast Fourier Transform (FFT) with 1024 points. Dominant frequencies and energy distributions are extracted. The dominant frequency corresponds to the position of the maximum value in the frequency-domain amplitude spectrum, and the energy distribution is calculated by summing the squares of each frequency component and normalizing it to the 0-1 range. In some embodiments, the time-domain feature set also includes skewness and kurtosis indices to describe the shape of the data distribution, and the frequency-domain feature set is supplemented with bandwidth and center frequency indices to enhance feature expressiveness. In practical implementation, feature weighting fusion is performed using a convolutional neural network (CNN). The CNN is constructed as a dual-input branch structure: the first branch receives the temporal feature set, and the second branch receives the frequency domain feature set. Each branch contains convolutional layers, batch normalization layers, and ReLU activation layers. The convolutional kernel size is uniformly 3x1, and the stride is set to 1. Feature weighting fusion is achieved through an attention mechanism. The attention module calculates the weight coefficients of each feature channel, and the weighted sums generate a fused feature map. The fused feature map is mapped to a fixed-dimensional production state evaluation vector through a fully connected layer. The CNN is trained using supervised learning, with cross-entropy loss as the loss function and stochastic gradient descent as the optimizer. The production state evaluation vector is set to 128 dimensions, representing the comprehensive score of the fermentation state. Optionally, residual connections are introduced in the feature weighting fusion stage to avoid gradient vanishing and improve network training stability.

[0023] In practical implementation, the specific parameter settings for the sliding window segmentation process rely on the technological knowledge of deer wine fermentation. The duration of each fermentation sub-stage is obtained through historical data statistical analysis. The window width is typically set to several minutes to tens of minutes, corresponding to the sampling interval of temperature readings, humidity readings, and appearance change data. The sliding window segmentation algorithm runs in real time on an embedded processor, with data block sequences cached in memory to ensure the continuity of the processing pipeline. In some embodiments, the size of the sliding window can be dynamically adjusted according to the fermentation stage. For example, a smaller window is used when changes are drastic in the early stages of fermentation, and a larger window is used when things are stable in the later stages. The window adjustment logic is based on the real-time calculated data variance value, and the window width is automatically reduced when the variance exceeds a threshold. In practical implementation, when implementing the wavelet transform algorithm, the selection of the Db4 wavelet basis function is based on its tight support and regularity, which is suitable for processing non-stationary sensor signals. The determination of the number of decomposition layers (5) considers the trade-off between computational complexity and frequency resolution. Noise power spectrum estimation is achieved by calculating the variance of the wavelet coefficients of each layer. The adaptive filtering threshold adopts a general threshold estimation method, and the threshold calculation formula is: the threshold equals the noise standard deviation multiplied by the square root of twice the logarithm of the signal length. It is understandable that the wavelet transform calculation uses the pyramid algorithm to reduce computational overhead, and the noise power spectrum estimation results are visualized for parameter tuning. In specific implementations, the design details of the digital filter bank include the selection of filter response types; the low-pass filter adopts a Chebyshev Type I design with passband ripple controlled within 0.1dB; the high-pass filter adopts an elliptic filter design with stopband attenuation greater than 40dB; the cutoff frequency is calculated based on the 3dB point of the noise power spectrum; frequency domain filtering is implemented through FFT and IFFT, with the number of FFT points matching the data block length; and the filtering operation is performed in the frequency domain using dot multiplication to avoid the computational burden of time-domain convolution. Optionally, the digital filter bank can be expanded into a multi-band filter to handle more complex noise environments, such as adding a band-stop filter to eliminate power frequency interference.

[0024] In specific implementations, during the time-domain feature extraction process, the mean is calculated by summing all sampling points in the data block and then dividing by the number of points. The variance is calculated using the standard deviation formula. The peak value is obtained by finding the point with the largest absolute value in the data block and calculating its ratio to the mean. In the frequency-domain feature extraction, a Hanning window is applied to the data block before the Fast Fourier Transform to reduce spectral leakage. The dominant frequency is identified as the frequency value corresponding to the global maximum value in the amplitude spectrum. The energy distribution is calculated as the sum and normalization of the squares of the amplitudes at each frequency point. In some embodiments, the time-domain features may also include a zero-crossing rate index, reflecting the signal oscillation frequency, and the frequency-domain features may include a spectral entropy index, describing the complexity of the frequency distribution. In specific implementations, the structural parameters of the convolutional neural network are as follows: the first branch input time-domain feature set has a 3-dimensional dimension corresponding to the mean, variance, and peak value; the second branch input frequency-domain feature set has a 2-dimensional dimension corresponding to the dominant frequency and energy distribution; the number of channels in the output feature map of the convolutional layer is set to 32; the number of neurons in the fully connected layer is set to 256; the attention mechanism uses global average pooling to generate channel weights; and the output production state evaluation vector is obtained after feature weighted fusion. It's understandable that convolutional neural networks are implemented using deep learning frameworks like TensorFlow, with model parameters trained on historical fermentation data. Optionally, feature-weighted fusion can be replaced with other fusion strategies such as concatenation or additive fusion, but the attention mechanism offers the advantage of dynamic weight adjustment.

[0025] In practical implementation, the entire machine learning fusion engine's software implementation is based on a multi-threaded architecture. The data stream processing thread is responsible for sliding window segmentation and wavelet transform, the feature extraction thread runs feature calculation algorithms, and the convolutional neural network thread performs model inference. Threads synchronize via message queues to ensure real-time performance. In some embodiments, the engine is deployed on an industrial computer, supporting online learning to update model parameters and adapt to fermentation process drift. In this implementation, the output format of the production status evaluation vector is a floating-point array, transmitted to the virtual production simulation platform via shared memory or a network interface. Vector elements are normalized to the 0-1 range for easy use by subsequent models. It is understood that performance optimization of the machine learning fusion engine involves algorithm acceleration, such as using GPUs to accelerate convolution operations and reduce processing latency. Optionally, the engine can integrate an anomaly detection module to identify anomalous data blocks during the feature extraction stage, preventing contamination of the fusion results.

[0026] In practical implementation, the sliding window segmentation algorithm uses a circular buffer to manage the data flow, updating the buffer pointer when new data arrives. Window triggering conditions are based on timestamps or sampling point counts to ensure correct data block timing. In some embodiments, window parameters can be configured to support flexible adjustments for different fermentation recipes. In practical implementation, the wavelet transform algorithm's computational optimization employs an enhancement scheme to reduce memory usage. Noise power spectrum estimation results are displayed in real-time for monitoring, and the adaptive filtering threshold is updated once per data block to adapt to non-stationary noise. It can be understood that the wavelet transform boundary handling uses a symmetrical extension method to avoid boundary effects. In practical implementation, the digital filter bank is implemented using iterative calculations with difference equations, and coefficients are stored in a lookup table. The FFT calculation for frequency domain filtering uses library functions to ensure accuracy, and the quality of clean data blocks is evaluated online using the signal-to-noise ratio (SNR) metric. Optionally, the filter bank can be designed as an adaptive filter, automatically updating coefficients based on noise statistics.

[0027] In specific implementations, the computation of time-domain and frequency-domain features is parallelized, utilizing multi-core processors to simultaneously compute the mean and FFT. Peak index detection uses the derivative method to locate extreme points, and the energy distribution integral frequency range of the frequency-domain features is preset based on the characteristics of the fermentation signal. In some embodiments, smoothing processing is added to feature computation, using moving average filtering of feature sequences to reduce fluctuations. In specific implementations, the attention mechanism of the convolutional neural network is implemented as an SE module, compressing the activation network to generate weights. L2 normalization is performed before feature weighted fusion to improve numerical stability, and the activation function of the fully connected layer uses a linear function, outputting a production state evaluation vector. It can be understood that the learning of attention weights is completed through end-to-end training, without the need for manual setting. Optionally, a dropout layer can be introduced into the convolutional neural network to prevent overfitting and improve generalization ability.

[0028] In practical implementation, the input data preprocessing of the machine learning fusion engine includes descaling and scaling to achieve zero mean and unit variance. Post-processing of the output production status evaluation vector includes clamping to a reasonable range to avoid overflow. In some embodiments, the engine logs feature values ​​and fusion weights for offline analysis. The entire processing chain is subject to strict time constraints, and worst-case execution time analysis ensures real-time response, supporting the closed-loop control requirements of deer wine fermentation. The engine's modular design facilitates maintenance and upgrades, and standardized interfaces ensure system compatibility. Optionally, the engine can be expanded to support more sensor types by simply adjusting the feature extraction branch.

[0029] Example 2: See Figure 3In specific implementation, the time-domain feature set in the clean data block is obtained by calculating the mean, variance, and peak index. The mean is calculated by arithmetically averaging the values ​​of all sampling points within the data block, reflecting the central trend of the data over the entire time window. The variance is calculated using the sample variance formula to measure the dispersion of data points relative to the mean. The peak index is obtained by scanning the data block to identify the global maximum and minimum values ​​and calculating the maximum absolute deviation from the mean. The frequency-domain feature set is converted into a frequency-domain representation of the clean data block using Fast Fourier Transform (FFT). The number of FFT points is set to 1024, and the sampling frequency is set according to the sensor hardware specifications. The dominant frequency and energy distribution are extracted. The dominant frequency is the frequency component corresponding to the largest amplitude in the frequency-domain amplitude spectrum. The energy distribution is calculated by summing the squares of the amplitudes of each frequency point within the 0-Nyquist frequency range and normalizing it to the [0,1] interval. In some embodiments, the time-domain feature set can be supplemented with a skewness index to describe the distribution asymmetry, and the frequency-domain feature set can be supplemented with a spectral centroid to describe the centroid of the frequency distribution. In practical implementation, a multi-branch convolutional neural network (CNN) structure is constructed. This structure contains two independent forward propagation branches: the first branch processes the temporal feature set, and the second branch processes the frequency domain feature set. The input layer dimension of the first branch is consistent with the number of features in the temporal feature set, and the input layer dimension of the second branch is consistent with the number of features in the frequency domain feature set. Each branch contains several convolutional layers, batch normalization layers, and modified linear unit activation layers. The parameters of the multi-branch CNN structure are initialized using the He normal initialization method, and the convolutional layer padding is set to maintain the same padding size to preserve the feature map dimensions.

[0030] In implementation, an attention mechanism is used to dynamically adjust the feature weights of the first and second branches. A channel attention map is calculated on the output feature map of the first branch. This channel attention map is compressed using global average pooling to generate channel descriptors. The channel descriptors learn the correlation between channels through two fully connected layers, outputting temporal feature weight coefficients. A spatial attention map is calculated on the output feature map of the second branch. This spatial attention map is processed by applying max pooling and average pooling along the channel dimension. The pooling results are concatenated and then passed through a convolutional layer to learn spatial importance, outputting frequency domain feature weight coefficients. The temporal and frequency domain feature weight coefficients are normalized using the Softmax function to ensure that the sum of all weight coefficients is 1, resulting in a fused weight matrix. The feature maps of the first and second branches are then weighted and summed based on the fused weight matrix. This weighted summation is performed element-wise to complete the cross-modal feature concatenation. Optionally, the fully connected layers in the attention mechanism can be replaced with one-dimensional convolutional layers to reduce the number of parameters.

[0031] In practical implementation, the calculation process for the time-domain feature set involves: mean calculation by summing all sampled values ​​within the data block and dividing by the total number of points; variance calculation using the unbiased estimator formula, i.e., sum of squared deviations divided by the number of points minus one; and peak index identification by comparing the values ​​of each sampled point within the data block with their neighboring points to identify local extrema, calculating the ratio of the extremum point with the largest absolute value to the mean. The frequency-domain feature set conversion process involves applying a Hanning window function to the clean data block before the Fast Fourier Transform to reduce spectral leakage; the dominant frequency is identified as the frequency value corresponding to the global maximum value in the amplitude spectrum; and the energy distribution is calculated as the sum of the squared amplitudes at each frequency point, divided by the total energy for normalization. In some embodiments, the time-domain features can be supplemented with the calculation of the root mean square value to characterize the effective amplitude, and the frequency-domain features can be supplemented with the calculation of the frequency standard deviation to describe the degree of spectral spread. In its implementation, the first branch of the multi-branch convolutional neural network has the following structure: the input layer receives a 3D temporal feature vector; the first convolutional layer uses 32 1x1 convolutional kernels with a stride of 1; a batch normalization layer standardizes the convolutional output and corrects the nonlinearity introduced by the linear unit activation function; the output feature map is then passed to the attention module. The second branch has the following structure: the input layer receives a 2D frequency domain feature vector; the first convolutional layer uses 32 1x1 convolutional kernels; and subsequent layers are symmetrical to the first branch to ensure matching feature map sizes for easy subsequent fusion. It can be understood that the 1x1 kernel size is chosen to achieve fully connected equivalent computation across features, reducing spatial information loss.

[0032] In specific implementation, the computation details of the channel attention map are as follows: Global average pooling is performed along the height and width dimensions of the feature map output from the first branch, compressing the two-dimensional feature map of each channel into scalar values ​​to form a channel descriptor vector. This channel descriptor vector is then dimensionality-reduced through a first fully connected layer with a 16:1 ratio, followed by a second fully connected layer to restore the original number of channels. Finally, a sigmoid activation function is applied to generate temporal feature weight coefficients. The computation details of the spatial attention map are as follows: Max pooling and average pooling are performed along the channel dimensions of the feature map output from the second branch. The two pooling results are concatenated by channel to form a two-dimensional feature map. This feature map is then used for spatial feature learning through a 7x7 convolutional layer, with the convolution output having 1 channel. Finally, a sigmoid activation function is applied to generate frequency domain feature weight coefficients. During normalization, the temporal and frequency domain feature weight coefficients are concatenated into a one-dimensional vector, input into a Softmax function for probability distribution transformation, with the temperature parameter of the Softmax function set to 1. The output is a fused weight matrix. Optionally, the attention mechanism can be extended to hybrid attention, simultaneously calculating channel and spatial attention.

[0033] In practical implementation, the cross-modal feature concatenation operation involves decomposing the fusion weight matrix into weight sub-matrices corresponding to the first branch and the second branch. These weight sub-matrices are expanded to the same dimension as the original feature maps through a broadcast mechanism. The first branch feature map is element-wise multiplied by the temporal weight sub-matrix, and the second branch feature map is element-wise multiplied by the frequency domain weight sub-matrix. The two weighted feature maps are then element-wise summed to achieve weighted summation, completing the cross-modal feature concatenation. The concatenated fusion feature map is then mapped to a fixed-dimensional production state evaluation vector through a fully connected layer. The input dimension of the fully connected layer is the flattened size of the fusion feature map, and the output dimension is fixed at 128 dimensions. The fully connected layer uses a linear activation function. Essentially, the weighted summation operation is equivalent to adaptively weighting and fusing the features from two modalities, with the weights dynamically generated by the attention mechanism. Optionally, a normalization layer can be added after cross-modal feature concatenation to stabilize the training process.

[0034] In the specific implementation, the training configuration of the multi-branch convolutional neural network uses sensor data and corresponding state labels collected during the historical deer wine fermentation process as training data. The loss function chosen is mean squared error loss, which measures the difference between the predicted and true vectors. The optimizer is the Adam optimizer, with an initial learning rate of 0.001 and a batch size of 32 based on memory capacity. An early stopping strategy is employed during training to prevent overfitting; training terminates when the validation set loss does not decrease for 10 consecutive epochs. In some embodiments, training data augmentation techniques, such as adding Gaussian noise and random scaling, enhance model robustness. In the specific implementation, the attention mechanism's parameter learning is performed end-to-end along with the entire network. The weights of the fully connected layers in the channel attention map are updated through backpropagation, while the weights of the convolutional layers in the spatial attention map are optimized synchronously. It is understood that the effectiveness of the attention mechanism depends on sufficient training data to avoid weight learning getting trapped in local optima. Optionally, L1 regularization constraints can be added to the attention weights to induce sparsity.

[0035] In practical implementation, the real-time performance of the entire feature fusion process is ensured. TensorRT is used to accelerate and optimize the forward inference of the multi-branch convolutional neural network. Operator fusion technology reduces memory access frequency, and the attention mechanism is implemented in parallel. The calculation of time-domain and frequency-domain feature weight coefficients is distributed to different computing units and performed synchronously. In some embodiments, network weights are stored as FP16 half-precision floating-point numbers to balance computational accuracy and speed. In practical implementation, the output interface of the production status evaluation vector is designed as a shared memory mapping method. Vector data is cyclically written to a designated memory area in binary format for low-latency reading by the virtual production simulation platform. It can be understood that the physical meaning of the vector elements corresponds to different dimensions of the fermentation state, such as temperature stability, humidity uniformity, and appearance change rate. Optionally, timestamps and confidence indices can be added to the output vector to enhance reliability.

[0036] In specific implementations, the numerical processing for time-domain feature calculations includes: mean calculation using double-precision floating-point accumulation to avoid rounding errors; variance calculation employing a two-pass algorithm to improve numerical stability; and peak detection using a local window scanning method with a window size of 5 sampling points. In frequency-domain feature calculations, the Fast Fourier Transform is implemented using the Cooley-Tukey algorithm; dominant frequency identification uses parabolic interpolation to improve frequency resolution; and energy distribution calculation ignores the DC component and focuses on the AC signal. In some embodiments, outlier filtering is added to feature calculations, removing data points exceeding three standard deviations. The deployment details of the multi-branch convolutional neural network include: model files using the ONNX format for cross-platform compatibility; pre-allocating input and output buffers after the inference engine loads the model; and the attention mechanism module being implemented as a custom operator registered with the inference framework. The model hot update mechanism allows switching model versions without interrupting service. Optionally, the network can integrate Monte Carlo dropout techniques to estimate prediction uncertainty.

[0037] In practical implementation, attention weights are visualized and monitored. During the training phase, the average weight distribution of the channel attention map and spatial attention map is recorded. During the inference phase, the feature channel index and spatial location corresponding to the maximum weight are output in real time for diagnostic model decision-making. In some embodiments, the attention weight distribution can be clustered to identify typical attention patterns at different fermentation stages. In practical implementation, alternative schemes for cross-modal feature concatenation are evaluated, including dimensionality reduction through fully connected layers after feature concatenation and taking the maximum value of feature elements, but the weighted summation scheme shows the best performance. It can be understood that the selection of feature concatenation scheme is based on objective comparison of ablation experiment results. Optionally, the concatenation operation can be replaced by a gating mechanism to control the information flow.

[0038] In specific implementations, the post-processing flow of the production status assessment vector involves mini-maximum normalization of vector elements to the [0,1] interval, and outliers are handled using the Winsorizing method, limiting them to the 5th and 95th percentile ranges. In some embodiments, a moving average filter is applied to the vector sequence to smooth instantaneous fluctuations. In specific implementations, the computational resource management of the entire system prioritizes the multi-branch convolutional neural network inference task to the highest priority, binds the attention mechanism computation task to the high-performance computing core, and uses DMA technology for feature map transfer to reduce CPU usage. It is understood that real-time performance assurance requires determining the system load limit through worst-case execution time analysis. Optionally, the system can dynamically adjust the computational accuracy to adapt to resource fluctuations.

[0039] Example 3: In specific implementation, the construction of the virtual production simulation platform begins with collecting equipment and process parameters of the deer wine production line. Equipment parameters include the geometric dimensions of the fermentation tank, the model and specifications of the agitator, the power rating of the heating device, and the installation coordinates of the sensors. Process parameters include the temperature control range, humidity setpoint, total fermentation time, and feed ratio. These parameters are obtained through on-site measurements, equipment manual queries, and summarization of historical production records. A three-dimensional geometric model is established using computer-aided design software such as SolidWorks. The modeling process involves drawing a three-dimensional solid model of the fermentation tank based on the equipment parameters, including the precise geometry and relative positions of the tank body, jacket, agitator, temperature sensor probe, and humidity sensor module. The three-dimensional geometric model is saved in a standard format such as STEP or IGES for easy import into the simulation environment. In some embodiments, the collection of equipment parameters can be extended to pipeline layout and valve types. Process parameters may include pH control range and pressure parameters. The establishment of the three-dimensional geometric model includes the addition of material property definitions such as color and texture to enhance visualization. In practical implementation, an actuator interface module is embedded in the 3D geometric model. This module employs industrial communication protocols such as Modbus TCP / IP or PROFIBUS. The hardware uses an embedded controller such as a PLC or industrial PC, while the software handles protocol stack parsing and data exchange. It connects to actual actuators, including temperature control valves, humidity regulators, and agitator motor drivers, and injects control signals. These signals are in analog voltage (0-10V) or digital (switching) signal format, injected through the digital or analog output channels of the actuator interface module. It is understood that the embedding of the actuator interface module must ensure electrical isolation and signal matching to avoid interference with the actual equipment.

[0040] In practical implementation, a multi-channel signal synthesizer is configured. This synthesizer is implemented based on software algorithms and hardware boards. The software runs on a real-time operating system, and the algorithm library includes a function generator, a noise generator, and a sequence player. It can generate simulated temperature curves, simulated humidity curves, and simulated image sequences. The simulated temperature curve uses mathematical functions such as piecewise linear functions or stochastic processes to simulate real temperature fluctuations. The simulated humidity curve is calculated based on a humidity transfer model. The simulated image sequence uses computer graphics to render the appearance changes of the fermentation broth, such as color gradients and foam dynamics. The number of output channels of the multi-channel signal synthesizer matches the sensor type. The temperature curve output is a voltage signal corresponding to the thermocouple reading, the humidity curve output is a current signal corresponding to the transmitter output, and the image sequence output is a video stream format. In some embodiments, the multi-channel signal synthesizer can integrate a physics engine such as ODE (OpenDynamicsEngine) to simulate fluid motion, enhancing the realism of the image sequence. In practical implementation, a remote control agent is installed and deployed on a web server using a client-server architecture. The agent software uses WebSocket or MQTT protocols for bidirectional communication. It receives external control commands, such as setpoint adjustment or mode switching commands, and forwards them to the simulation platform. Commands are in JSON or XML format, including timestamps, command types, and parameter values. The forwarding mechanism is implemented through message queues or remote procedure calls. It is understood that the remote control agent must have authentication and encryption functions to ensure control security.

[0041] In practical implementation, a unified simulation environment is formed by integrating a 3D geometric model, actuator interface module, multi-channel signal synthesizer, and remote control agent. The integration process utilizes simulation platform development frameworks such as Unity3D or Gazebo. After importing, the 3D geometric model is set as a dynamic object in the scene. The actuator interface module is bound to the virtual actuator model via a plug-in interface. The output signal of the multi-channel signal synthesizer is connected to a virtual sensor display. The network interface of the remote control agent is integrated with the simulation event system. The unified simulation environment runs on a high-performance computer, supporting real-time rendering and physics calculations. The user interface of the virtual production simulation platform provides a dashboard and a 3D view. The dashboard displays real-time data curves such as temperature and historical control commands, while the 3D view can be interactively rotated and zoomed for easy monitoring and debugging. In some embodiments, the unified simulation environment can add a data logging function to save simulation operation logs for playback analysis. Optionally, the simulation environment can be expanded to support hardware-in-the-loop testing by connecting actual actuators to the simulation loop.

[0042] In practical implementation, the details of establishing the 3D geometric model are as follows: the modeling software uses a parametric design method; the cylindrical model of the fermenter is driven by diameter and height parameters; internal components such as baffles are added using Boolean operations; the model mesh uses tetrahedral elements, and the mesh size is set according to the required computational accuracy; the exported model file includes hierarchical information, such as the relationships between sub-components in the assembly, facilitating identification and operation in the simulation environment. The implementation of the actuator interface module involves the module's communication protocol stack implementing Modbus function code processing, such as reading holding registers (function code 03) and writing individual registers (function code 06); a data mapping table maps control signal values ​​to register addresses; when a control signal is injected, the module checks if the signal range exceeds limits, triggering an alarm if it does. The algorithm for the multi-channel signal synthesizer generates the simulated temperature curve based on a regression model fitted from historical fermentation data, for example, using a multinomial regression function.

[0043] in: Indicates the simulated temperature over time The value, , , The regression coefficients are learned from historical data using the least squares method. The random noise term simulates environmental disturbances. The simulated humidity curve uses a first-order hysteresis model to simulate the humidity response. The rendering of the simulated image sequence uses the OpenGL library, generating a physically based lighting model and texture mapping for each frame. For remote control agent deployment, the agent server is configured with a static IP address and port number. Client authentication uses a username and password mechanism, and a sequence number is added during command forwarding to ensure order.

[0044] In practical implementation, the virtual production simulation platform operates as follows: upon platform startup, the 3D geometric model is loaded into the scene graph; the network connection of the actuator interface module is initialized; the multi-channel signal synthesizer begins generating signal streams; the remote control agent listens to a designated port; the platform's main loop updates the physical state every frame, such as calculating the temperature distribution within the fermenter, updating the 3D model's posture, and rendering images for display. The actuator interface module injects control signals; upon receiving a control command, the module parses the command parameters, converts them into electrical signal values, and outputs them to the hardware interface via a digital-to-analog converter. Simultaneously, the virtual actuator model synchronizes its actions, such as visualizing valve opening changes in the 3D model. Signal synchronization by the multi-channel signal synthesizer ensures alignment of the timestamps of temperature curves, humidity curves, and image sequences. A high-precision timer ensures timing consistency, and the frame rate of the image sequence matches the display refresh rate. It can be understood that the platform's real-time performance depends on the allocation of computing resources, and the physical calculation thread is set to high priority. In some embodiments, the platform can integrate a script engine to support automated test script execution.

[0045] In practical implementation, the integration of the unified simulation environment involves several technical details. After the 3D geometric model is imported into the simulation engine, rigid body physical properties such as mass and collision shape are added. The actuator interface module is bound to simulation events via callback functions, such as control signal updates triggering model animations. The output signals of the multi-channel signal synthesizer are mapped to simulation variables via shared memory. The command parser of the remote control agent converts JSON commands into internal events. The user interface is implemented using UI toolkits such as Qt, with dashboard controls bound to simulation data variables for real-time updates. The virtual production simulation platform includes debugging functions, providing variable monitors and breakpoint settings for easy troubleshooting. During simulation environment runtime resource management, memory allocation pool optimization reduces fragmentation, and the rendering pipeline uses layered detail technology to balance image quality and performance.

[0046] See Figure 4 This diagram showcases a complete monitoring system for the fermentation process of deer wine, including the temporal changes of three key parameters: temperature, humidity, and control signals. The red curve represents the dynamic changes in fermentation temperature, reflecting the thermal effect of microbial activity during fermentation; the blue curve shows the adjustment process of environmental humidity to ensure suitable fermentation conditions; the green dashed line represents the control signals automatically generated by the system, which issue corresponding adjustment commands when parameters deviate from the set range. The entire diagram intuitively demonstrates the real-time monitoring and intelligent control capabilities of the virtual production simulation platform for the fermentation process.

[0047] Example 4: In specific implementation, the virtual production simulation platform loads the production status evaluation vector and then runs the physical mechanism model. The production status evaluation vector is received from the machine learning fusion engine. The vector dimension is 128, containing comprehensive state information of the deer wine fermentation process. The production status evaluation vector is mapped to the initial condition parameters of the physical mechanism model. The mapping process is achieved through a linear transformation matrix. The matrix elements are trained based on historical data. The vector elements are mapped to the initial values ​​of temperature, humidity, substrate concentration, and product concentration. The initial condition parameters are set at the solver entry of the physical mechanism model. The partial differential equations in the physical mechanism model are solved. The partial differential equations are established based on the laws of conservation of energy and mass, describing the temperature field distribution and concentration field distribution in the fermentation tank. The solution method uses the finite difference method for numerical calculation. The discretization grid step size is set according to the tank size, and the time step satisfies the CFL condition to ensure stability. The spatiotemporal distribution data of the temperature field and concentration field are output. The control deviation index is calculated based on the temperature and concentration field distributions. The control deviation index is defined as the Euclidean distance between the actual distribution and the target distribution, which is obtained from the process specifications. The deviation value for each grid point is calculated and then averaged to obtain the overall control deviation index. A preset control rule base is queried based on the control deviation index. The control rule base is stored as a relational database table, where the rule conditions are deviation ranges and the rule actions are control parameter adjustments. The query uses SQL statements to match the range in which the deviation index falls. Matching control action sequences are retrieved from the control rule base. These control action sequences include temperature setpoint adjustments, stirring rate changes, and feed valve opening control, encoded as analog control signal sequences in PWM waveform parameters or digital instruction strings, and output through the actuator interface module.

[0048] In practice, the production status assessment vector is mapped to the initial condition parameters of the physical mechanism model. The mapping matrix has a dimension of 128 rows and 4 columns, corresponding to the 128 elements of the production status assessment vector being mapped to four initial condition parameters (temperature, humidity, substrate concentration, and product concentration). Matrix multiplication is performed during the simulation platform initialization phase, and the resulting vector serves as the initial condition for the partial differential equation system. The units of the initial condition parameters are converted to the standard units required by the model, such as temperature to Kelvin and concentration to moles per cubic meter. The partial differential equation system of the physical mechanism model is solved. The equation system includes heat conduction equations and reaction-diffusion equations. The heat conduction equations describe the temperature change over time, and the reaction-diffusion equations describe the concentration changes of each chemical component. The discretized mesh divides the fermenter into cubic elements, and the number of mesh nodes is set based on the tank volume and accuracy requirements. The solver iteratively calculates the field distribution at each time step, using an implicit difference scheme to avoid numerical oscillations. The calculation of the control deviation index involves loading target distribution data from the database, representing the field distribution under the ideal fermentation curve, while the actual distribution is obtained from the model output. A parallel algorithm is used to accelerate grid point processing for deviation calculation, and index values ​​are normalized to the 0-1 range for easy comparison. A pre-set control rule base is queried; the rule base table structure includes fields such as rule ID, lower deviation limit, upper deviation limit, action type, and action amplitude. The query logic is: SELECT action sequence WHERE deviation index BETWEEN lower deviation limit AND upper deviation limit. After retrieving the control action sequence, the encoding process converts the action amplitude into an electrical signal value; for example, temperature adjustment is mapped to a 0-10V voltage. The sequence is organized by timestamp.

[0049] In practical implementation, the physical mechanism model is constructed by collecting the geometric and physical property parameters of the deer wine fermentation tank. Geometric parameters are obtained by measuring the tank volume and internal component layout using a 3D scanner. Physical property parameters, such as thermal conductivity and specific heat capacity, are retrieved from material data sheets and recorded in a configuration file. A system of partial differential equations is established based on the laws of energy and mass conservation. The energy conservation equations consider heat generation and heat loss terms, while the mass conservation equations include reaction rate and diffusion terms. The equations are in the form of nonlinear partial differential equations. The partial differential equation system is spatially discretized using the finite difference method to divide the computational domain into uniform grid nodes. A structured grid is used for mesh generation, and the node spacing is determined based on the Courant number. Initial state parameters and boundary condition parameters for the fermentation process are set. The initial state parameters are based on the measured temperature at the time of material feeding, while the boundary condition parameters, including the heat convection coefficient of the tank wall and the ambient temperature, are obtained through sensor calibration. Based on the initial state parameters and boundary condition parameters, the parameters in the partial differential equation system are iteratively corrected using historical fermentation operation data. The correction algorithm uses the Levenberg-Marquardt optimization method to minimize the root mean square error between the model output and the historical data.

[0050] In some embodiments, parameter acquisition for the physical mechanism model can be extended to agitator power and jacket heat transfer coefficient, and the establishment of partial differential equations can add a turbulence model to enhance accuracy. Queries from the control rule base can support fuzzy matching to handle boundary conditions.

[0051] Optionally, a nonlinear neural network can be used instead of a linear matrix in the mapping process to improve mapping accuracy. The discretized mesh can be a non-uniform mesh to adapt to boundary layer effects.

[0052] Optionally, the calculation of control deviation indicators can incorporate a weight matrix to prioritize critical areas, and the rule base can be dynamically updated to adapt to process changes.

[0053] It is understandable that solving physical mechanism models requires high-performance computing resources due to their complexity, and model calibration requires sufficient historical data to ensure reliability.

[0054] In practical implementation, the construction details of the pre-defined control rule base are as follows: the rule base is designed based on expert knowledge and experimental data; the rule conditions cover the full range of deviation indicators from 0 to 1; the rule actions are derived based on PID control principles; and the rule base is loaded into memory when the simulation platform starts to accelerate querying. Refer to Table 1 for an example structure of the control rule base, showing some rule entries: Table 1: Control Rule Base Rule ID Lower limit of deviation Upper limit of deviation Action type Range of motion Effective time (s) R001 0.00 0.10 Maintain the current 0.0 0 R002 0.10 0.30 Temperature adjustment +0.5°C 60 R003 0.30 0.60 Stirring acceleration +10% 30 R004 0.60 1.00 Adding ingredients +5% 120 The query implementation of the rule base is a real-time matching algorithm. After the deviation index is input, the rule table is traversed to find the matching interval and the corresponding action sequence is returned. The action sequence contains multiple action instructions and their timing. When encoded as an analog control signal sequence, each action is converted into a specific voltage value or digital code. The sequence is sent to the actuator through a message queue.

[0055] In practical implementation, the specific forms of the partial differential equations of the physical mechanism model are as follows: the heat conduction equation includes a partial derivative of temperature with respect to time equal to the thermal diffusivity multiplied by the temperature Laplace term plus the reaction heat source term; the reaction diffusion equation includes a partial derivative of concentration with respect to time equal to the diffusion coefficient multiplied by the concentration Laplace term minus the reaction consumption term. Parameters in the equations, such as the diffusion coefficient and reaction rate constant, are determined experimentally. During discretization, the spatial derivative is approximated using the central difference method, and the temporal derivative is approximated using the backward difference method. The linear equation system is solved using the conjugate gradient method. Initial state parameters are set using measured feed temperatures obtained from temperature sensors and converted into initial model conditions. The thermal convection coefficient in the boundary condition parameters is obtained by fitting experimental data, and the ambient temperature is obtained from an environmental monitoring system. In the iterative calibration process, historical data includes temperature and concentration curves from multiple fermentation batches. The calibration algorithm adjusts the parameter values ​​to minimize the error between the model output curve and the historical curves. The calibrated parameters are saved for subsequent simulations.

[0056] In its implementation, the virtual production simulation platform operates as follows: after startup, it loads the physical mechanism model and rule base, receives the production status evaluation vector, and initiates model solving. During the solving process, it monitors changes in field distribution in real time, calculates control deviation indicators every time step, triggers rule queries, generates control signal sequences, and injects them into the actuator interface. Simultaneously, the platform visualizes and displays the field distribution map and signal sequences. Parallel computation for model solving utilizes multi-threading technology, with grid computing distributed across different threads, and rule queries using hash indexes to accelerate matching. This entire implementation ensures accurate simulation and optimized control of the deer wine fermentation process.

[0057] See Figure 5 The graph displays the spatial distribution of temperature within the tank, with colors ranging from warm to cool representing temperature variations. Blue contour lines reflect the distribution characteristics of reactant concentrations, and the density of the contour lines indicates the steepness of the concentration gradient. The graph also labels overall control deviation indicators and corresponding control actions; this information is crucial for optimizing fermentation process parameters. Through this refined field distribution analysis, the fermentation process can be predicted and controlled more accurately, improving the quality of deer-made wine.

[0058] Example 5: In specific implementation, the design of the edge computing control grid begins with precise measurement of the spatial dimensions and equipment layout of the deer wine production area. Spatial dimensions are obtained using a laser rangefinder to acquire data on the length, width, and height of the production workshop. The equipment layout records the specific physical coordinates of each fermentation tank, sensor node, and actuator, forming a digital map of the workshop. Based on the measurement data, the coverage radius and deployment density of the edge nodes are determined. The coverage radius is set according to the wireless communication capabilities of the edge computing devices; for example, the coverage radius of devices based on the Wi-Fi 6 standard is approximately 50 meters. The deployment density calculation is based on the principle of ensuring no blind spots in data acquisition, using a circular coverage model for geometric calculation. The optimal set of node locations is calculated based on the coverage radius and deployment density. This optimal set of node locations is obtained by solving the facility location problem. The objective function is to maximize the coverage area and minimize the number of nodes. Genetic algorithms or particle swarm optimization algorithms are used to calculate coordinate points. Edge computing devices are installed at each location. These edge computing devices are industrial-grade embedded systems, such as ARM architecture industrial control computers, equipped with multiple digital input / output interfaces and wireless communication modules. Each edge computing device is configured with a network address and a communication protocol stack. The network address adopts the IPv6 static allocation scheme to ensure address uniqueness. The communication protocol stack includes physical layer, data link layer, network layer and transport layer protocols. The physical layer adopts the IEEE 802.11ac standard, the data link layer implements the TDMA mechanism, the network layer uses the IPv6 routing protocol, and the transport layer uses the UDP protocol to ensure real-time performance. An edge computing control grid with a mesh topology is constructed. The mesh topology is realized by each edge computing device establishing multiple wireless links with neighboring nodes, supporting dynamic routing and self-healing functions.

[0059] In practical implementation, each edge node derives a set of control strategies in real time based on its local data stream. A data buffer is set up at each edge node, employing a circular buffer structure. The capacity of the buffer is calculated based on the sampling frequency and retention time. Local sensor data streams are continuously collected, including real-time readings from temperature, humidity, and vision sensors. A time series forecasting algorithm is used to predict data trends over a future period. The algorithm selects either an autoregressive integral moving average (ARM) model or a long short-term memory (LSM) network. The parameters of the ARM model are determined using the AIC criterion, and the number of hidden layer neurons in the LSM network is set based on data complexity. A predicted data sequence is generated, containing predicted sensor values ​​for multiple future time points. This predicted data sequence is then input into a strategy optimization algorithm. The algorithm evaluates the utility of different control actions based on a reward function. The reward function is designed considering control objectives such as temperature stability, energy minimization, and product quality indicators. Utility value calculation uses dynamic programming or Monte Carlo tree search methods. The control action with the highest utility value is selected to form a control strategy set. This set contains control command sequences with multiple time steps, such as temperature setpoint adjustment sequences and stirring rate control sequences.

[0060] In some embodiments, data cache management can employ dynamic memory allocation to adapt to fluctuations in data volume, and prediction algorithms can integrate multiple models in parallel to dynamically select the optimal predictor. Policy optimization algorithms can incorporate deep reinforcement learning techniques to handle complex decision-making scenarios.

[0061] In practical implementation, a data exchange middleware is established to integrate production status evaluation vectors and control strategy sets, parse out optimized control command streams, and deploy a message queue service. The message queue service uses RabbitMQ or Apache Kafka middleware, deployed on a central server. A topic switch is configured to receive production status evaluation vectors from the machine learning fusion engine and control strategy sets from the edge computing control mesh. The message format uses Protocol Buffers serialization to ensure efficient transmission. Multi-objective optimization calculations are performed on the production status evaluation vectors and control strategy sets, using a non-dominated sorting genetic algorithm or a weighted sum method. Optimization objectives include control accuracy, response speed, and energy consumption, generating candidate command sets containing multiple control command schemes that meet the constraints. The feasibility and consistency of the candidate command sets are verified. Feasibility checks whether the control commands exceed the physical limits of the actuators, and consistency checks for logical conflicts between commands, such as simultaneous heating and cooling. After eliminating conflicting commands, the optimized control command stream is packaged and encoded using TLV (Type-Length-Value) format, and then distributed to the field actuators via industrial Ethernet.

[0062] Optionally, the message queue service can be configured with persistent storage to ensure messages are not lost, and fuzzy logic can be introduced into multi-objective optimization calculations to handle uncertainties. Timing consistency checks can be added to the instruction verification process to ensure the logical consistency of instruction timing.

[0063] Optionally, the data exchange middleware can implement data encryption and digital signature functions to ensure communication security, and optimize the control command flow by adding cyclic redundancy check codes to detect transmission errors.

[0064] It is understandable that the real-time requirements of edge computing control grids are strict clock synchronization, and the throughput of data exchange middleware needs to match the system data scale.

[0065] In practical implementation, the deployment details of the edge computing control mesh include: spatial dimension measurement generating point cloud data, which is then converted into a 3D model; power supply locations and network interfaces are marked on the equipment layout diagram; coverage radius calculation considers wall attenuation factors to correct theoretical values; and deployment density optimization uses the Voronoi diagram method to divide areas. Edge computing devices are installed using rail fixing, power is connected to an industrial distribution box, network address configuration is implemented via the DHv6 protocol, the communication protocol stack is implemented using a ported open-source network stack, and the mesh topology is built based on the OLSR routing protocol. The data buffer uses double-buffer technology to avoid data loss, and sensor data streams undergo timestamp alignment. The input data for the time series prediction algorithm is standardized using a sliding window, and the prediction results are smoothed using a Kalman filter, with the prediction horizon set to the next 5 control cycles. The reward function weights of the policy optimization algorithm are initialized using expert experience, utility value calculation uses the Q-learning algorithm, and control actions are discretized into multiple levels. The message queue service's topic partitioning is based on data type, and messages for the production status evaluation vector and control policy set carry priority tags. The population size for multi-objective optimization computation is set to 100, and the number of iterations is 500 generations. The feasibility check rule base is established based on the equipment manual, and the consistency verification uses a time-series logic model. Frame headers and checksums are added to the packaging of the optimization control command flow, and ModbusTCP is used as the transmission protocol.

[0066] In some embodiments, edge computing devices can be equipped with watchdog timers to enable self-recovery, and prediction algorithms can update model parameters online to adapt to the process dynamics. Instruction verification can incorporate formal verification methods to improve reliability.

[0067] In practical implementation, real-time performance is guaranteed for edge node control strategy derivation. Data buffers utilize ping-pong operations for lock-free read / write operations. Prediction algorithms use fixed-point arithmetic to accelerate inference, and strategy optimization algorithms apply action pruning to reduce the search space. The number of consumer worker threads in the message queue service is configured based on the number of CPU cores, and multi-objective optimization calculations are accelerated using GPU parallelism. The transmission priority of the optimized control command stream is set to the highest, and network bandwidth is reserved to ensure real-time channels. The entire system undergoes worst-case response time analysis to ensure that control cycle requirements are met. Node fault detection in the edge computing control mesh is achieved through a heartbeat mechanism; faulty nodes are automatically isolated, and the mesh reconfigures its routing. The data exchange middleware implements persistent data storage, and unprocessed messages are reloaded after power failure recovery. It can be understood that system reliability is enhanced through redundant design and failover mechanisms.

[0068] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for optimizing control of the production process of deer liquor based on artificial intelligence, characterized in that, The method comprises: configuring a multi-source perception unit array integrated with a heat-sensitive sensor, a humidity-sensitive sensor and an image collector for synchronously capturing temperature readings, humidity readings and appearance change data of the deer liquor fermentation link; applying a machine learning fusion engine to perform multi-modal alignment and feature-level fusion on the temperature readings, humidity readings and appearance change data, and output a production state evaluation vector; constructing a virtual production simulation platform that runs a physical mechanism model after loading the production state evaluation vector to generate a simulated control signal sequence; designing an edge computing control grid composed of multiple edge nodes, each of which derives a control strategy set in real time according to local data flow; establishing a data exchange middleware that integrates the production state evaluation vector and the control strategy set to analyze an optimized control instruction stream.

2. The artificial intelligence-based deer distiller production process optimization control method of claim 1, wherein, The application of the machine learning fusion engine to the multi-modal alignment and feature-level fusion of the temperature readings, humidity readings and appearance change data to output the production state evaluation vector comprises: receiving raw data streams from the multi-source perception unit array, performing sliding window segmentation on the raw data streams to form time series data blocks; using a wavelet transform algorithm to estimate the noise power spectrum of each time series data block to determine an adaptive filtering threshold; designing a digital filter bank according to the adaptive filtering threshold to perform frequency domain filtering processing on the time series data blocks to obtain clean data blocks; extracting time domain features and frequency domain features from the clean data blocks, performing feature weighted fusion through a convolutional neural network to generate the production state evaluation vector. 3.The AI-based optimization control method for producing deer-based liquor according to claim 2, wherein, The extraction of time domain features and frequency domain features from the clean data blocks and the generation of the production state evaluation vector through convolutional neural network feature weighted fusion comprise: calculating the mean, variance and peak value indicators of each clean data block as a time domain feature set; applying a fast Fourier transform to convert the clean data block to a frequency domain representation and extracting the dominant frequency and energy distribution as a frequency domain feature set; constructing a multi-branch convolutional neural network structure, wherein a first branch processes the time domain feature set and a second branch processes the frequency domain feature set; using an attention mechanism to dynamically adjust the feature weights of the first branch and the second branch for cross-modal feature splicing; mapping the spliced features through a fully connected layer to the production state evaluation vector of a fixed dimension. 4.The AI-based optimization control method for producing deer-based liquor according to claim 3, wherein, The use of the attention mechanism to dynamically adjust the feature weights of the first branch and the second branch for cross-modal feature splicing comprises: calculating a channel attention map on the output feature map of the first branch to generate time domain feature weight coefficients; calculating a spatial attention map on the output feature map of the second branch to generate frequency domain feature weight coefficients; normalizing the time domain feature weight coefficients and the frequency domain feature weight coefficients to obtain a fusion weight matrix; performing weighted summation on the feature maps of the first branch and the second branch according to the fusion weight matrix to complete cross-modal feature splicing. 5.The AI-based optimization control method for producing deer-based liquor according to claim 1, wherein, The construction of the virtual production simulation platform comprises: collecting equipment parameters and process parameters of the deer liquor production line to establish a three-dimensional geometric model; Embedding an actuator interface module in the three-dimensional geometric model for connecting actual actuators and injecting control signals; Configuring a multi-channel signal synthesizer capable of generating analog temperature curves, analog humidity curves, and analog image sequences; Installing a remote control agent through which external control commands are received and forwarded to the simulation platform; Integrating the three-dimensional geometric model, the actuator interface module, the multi-channel signal synthesizer, and the remote control agent to form a unified simulation environment as the virtual production simulation platform. 6.The AI-based optimization control method of a deer distillate production process according to claim 5, wherein, The virtual production simulation platform loads the production state evaluation vector to run the physical mechanism model to generate an analog control signal sequence, including: Mapping the production state evaluation vector to the initial condition parameters of the physical mechanism model; Solving the partial differential equation system in the physical mechanism model to obtain the temperature field and concentration field distribution in the fermenter; Based on the temperature field and concentration field distribution, calculating the control deviation index and querying the pre-configured control rule library according to the control deviation index; Retrieving the matched control action sequence from the control rule library and encoding it as the analog control signal sequence. 7.The AI-based optimization control method of a deer distillate production process according to claim 6, characterized in that, The construction method of the physical mechanism model includes: Collecting the geometric structure parameters and physical property parameters of the deer liquor fermenter, including the tank volume and internal component layout, and the material thermal conductivity and specific heat capacity; Based on the energy conservation law and the mass conservation law, a partial differential equation system is established to describe the temperature distribution and reactant concentration distribution in the fermenter; Spatially discretizing the partial differential equation system and dividing the calculation region into uniform grid nodes using the finite difference method; Setting the initial state parameters and boundary condition parameters of the fermentation process, the initial state parameters are based on the measured temperature values at the time of feeding, and the boundary condition parameters include the heat convection coefficient of the tank wall and the environmental temperature value; Based on the initial state parameters and boundary condition parameters, iteratively correct the parameters in the partial differential equation system based on historical fermentation operation data. 8.The AI-based optimization control method of a deer distillate production process according to claim 1, wherein, The design of the edge computing control grid includes: Measuring the spatial dimensions and equipment layout of the deer liquor production area to determine the coverage radius and deployment density of the edge nodes; According to the coverage radius and deployment density, calculate the optimal node position set, and install edge computing devices at each position; Configure network addresses and communication protocol stacks for each edge computing device to build the edge computing control grid with a mesh topology. 9.The AI-based optimization control method of a deer distillate production process according to claim 8, wherein, Each edge node derives a control strategy set in real time based on local data streams, including: Setting a data buffer area in each edge node to continuously collect local sensor data streams; Using a time series prediction algorithm to predict the data trend in the future period to generate a predicted data sequence; Input the predicted data sequence into the strategy optimization algorithm, which evaluates the utility value of different control actions based on a reward function; Select the control action with the highest utility value to form the control strategy set. 10.The AI-based optimization control method of a deer distillate production process according to claim 1, wherein, The establishment of the data exchange middleware integrates the production state evaluation vector and the control strategy set through the data exchange middleware to analyze the optimized control instruction stream, including: Deploying a message queue service to receive the production state evaluation vector from the machine learning fusion engine and the control policy set from the edge computing control grid; Performing multi-objective optimization calculation on the production state evaluation vector and the control policy set to generate a candidate instruction set; Verifying the feasibility and consistency of the candidate instruction set and packing the optimized control instruction stream after eliminating conflicting instructions.

Images