A method for constructing a tobacco leaf moisture content prediction model

Abstract

The embodiment of the present application discloses a kind of tobacco moisture content prediction model construction methods, the method comprises: to the echo sequence of downsampling is decomposed to extract free water dielectric characteristics and bound water dielectric characteristics;Obtain packing physical parameters and the three-dimensional space coordinates of tobacco bale, calculate bale core effective stress;Based on bale core temperature and bale core effective stress, by conditional generative adversarial network processing generates virtual moisture feature data, virtual moisture feature data and real feature data are merged to constitute balanced space-time data set;Balanced space-time data set is used to construct high-dimensional feature matrix, and the intrinsic time sequence feature set is output by dimension reduction processing to high-dimensional feature matrix;Intrinsic time sequence feature set is used to update and train network weight for prediction model, and the construction of tobacco moisture content prediction model is completed.The technical scheme of the embodiment of the present application generates virtual data by physical state constraint, solves the data imbalance problem caused by edge sample, and excludes the covering of total moisture parameter to moisture active state.

Description

Technical Field

[0001] This invention relates to the field of tobacco leaf storage technology, and in particular to a method for constructing a tobacco leaf moisture content prediction model. Background Technology

[0002] After threshing and re-curing, tobacco leaves require several years of aging and storage. Changes in moisture content directly affect the quality of aging and the risk of mold growth. Existing storage management systems typically deploy temperature and humidity sensors and moisture detection probes to collect parameters such as core temperature and storage environment temperature and humidity, and then use empirical formulas or statistical models to predict moisture content. Furthermore, to address the issue of scarce abnormal state samples, some solutions employ data augmentation techniques such as generative adversarial networks to expand the training sample.

[0003] However, the existing technologies still have the following drawbacks: First, existing methods rely solely on the macroscopic total moisture content index, ignoring the different physical binding states of free and bound water. The total moisture parameter masks the differences in internal water activity, resulting in insufficient distinguishability of features for mold risk. Second, when generating minority class samples, data augmentation techniques ignore the constraints of internal effective stress changes caused by tobacco pack compaction and the thermodynamic transformation mechanism of moisture. The generated virtual data lacks physical laws and is prone to producing invalid samples that do not conform to actual conditions. Summary of the Invention

[0004] This invention provides a method for constructing a tobacco leaf moisture content prediction model. By extracting moisture characteristics under different physical states, the method improves the distinguishability of moisture characteristic data for mold risk. By generating virtual data that conforms to physical state constraints and merging it with real data, the method alleviates the data imbalance problem caused by the scarcity of abnormal samples.

[0005] In a first aspect, embodiments of the present invention provide a method for constructing a tobacco leaf moisture content prediction model, comprising: Acquire the core temperature of tobacco leaves, time-series parameters of storage environment, and downsampled echo sequences; The downsampled echo sequence is decomposed to extract the dielectric characteristics of free water and bound water. Obtain the physical parameters of the outer packaging box of the cigarette pack and the three-dimensional spatial coordinates of the cigarette pack in the warehouse stacking, and calculate the effective stress of the pack core; Based on the core temperature and the effective core stress, virtual moisture feature data is generated and output through a conditional generative adversarial network. Real feature data uploaded by the edge gateway is obtained, and the virtual moisture feature data and the real feature data are merged to form a balanced spatiotemporal dataset. A high-dimensional feature matrix is ​​constructed using the free water dielectric characteristics, bound water dielectric characteristics, core temperature, core effective stress, and storage environment time series parameters contained in the balanced spatiotemporal dataset. The high-dimensional feature matrix is ​​then dimensionality-reduced to output an intrinsic time series feature set. The intrinsic temporal feature set is used to update and train the network weights of the dual-branch gated recurrent unit prediction model, thus completing the construction of the tobacco leaf moisture content prediction model.

[0006] Optionally, the step of calculating the effective stress of the tobacco pack core specifically includes: obtaining the three-dimensional spatial coordinates of the tobacco pack in the warehouse stack, the total height of the tobacco stack, and the equivalent density parameters of the tobacco pack at different height levels; extracting the height coordinates from the three-dimensional spatial coordinates and performing an integral operation with the total height of the tobacco stack and the equivalent density parameters of the tobacco pack at different height levels to obtain the macroscopic stacking integral; obtaining the structural yield threshold and internal stress transmission coefficient of the outer packaging box; and calculating the effective stress of the tobacco pack core by multiplying the macroscopic stacking integral, the structural yield threshold of the outer packaging box, and the internal stress transmission coefficient.

[0007] Optionally, the step of generating and outputting virtual moisture feature data based on the core temperature and the effective core stress through a conditional generative adversarial network includes: combining the core temperature and the effective core stress with time-stamp alignment and one-dimensional formatting preprocessing to form condition variables; concatenating a Gaussian distributed random noise vector with the condition variables by feature dimension as input data to the generator network in the conditional generative adversarial network; calculating the dynamic conversion rate of bound water to free water using the effective core stress and the core temperature; and introducing the dynamic conversion rate as a physical penalty term into the training process of the conditional generative adversarial network to generate the virtual moisture feature data.

[0008] Optionally, the specific process of constructing a physical penalty term and using the physical penalty term to train the conditional generative adversarial network includes: calculating the rate of change of the free water dielectric feature over time; subtracting the product of the dynamic conversion rate and the bound water dielectric feature from the rate of change of the free water dielectric feature over time, squared, and multiplied by the penalty term weight coefficient to construct the physical penalty term; adding the physical penalty term to the binary cross-entropy loss function of the discriminator network in the conditional generative adversarial network to form a joint loss function; and using the joint loss function to train the conditional generative adversarial network to generate and output the virtual water feature data composed of virtual free water dielectric features and virtual bound water dielectric features.

[0009] Optionally, obtaining the downsampled echo sequence includes: sending an initial broadband pulse sequence with an initial pulse fall edge time to a flexible microstrip probe arranged inside the cigarette pack via a device to obtain an initial echo sequence; and performing downsampling processing on the initial echo sequence using equivalent time sampling to generate the downsampled echo sequence.

[0010] Optionally, the step of decomposing the downsampled echo sequence to extract the dielectric features of free water and bound water includes: repeatedly adding positive white noise sequences and negative white noise sequences to the downsampled echo sequence to construct multiple reconstruction analysis sequences; performing empirical mode decomposition on each reconstruction analysis sequence to generate modal components; and performing an overall average calculation on the modal components obtained from the decomposition of the reconstruction analysis sequence to obtain multiple intrinsic mode function components; calculating the center frequency of each intrinsic mode function component; and based on a preset frequency cutoff threshold, classifying the intrinsic mode function components with a center frequency greater than the preset frequency cutoff threshold into high-frequency intrinsic mode function components and extracting them as the dielectric features of free water; and classifying the intrinsic mode function components with a center frequency less than or equal to the preset frequency cutoff threshold into low-frequency intrinsic mode function components and extracting them as the dielectric features of bound water.

[0011] Optionally, the step of reducing the dimensionality of the high-dimensional feature matrix to output the intrinsic time-series feature set includes: calculating the Euclidean distance between sample points in the high-dimensional feature matrix using a two-stage equidistant mapping algorithm, selecting the nearest neighbor points to establish connecting edges and construct a local neighborhood graph; calculating the shortest path length between any two sample points in the local neighborhood graph using a shortest path algorithm as the geodesic distance; combining the geodesic distances between all the sample points to form a distance matrix; performing an operation on the distance matrix using a centering transformation matrix to obtain a double-centered inner product matrix; performing eigenvalue decomposition on the double-centered inner product matrix to extract eigenvalue sequences and eigenvectors; and constructing and outputting the intrinsic time-series feature set based on the eigenvalue sequences and eigenvectors.

[0012] Optionally, the step of updating and training the network weights of the dual-branch gated recurrent unit prediction model using the intrinsic temporal feature set includes: dividing the intrinsic temporal feature set into a temporal feature training set and a temporal feature test set; training the dual-branch gated recurrent unit prediction model using the temporal feature training set; concatenating the macroscopic evolution feature vector output by the long-term trend branch network and the microscopic mutation feature vector output by the short-term perturbation branch network in the dual-branch gated recurrent unit prediction model by feature dimension through a dual-branch feature fusion layer; and updating the network weights of the dual-branch gated recurrent unit prediction model using a binary cross-entropy loss function.

[0013] Optionally, after completing the construction of the tobacco leaf moisture content prediction model, the method further includes: outputting the predicted mold probability value through the prediction model of the dual-branch gated loop unit; obtaining the historical high-frequency intrinsic mode function components of the non-moldy tobacco leaves, calculating the energy mean of the historical high-frequency intrinsic mode function components for calibration, and obtaining the free water safety benchmark threshold; subtracting the free water safety benchmark threshold from the free water dielectric characteristics at the current time step to obtain the difference between the free water and the safety benchmark, and obtaining the effective free water deviation through a non-negative truncation function; and adding the product of the predicted mold probability value and the preset prediction probability weight coefficient to the product of the effective free water deviation and the preset current state deviation weight coefficient to calculate the dynamic intervention index.

[0014] Optionally, after calculating and obtaining the dynamic intervention index, the method further includes: triggering a pulse reconstruction operation based on the dynamic intervention index, dynamically compressing the initial pulse fall time of the initial broadband pulse sequence sent to the flexible microstrip probe arranged inside the cigarette pack, and generating a reconstructed broadband pulse sequence; sending the reconstructed broadband pulse sequence to obtain a reconstructed echo sequence, and performing downsampling processing on the reconstructed echo sequence using equivalent time sampling to generate an updated downsampled echo sequence; supplementing the updated downsampled echo sequence and the corresponding storage environment time series parameters at the current time step into the balanced spatiotemporal dataset; and using the supplemented balanced spatiotemporal dataset to perform incremental training and weight updates on the dual-branch gated recurrent unit prediction model.

[0015] Secondly, embodiments of the present invention provide a device for constructing a tobacco leaf moisture content prediction model, comprising: The parameter acquisition module is used to acquire the core temperature of tobacco leaves, storage environment time-series parameters, and downsampled echo sequences. The feature extraction module is used to decompose the downsampled echo sequence and extract the dielectric features of free water and the dielectric features of bound water. The core effective stress calculation module is used to obtain the packaging physical parameters of the outer packaging box of the cigarette pack and the three-dimensional spatial coordinates of the cigarette pack in the warehouse stacking, and to calculate the core effective stress. The equilibrium spatiotemporal dataset composition module is used to process the core temperature and the core effective stress through a conditional generative adversarial network to generate and output virtual moisture feature data, obtain real feature data uploaded by the edge gateway, and merge the virtual moisture feature data and the real feature data to form an equilibrium spatiotemporal dataset. The intrinsic time series feature set output module is used to construct a high-dimensional feature matrix using the free water dielectric characteristics, the bound water dielectric characteristics, the core temperature, the core effective stress, and the storage environment time series parameters contained in the equilibrium spatiotemporal dataset, and to perform dimensionality reduction processing on the high-dimensional feature matrix to output the intrinsic time series feature set. The model building module is used to update and train the network weights of the dual-branch gated recurrent unit prediction model using the intrinsic temporal feature set, thereby completing the construction of the tobacco leaf moisture content prediction model.

[0016] Thirdly, embodiments of the present invention provide an electronic device, the electronic device comprising: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the tobacco leaf moisture content prediction model construction method provided in any embodiment of the present invention.

[0017] Fourthly, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for constructing a tobacco leaf moisture content prediction model as provided in any embodiment of the present invention.

[0018] Fifthly, embodiments of the present invention provide a computer program product, including a computer program that, when executed by a processor, implements the tobacco leaf moisture content prediction model construction method provided in any embodiment of the present invention.

[0019] The technical solution of this invention provides a time-consistent data foundation for subsequent multi-source feature fusion and physical model construction by acquiring the core temperature, storage environment time-series parameters, and downsampled echo sequences of tobacco leaves. By decomposing the downsampled echo sequences and extracting free water dielectric features and bound water dielectric features, the single macroscopic moisture content index is broken down into moisture features of different physical binding states, eliminating the masking effect of total moisture parameters on the internal active state of moisture and improving the distinguishability of moisture feature data for mold risk. By acquiring the packaging physical parameters of the outer packaging box of the tobacco bales and the three-dimensional spatial coordinates of the tobacco bales in the storage stack, the effective stress of the core is calculated, providing a physically realistic stress input for calculating the moisture kinetic conversion rate. Virtual moisture feature data is generated and output by processing the core temperature and the effective core stress through a conditional generative adversarial network, while acquiring the real feature data uploaded by the edge gateway. The virtual moisture feature data is merged with the real feature data to form a balanced spatiotemporal dataset. The generation of virtual data is constrained by physical laws, which solves the problem of severe imbalance in model training data distribution caused by the scarcity of marginal abnormal state samples and avoids the generation of invalid data that does not conform to physical laws. By using the free water dielectric features, bound water dielectric features, core temperature, core effective stress, and storage environment time series parameters contained in the balanced spatiotemporal dataset to construct a high-dimensional feature matrix, the high-dimensional feature matrix is ​​dimensionality reduced to output an intrinsic time series feature set, which effectively removes multicollinearity redundant features and noise interference, and extracts pure intrinsic variables that reflect the thermodynamic transformation state of moisture. By using the intrinsic time series feature set to update and train the network weights of the bi-branch gated recurrent unit prediction model, the tobacco leaf moisture content prediction model is completed, realizing accurate prediction of tobacco leaf moisture content and its mold risk at multiple time scales.

[0020] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart of a method for constructing a tobacco leaf moisture content prediction model according to Embodiment 1 of the present invention; Figure 2This is a flowchart of a method for constructing a tobacco leaf moisture content prediction model provided in Embodiment 2 of the present invention; Figure 3 This is a schematic diagram of the structure of a tobacco leaf moisture content prediction model construction device provided in Embodiment 3 of the present invention; Figure 4 This is a schematic diagram of the structure of an electronic device that implements the tobacco leaf moisture content prediction model construction method of the present invention. Detailed Implementation

[0023] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are 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 should fall within the scope of protection of the present invention.

[0024] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0025] Example 1 Figure 1 This invention provides a flowchart of a method for constructing a tobacco leaf moisture content prediction model, according to Embodiment 1. This embodiment is applicable to predicting the internal moisture content of tobacco bales and assessing the risk of mold growth in tobacco leaves during storage. The method can be executed by a tobacco leaf moisture content prediction model construction device, which can be implemented in hardware and / or software and can be configured in an electronic device. Figure 1 As shown, the method includes: S110: Obtain the core temperature of the tobacco leaves, the time series parameters of the storage environment, and the downsampling echo sequence.

[0026] In this embodiment, tobacco leaves refer to sheet tobacco that has been threshed, re-dried, packaged, and placed in boxes for storage and aging. After threshing and re-drying, tobacco leaves undergo a natural aging and storage process lasting several years. Changes in moisture content during storage directly affect the aging quality and the risk of mold growth. The core temperature refers to the temperature value of the geometric center region of the tobacco bale. Since tobacco leaves are porous materials with low thermal conductivity, there is a significant difference between the external temperature and the internal core temperature of the tobacco bale. The core temperature directly reflects the thermodynamic state inside the tobacco leaf and is a key parameter affecting moisture activity and microbial metabolism. The storage environment time-series parameters refer to a sequence of parameters formed by arranging the dry-bulb temperature and relative humidity values ​​of the air inside the storage area in chronological order. The temperature and humidity of the storage environment are external boundary conditions affecting heat and moisture exchange inside the tobacco bale, and their changes over time affect the moisture state inside the tobacco leaf through the slow permeation of the packaging box. The downsampled echo sequence refers to discrete time-series waveform data generated after downsampling the initial echo sequence using equivalent time sampling technology. This sequence contains information reflecting the dielectric properties inside the tobacco leaf, which is the basic data for subsequent extraction of moisture characteristics.

[0027] Specifically, during the leaf re-drying and packaging process, a flexible microstrip probe is implanted into the core area of ​​the tobacco pack. After warehousing, the data acquisition terminal (i.e., the device) sends an initial broadband pulse sequence with the initial pulse falling edge time to the inside of the tobacco pack via an RF interface, simultaneously recording the core temperature and collecting time-series parameters of the storage environment using temperature and humidity sensors. To ensure the reliability of multi-source data fusion processing, the data acquisition terminal uses a built-in hardware real-time clock to uniformly timestamp and align the initial electromagnetic wave reflection signal data, core temperature, and storage environment time-series parameters corresponding to the initial broadband pulse sequence, completing the time-domain synchronization of heterogeneous data. The data acquisition terminal uses equivalent time sampling technology to downsample the initial echo sequence and sends the processed downsampled echo sequence to the edge gateway.

[0028] This step involves simultaneously collecting three types of heterogeneous data: core temperature, storage environment time-series parameters, and downsampled echo sequences, and aligning them with unified timestamps. This provides a time-consistent data foundation for subsequent multi-source feature fusion and physical model construction. Core temperature reflects the thermodynamic state inside the tobacco leaf, storage environment time-series parameters characterize changes in external boundary conditions, and downsampled echo sequences carry information about the dielectric properties inside the tobacco leaf. Together, these three constitute a complete physical information chain encompassing the external environment, packaging boundaries, and the interior of the tobacco bale.

[0029] As an optional implementation of this disclosure, obtaining the downsampled echo sequence may specifically include: sending an initial broadband pulse sequence with an initial pulse falling edge time to a flexible microstrip probe arranged inside the cigarette pack via a device to obtain an initial echo sequence; and performing downsampling processing on the initial echo sequence using equivalent time sampling to generate a downsampled echo sequence.

[0030] In this embodiment, the device end can refer to a data acquisition terminal arranged inside the tobacco bale and in the storage area. The device end includes a microwave pulse generation circuit, an equivalent time sampling circuit, and a wireless communication module. The device end is responsible for sending detection signals into the tobacco bale and receiving reflected echoes. The flexible microstrip probe can refer to a detection device based on a coplanar waveguide structure using a flexible insulating polymer substrate. In this structure, the flexible insulating polymer substrate (such as a polyimide film) provides flexible physical support, and the coplanar waveguide structure places the central signal conductor and the two side ground planes on the same plane, allowing the probe to bend with the deformation of the tobacco leaves under the high pressure stress of the tobacco bale, effectively reducing the air gap between the central signal conductor and the tobacco material, achieving a tight fit, and reducing the physical obstruction interference of the external packaging box frame on the high-frequency detection signal. The initial pulse fall edge time can refer to the transition time required for the pulse to fall from a high level to a low level in the initial broadband pulse sequence transmitted by the data acquisition terminal. The pulse fall edge time determines the spectral energy distribution of the transmitted signal. The steeper the fall edge, the higher the frequency of the microwave harmonic components excited in the frequency domain.

[0031] In this embodiment, the initial broadband pulse sequence can refer to an electromagnetic pulse signal with broadband characteristics sent by the acquisition terminal to the inside of the tobacco bale through the radio frequency interface. Based on the principle of microwave time-domain reflectometry, this pulse sequence will generate electromagnetic wave reflection when it encounters a region where the dielectric constant inside the tobacco material changes abruptly during transmission along the flexible microstrip line probe, forming an electromagnetic wave reflection signal that returns along the original path. The initial echo sequence can refer to the electromagnetic wave signal sequence that returns after the initial broadband pulse sequence propagates inside the tobacco bale and is reflected back through the region where the dielectric constant changes abruptly. This sequence belongs to the high-frequency analog signal in the gigahertz band and carries the distribution information of the dielectric properties inside the tobacco bale. Equivalent time sampling can refer to a sampling technique that converts a high-frequency periodic signal into a low-frequency slow signal for acquisition. Equivalent time sampling can be used to delay each sample relative to the trigger signal by a fixed step time increment in multiple consecutive repetition cycles of a periodic signal, and to sequentially splice the discrete voltage data points extracted in multiple cycles according to the sampling order to reconstruct the complete high-frequency echo waveform characteristics on the equivalent macroscopic time axis. Downsampling can refer to the signal processing process of converting a high-frequency initial echo sequence into a low-frequency downsampled echo sequence using equivalent time sampling techniques.

[0032] Specifically, the acquisition terminal is equipped with a microwave pulse generation circuit. Based on the principle of microwave time-domain reflectometry, the acquisition terminal sends an initial broadband pulse sequence with an initial pulse falling edge time to the inside of the tobacco bale via an RF interface. During transmission along the flexible microstrip probe, the initial broadband pulse sequence encounters regions where the dielectric constant of the tobacco material changes abruptly, generating electromagnetic wave reflections and forming an initial electromagnetic wave reflection signal that returns along the original path. The initial echo sequence belongs to a high-frequency analog signal in the gigahertz band; directly performing real-time analog-to-digital conversion and wireless transmission of the high-frequency analog signal would cause excessive communication bandwidth load. The acquisition terminal is equipped with an equivalent time sampling circuit, which uses a stepped delay triggering mechanism to discretize and map the actual high-frequency time axis to an equivalent low-frequency time axis. For example, in multiple consecutive repetition cycles of the initial echo sequence, the equivalent time sampling circuit sets a fixed phase offset time increment based on the internal stepped clock resolution. At the arrival of each new signal cycle, the equivalent time sampling circuit controls the trigger to delay the trigger position compared to the previous cycle by one phase offset time increment and extracts a discrete voltage data point. The equivalent time sampling circuit splices together discrete voltage data points extracted sequentially from multiple cycles according to the sampling order, thereby restoring the complete high-frequency echo waveform characteristics on the equivalent macroscopic time axis.

[0033] This optional implementation converts high-frequency echo signals in the gigahertz band into low-frequency sampling sequences using equivalent time sampling technology, which significantly reduces the data transmission bandwidth requirements and analog-to-digital conversion difficulty, making it possible to transmit echo data in real time in an industrial IoT environment. At the same time, it completely preserves the waveform characteristics of the high-frequency echo, providing a high-quality data foundation for subsequent signal decomposition and moisture feature extraction.

[0034] S120. Decompose the downsampled echo sequence to extract the dielectric characteristics of free water and bound water.

[0035] In this embodiment, the dielectric characteristics of free water can refer to the characteristic quantitative index extracted from the downsampled echo sequence, reflecting the dipole orientation polarization response of free water molecules inside the tobacco leaf. Free water molecules have a large dipole orientation space, and their response rate to external alternating electromagnetic fields is fast, corresponding to high-frequency polarization relaxation characteristics. The dielectric characteristics of bound water can refer to the characteristic quantitative index extracted from the downsampled echo sequence, reflecting the interfacial polarization response of bound water molecules inside the tobacco leaf. Bound water molecules are subject to strong intermolecular forces (such as rotationally hindered water molecules near solid surfaces), and their response rate to external alternating electromagnetic fields is slow, mainly exhibiting low-frequency interfacial polarization characteristics.

[0036] Specifically, water molecules in different binding states exhibit corresponding relaxation response timescales when excited by electromagnetic pulses. Free water molecules have a large dipole orientation space, and their response rate to external alternating electromagnetic fields is fast, corresponding to high-frequency polarization relaxation characteristics. Bound water molecules are subject to strong intermolecular forces, and their response rate to external alternating electromagnetic fields is slow, mainly exhibiting low-frequency interface polarization characteristics. The edge gateway uses a complementary integrated empirical mode decomposition algorithm to decompose the downsampled echo sequence in the time and scale domains to obtain multiple intrinsic mode function components. Based on the dispersive polarization theory of porous media, the edge gateway extracts the high-frequency intrinsic mode function components as the dielectric characteristics of free water and extracts the low-frequency intrinsic mode function components as the dielectric characteristics of bound water.

[0037] This step decomposes the single macroscopic moisture content index into two different physical states: free water dielectric characteristics and bound water dielectric characteristics. Free water is a direct participant in microbial metabolic activities and is highly correlated with mold risk, while bound water is bound by strong intermolecular forces, making it difficult for microorganisms to utilize directly. By distinguishing between these two moisture states, the masking effect of total moisture parameters on the internal active state of water is eliminated, improving the discriminative power of moisture characteristic data input into the prediction model for mold risk.

[0038] As an optional implementation of this disclosure, decomposing the downsampled echo sequence to extract free water dielectric features and bound water dielectric features may specifically include: repeatedly adding positive white noise sequences and negative white noise sequences to the downsampled echo sequence to construct multiple reconstruction analysis sequences; performing empirical mode decomposition on each reconstruction analysis sequence to generate modal components; and performing an overall average calculation on the modal components obtained from the decomposition of the reconstruction analysis sequence to obtain multiple intrinsic mode function components; calculating the center frequency of each intrinsic mode function component; based on a preset frequency cutoff threshold, classifying intrinsic mode function components with a center frequency greater than the preset frequency cutoff threshold into high-frequency intrinsic mode function components and extracting them as free water dielectric features; and classifying intrinsic mode function components with a center frequency less than or equal to the preset frequency cutoff threshold into low-frequency intrinsic mode function components and extracting them as bound water dielectric features.

[0039] In this embodiment, a positive white noise sequence can refer to a sequence of noise samples with positive amplitude in a random signal sequence with zero mean and constant power spectral density. A negative white noise sequence can refer to a sequence of noise samples with negative amplitude in a random signal sequence with zero mean and constant power spectral density. A reconstruction analysis sequence can refer to multiple signal sequences to be decomposed formed by adding positive and negative white noise sequences to the original downsampled echo sequence, respectively. Empirical mode decomposition (EMD) can refer to a signal processing method that adaptively decomposes a nonlinear, non-stationary signal into a series of modal components according to local characteristic time scales. A modal component can refer to a signal component with a single local characteristic time scale generated during EMD. Overall averaging can refer to performing an arithmetic average operation on the corresponding order modal components obtained from the decomposition of multiple reconstruction analysis sequences. Overall averaging can be used to cancel artificially added white noise interference.

[0040] In this embodiment, the Intrinsic Mode Function (IMF) component can refer to an oscillating signal component obtained after complementary integrated empirical mode decomposition that satisfies specific mathematical conditions (the number of extrema and zero-crossings are equal or differ by at most one, and the local mean is zero). The center frequency can refer to the weighted average frequency value calculated by energy weight in the spectrum of the IMF component. The center frequency can reflect the dominant oscillation frequency of the component signal. The frequency cutoff threshold can refer to the frequency boundary value used to distinguish between high-frequency IMF components (free water dielectric characteristics) and low-frequency IMF components (bound water dielectric characteristics), such as 200 MHz. This threshold is obtained through a pre-conducted calibration experiment using a thermogravimetric and microwave vector network analyzer on the target tobacco leaf sample. For example, this threshold can be set in the frequency range of 100 MHz to 300 MHz. The high-frequency IMF component can refer to the IMF component with a center frequency greater than the frequency cutoff threshold, corresponding to the high-frequency relaxation response of the free water molecule dipole orientation polarization. The low-frequency intrinsic mode function components can refer to the intrinsic mode function components with a center frequency less than or equal to the frequency cutoff threshold, corresponding to the low-frequency relaxation response of the interface polarization of bound water molecules.

[0041] Specifically, the edge gateway employs a complementary ensemble empirical mode decomposition (EMD) algorithm to process the downsampled echo sequence. This algorithm eliminates mode aliasing, a common phenomenon in EMD, by adding paired positive and negative white noise sequences to the original signal. For example, the edge gateway repeatedly adds positive and negative white noise sequences with fixed standard deviations to the downsampled echo sequence to construct multiple reconstruction analysis sequences. Subsequently, EMD is performed on each reconstruction analysis sequence, generating a series of modal components with different local feature time scales. To counteract the artificially added white noise interference, the corresponding modal components obtained from the decomposition of all reconstruction analysis sequences are averaged to obtain multiple intrinsic mode function (IMF) components. In this embodiment, the standard deviation of the added white noise is set to 0.1 to 0.2 times the standard deviation of the original data of the downsampled echo sequence, and the number of ensemble calculations for the overall average is set to 50 to 100. The decomposition relationship of the downsampled echo sequence can be expressed as: .

[0042] in, This indicates a downsampled echo sequence. Represents the equivalent time variable. This represents the total number of intrinsic mode function components. The decomposition yields the first... Each intrinsic mode function component This represents the residual trend component after decomposition.

[0043] The edge gateway performs a Fast Fourier Transform on each intrinsic mode function component to obtain a discrete spectrum, and then performs centroid calculation based on spectral energy weights. The formula for calculating the center frequency is: .

[0044] in, This represents the center frequency of the intrinsic mode function component. This represents the total number of discrete sampling points in the Fast Fourier Transform. Indicates the first The actual frequency point value corresponding to each discrete frequency index number. Indicates the first The complex spectrum value corresponding to each discrete frequency index number. This represents the power spectral density amplitude.

[0045] Based on the magnitude of the center frequency, the edge gateway uses a preset frequency truncation threshold to divide multiple intrinsic mode function (IMF) components into high-frequency IMF components and low-frequency IMF components. The high-frequency IMF components are extracted as the free water dielectric characteristics. This is achieved by squaring the voltage amplitudes contained within these components and summing the sums on the discrete time axis corresponding to the equivalent time variable, thus obtaining the energy value of the high-frequency IMF components as the free water dielectric characteristics. Similarly, the low-frequency IMF components are extracted as the bound water dielectric characteristics.

[0046] This alternative implementation uses a complementary integrated empirical mode decomposition algorithm to separate the coupled dielectric response signal into multiple intrinsic mode function components with different characteristic time scales, effectively eliminating the mode aliasing phenomenon commonly found in traditional empirical mode decomposition. Based on this, and using the dispersive polarization theory of porous media, the intrinsic mode function components are precisely divided into high-frequency components reflecting the free water state and low-frequency components reflecting the bound water state using a frequency cutoff threshold obtained from a physical calibration experiment. This achieves a physically interpretable separation of the two water dielectric characteristics.

[0047] It should be noted that the edge gateway also uses core temperature and storage environment time-series parameters, as well as free water dielectric characteristics and bound water dielectric characteristics as independent variables, and uses random forest and neural network regression algorithms to output the real-time moisture content of tobacco leaves, and adaptively updates the parameters of the soft measurement model according to the changes in tobacco leaf state.

[0048] S130. Obtain the physical parameters of the outer packaging box of the cigarette pack and the three-dimensional spatial coordinates of the cigarette pack in the warehouse stack, and calculate the effective stress of the pack core.

[0049] In this embodiment, the outer packaging box of the tobacco pack can refer to a rigid cardboard box or wooden box used to hold tobacco leaves after re-drying. The outer packaging box of the tobacco pack has an absorption and shielding effect on external compressive stress, affecting the effective transmission of external loads to the interior of the tobacco pack. Packaging physical parameters can refer to the geometric dimensions, material properties, and structural mechanical parameters of the outer packaging box of the tobacco pack. Packaging physical parameters may include the total height of the tobacco leaf stack, the equivalent density parameters of tobacco packs at different height levels, the structural yield threshold of the outer packaging box, and the internal stress transmission coefficient, etc. Warehouse stacking can refer to the stacking structure formed by layering tobacco packs in a storage area according to certain rules. Three-dimensional spatial coordinates can refer to the spatial position coordinates of the tobacco pack in the warehouse stack. The three-dimensional spatial coordinates can include the horizontal plane coordinates (planar position) and the height coordinates (layer height) of the tobacco pack. The effective stress of the core can refer to the stress value actually transmitted to the core area of ​​the tobacco pack after absorption and shielding by the outer packaging box, such as 16.77 kPa. External compressive loads alter the porosity and capillary forces within the tobacco leaf, thereby reducing the activation energy required for bound hydrolysis and adsorption within the tobacco cell structure. The effective core stress is a key input parameter for calculating the water kinetic conversion rate.

[0050] Specifically, the central server extracts the three-dimensional spatial coordinates of the tobacco bales within the warehouse stack. These coordinates include the horizontal and vertical coordinates of the bales. Before calculating the stacking pressure, the central server retrieves the packaging physical parameters of the corresponding batch of tobacco leaves from the warehouse management database. These parameters include the pre-entered total height of the tobacco stack and the equivalent density parameters of the bales at different height levels. After parsing these parameters, the central server uses the extracted height coordinates, combined with the parsed total height of the tobacco stack and the equivalent density parameters of the bales at different height levels, to perform continuous integration along the direction of gravity, obtaining the macroscopic stacking pressure integral borne by the bales at the corresponding height coordinate. Considering the absorption and shielding effect of rigid outer packaging boxes on external compressive stress, the central server retrieves the structural yield threshold of the corresponding material outer packaging box from a pre-set material property database. It then uses this structural yield threshold to perform nonlinear transmission correction on the macroscopic stacking pressure integral, constructing an effective stress transmission model for the core of the packaging.

[0051] This step, by combining the precise position of the tobacco bales in the stack with the equivalent density distribution at different height levels through integral calculations, and considering the structural yield characteristics and stress transmission coefficient of the outer packaging box, can accurately calculate the effective stress actually transmitted to the core of the tobacco bales. This stress value reflects the compressive effect of the stacking gravity on the internal microstructure of the tobacco leaves and is a key physical input for subsequent calculations of the dynamic conversion rate of bound water to free water.

[0052] As an optional implementation of this disclosure, the step of calculating and obtaining the effective stress of the tobacco pack core specifically includes: obtaining the three-dimensional spatial coordinates of the tobacco pack in the warehouse stack, the total height of the tobacco stack, and the equivalent density parameters of tobacco packs at different height levels; extracting the height coordinates from the three-dimensional spatial coordinates and performing an integral operation with the total height of the tobacco stack and the equivalent density parameters of tobacco packs at different height levels to obtain the macroscopic stacking integral; obtaining the structural yield threshold and internal stress transmission coefficient of the outer packaging box; and calculating the effective stress of the tobacco pack core by multiplying the macroscopic stacking integral, the structural yield threshold of the outer packaging box, and the internal stress transmission coefficient.

[0053] In this embodiment, the total height of the tobacco stack can refer to the total vertical height from the ground to the top of the stack in the warehouse. Different height levels can refer to different height intervals in the warehouse stack, divided according to the number of layers of tobacco bales. The equivalent density parameter can refer to the average density value of tobacco bales at a certain height level. The equivalent density parameter can be pre-calibrated by extracting the actual weighing data and external dimension measurement data of tobacco bales sampled from the previous shift and performing a division operation. The height coordinate can refer to the coordinate components in three-dimensional space that characterize the height level of the tobacco bales in the stack. The integral operation can refer to the mathematical operation of continuously accumulating the equivalent density parameter with respect to the height coordinate along the direction of gravity. The macroscopic stacking pressure integral can refer to the sum of the weight of all tobacco bales above the tobacco bale located at a certain height coordinate, that is, the external stacking pressure load borne by the tobacco bale, such as 60 kPa. The structural yield threshold can refer to the critical stress value corresponding to when the outer packaging material begins to undergo significant plastic deformation, such as 50 kPa. The structural yield threshold can be obtained by calibrating the compressive failure physical test of the material of an unloaded outer packaging box of the same specification using a standard universal testing machine. The internal stress transmission coefficient refers to the efficiency coefficient by which a tobacco stack transmits external compressive stress inwards. The internal stress transmission coefficient can be based on empirical test data of the lateral pressure coefficient in granular material mechanics, and its value is limited to between 0.3 and 0.6, such as 0.4.

[0054] Specifically, the central server extracts the three-dimensional spatial coordinates of the tobacco bales within the warehouse stack. These coordinates include the horizontal and vertical coordinates of the bales. Before calculating the stacking pressure, the central server retrieves the packaging physical parameters of the corresponding batch of tobacco leaves from the warehouse management database. These parameters include the pre-entered total height of the tobacco stack and the equivalent density parameters of the bales at different height levels. After parsing these parameters, the central server uses the extracted height coordinates, combined with the parsed total height of the tobacco stack and the equivalent density parameters of the bales at different height levels, to perform continuous integration along the direction of gravity, obtaining the macroscopic stacking pressure integral borne above the bales at the corresponding height coordinate. The central server retrieves the structural yield threshold of the corresponding material's outer packaging box from a pre-set material property database and uses this threshold to perform nonlinear transfer correction on the macroscopic stacking pressure integral. The formula for calculating the effective stress of the core is: .

[0055] in, Indicates the effective stress of the core. Let denote the macroscopic stacking integral, and exp denote the exponential function with base to the natural constant. This indicates the structural yield threshold of the outer packaging box. This represents the internal stress transmission coefficient of the tobacco leaf stack. When the macroscopic compressive strength integral is lower than the structural yield threshold of the outer packaging box, the external compressive load is mainly borne by the box frame, and the effective core stress transmitted into the tobacco leaf is relatively small. When the macroscopic compressive strength integral approaches or exceeds the structural yield threshold of the outer packaging box, the outer packaging box undergoes slight deformation, and the stress corresponding to the macroscopic compressive strength integral is transferred to the interior of the tobacco leaf.

[0056] This optional implementation introduces a structural yield threshold of the outer packaging box to perform nonlinear correction on the macroscopic stacking integral, accurately characterizing the physical process of stress absorption and post-yield stress transmission by the outer packaging box, thus avoiding errors caused by a simple linear transfer model. The calculated effective core stress reflects the actual thermodynamic driving stress of moisture acting inside the tobacco leaf, providing physically accurate input parameters for subsequent calculations of the dynamic conversion rate.

[0057] S140. Based on the core temperature and core effective stress, a conditional generative adversarial network is used to process and generate and output virtual moisture feature data. Real feature data uploaded by the edge gateway is obtained, and the virtual moisture feature data and real feature data are merged to form a balanced spatiotemporal dataset.

[0058] In this embodiment, a Conditional Generative Adversarial Network (CGAN) can refer to a network architecture that introduces conditional variables into a standard generative adversarial network to control the types of generated data. In this invention, the conditional variables consist of core temperature and effective core stress, constraining the generator network to physical state conditions when generating virtual moisture characteristic data. Virtual moisture characteristic data can refer to data composed of virtual free water dielectric characteristics and virtual bound water dielectric characteristics output by the generator network of the CGAN, simulating the critical state of early mold growth. An edge gateway can refer to data aggregation and preprocessing equipment deployed in the storage area. The edge gateway is responsible for receiving node data packets uploaded by the acquisition terminals, performing signal decomposition and feature extraction, and uploading the extracted real feature data to the central server. Real feature data can refer to the free water dielectric characteristics and bound water dielectric characteristics extracted by the edge gateway from the actual acquired downsampled echo sequence, as well as the corresponding core temperature and storage environment time-series parameters. A balanced spatiotemporal dataset can be defined as a dataset created by merging virtual moisture feature data with real feature data in a predetermined class balance ratio (e.g., 1:1). This dataset eliminates the class imbalance bias in the real data distribution, providing a balanced sample distribution for the training of subsequent prediction models.

[0059] Specifically, in actual warehouse operations, early moldy sample data is scarce. Directly relying on a small amount of real feature data showing anomalies to train the prediction model can easily lead to overfitting to a few abnormal data points. The central server uses a conditional generative adversarial network (GAN) to generate virtual moisture feature data under the critical state of mold growth for data augmentation. To ensure that the generated features conform to actual physical laws and prevent the generation of abnormal data that does not conform to the changing patterns of tobacco leaves, the central server incorporates the effective stress and temperature of the core sheath as conditional variables within the GAN framework, subjecting the generated data to physical state conditions. The central server merges the real feature data uploaded from the edge gateway with the generated virtual moisture feature data to form a balanced spatiotemporal dataset, eliminating class imbalance bias in the real data distribution.

[0060] This step uses a conditional generative adversarial network to generate virtual moisture feature data at the early critical state of mold growth, and then merges it with real feature data in a proportional manner to form a balanced spatiotemporal dataset. This solves the problem of severely imbalanced model training data distribution caused by the scarcity of samples at the edge of abnormal states. Simultaneously, the generation process is constrained by core temperature and effective core stress, ensuring that the generated virtual data corresponds to specific physical stress distributions and thermodynamic states, thus avoiding the problem of non-physical consistency caused by purely data-driven generation.

[0061] It should be noted that the process of merging virtual moisture feature data with real feature data to form an equilibrium spatiotemporal dataset can also be as follows: merge virtual moisture feature data with real feature data, and supplement and merge with physical model experimental data based on constant temperature and humidity chamber and data from leaf re-drying.

[0062] S150. Construct a high-dimensional feature matrix using the free water dielectric characteristics, bound water dielectric characteristics, core temperature, core effective stress, and storage environment time series parameters contained in the equilibrium spatiotemporal dataset. Perform dimensionality reduction processing on the high-dimensional feature matrix to output the intrinsic time series feature set.

[0063] In this embodiment, a high-dimensional feature matrix can refer to a matrix formed by concatenating and combining multiple physical quantities in a balanced spatiotemporal dataset according to their time-series feature dimensions. In a multivariate coupled storage physical environment, there is a thermodynamic coupling relationship between core temperature, storage environment time-series parameters, and tobacco leaf moisture characteristics, resulting in multicollinearity redundancy features in the high-dimensional feature matrix. Dimensionality reduction can refer to the operation of mapping the high-dimensional feature matrix to a low-dimensional intrinsic space while maintaining the inherent geometric structure of the data. The intrinsic time-series feature set can refer to the feature set formed by arranging independent intrinsic variables in time sequence after nonlinear manifold dimensionality reduction. Each intrinsic dimension feature corresponds to the pure moisture thermodynamic transformation state quantity after eliminating environmental background noise interference.

[0064] Specifically, the central server uses the sampling period of the storage environment time-series parameters as a benchmark to align the time intervals. It then performs linear interpolation resampling on the free water dielectric characteristics, bound water dielectric characteristics, core temperature, and effective core stress in the equilibrium spatiotemporal dataset to unify the timestamps of each physical quantity. After time-series alignment, the central server extracts the free water dielectric characteristics, bound water dielectric characteristics, core temperature, effective core stress, and storage environment time-series parameters from the equilibrium spatiotemporal dataset, and concatenates and combines the feature dimensions according to the chronological order of the time series to construct a high-dimensional feature matrix. In the multivariate coupled storage physical environment, there is a thermodynamic coupling relationship between core temperature, storage environment time-series parameters, and tobacco leaf moisture characteristics, resulting in multicollinearity redundant features in the high-dimensional feature matrix. Directly using the high-dimensional feature matrix for deep learning model training leads to parameter space expansion and decreased generalization ability. The central server employs nonlinear manifold mapping technology to extract the independent intrinsic physical state dimensions within the high-dimensional feature matrix, performing nonlinear dimensionality reduction while maintaining the manifold geometry to filter out multicollinearity redundant features and output the intrinsic time-series feature set.

[0065] This step utilizes nonlinear manifold dimensionality reduction technology to effectively remove high-frequency, low-variance redundant dimensions caused by instrument thermal noise and minor environmental disturbances, as well as multicollinear redundant features arising from thermodynamic coupling between core temperature, ambient temperature and humidity, and moisture characteristics, while preserving the nonlinear dynamic evolution laws between the original high-dimensional physical features. The output intrinsic time-series feature set, while reducing data dimensionality, fully retains the core macroscopic dynamic energy components driving the evolution of the moisture phase state within the tobacco leaf, providing clean, independent, and low-redundancy input features for subsequent prediction model training.

[0066] As an optional implementation of this disclosure, the dimensionality reduction processing of the high-dimensional feature matrix to output the intrinsic time series feature set may specifically include: calculating the Euclidean distance between sample points in the high-dimensional feature matrix using a two-stage equidistant mapping algorithm, selecting the nearest neighbor points to establish connecting edges and construct a local neighborhood graph; calculating the shortest path length between any two sample points in the local neighborhood graph using a shortest path algorithm as the geodesic distance; combining the geodesic distances between all sample points to form a distance matrix; performing operations on the distance matrix using a centering transformation matrix to obtain a double-centered inner product matrix, performing eigenvalue decomposition on the double-centered inner product matrix to extract the eigenvalue sequence and eigenvector, and constructing and outputting the intrinsic time series feature set based on the eigenvalue sequence and eigenvector.

[0067] In this embodiment, the two-stage isomap algorithm can refer to a neighborhood-based nonlinear dimensionality reduction method. Its core principle is to approximate the geodesic distance on the local manifold surface using Euclidean distance within the local neighborhood, and then reconstruct the overall nonlinear topology through global graph search. A sample point can refer to the eigenvector corresponding to each time step in the high-dimensional feature matrix, i.e., each row of the high-dimensional feature matrix. Euclidean distance can refer to the straight-line distance between two sample points in the high-dimensional feature space. Nearest neighbors can refer to several other sample points whose Euclidean distance is closest to a given sample point. The number of nearest neighbors can be set to 5 to 15. Connecting edges can refer to the graph connections established between a sample point and its nearest neighbors. The local neighborhood graph can refer to a graph structure composed of all sample points and their connecting edges with their nearest neighbors. The shortest path algorithm can refer to a graph theory algorithm (such as Dijkstra's algorithm or Floyd's algorithm) that calculates the shortest connected path between any two nodes in the graph structure.

[0068] In this embodiment, the geodesic distance can refer to the shortest path length between two sample points in a local neighborhood graph. The geodesic distance can be used to approximate the true geodesic distance on a manifold. The distance matrix can be a symmetric matrix composed of the geodesic distances between all pairwise sample points. The centering transformation matrix can be a transformation matrix obtained by subtracting the mean matrix from the identity matrix. The centering transformation matrix can be used to eliminate the influence of the origin offset on the inner product calculation. The double-centered inner product matrix can be the inner product matrix obtained after performing a double-centering operation on the distance matrix. The double-centered inner product matrix can be used to transform the distance metric into an inner product geometric relationship that can be decomposed into eigenvalues. Eigenvalue decomposition can be a linear algebraic operation that decomposes a matrix into eigenvalues ​​and eigenvectors. The eigenvalue sequence can be a set of eigenvalues ​​arranged in descending order of numerical value. The eigenvector can be a set of orthogonal vectors corresponding to the eigenvalues.

[0069] Specifically, the central server calculates the Euclidean distance between any two sample points in the high-dimensional feature matrix. Based on the magnitude of the Euclidean distance, it selects the nearest neighbors (e.g., 5 to 15) for each sample point and establishes connecting edges between the sample point and its nearest neighbors, constructing a local neighborhood graph. The central server initializes the distances between connected sample points in the local neighborhood graph to their corresponding Euclidean distances, and initializes the distances between unconnected sample points to infinity. The central server uses a graph theory shortest path algorithm to calculate the shortest path length between any two sample points in the local neighborhood graph, and uses the calculated shortest path length as the geodesic distance between the two sample points. The update formula for the geodesic distance is: .

[0070] in, Represents sample points With sample points Geodesic distance between them This represents an intermediate sample point used for transfer jumps in a graph network.

[0071] The central server calculates and combines the geodesic distances between all the acquired sample points to form a distance matrix. To transform the geodesic distance relationships into inner product geometric relationships that can be decomposed using eigenvalues, the central server performs a double centering operation on the distance matrix using a centering transformation matrix. The formula for calculating the double-centered inner product matrix is ​​as follows: .

[0072] Where B represents the doubly centered inner product matrix, and H represents the centered transformation matrix. This represents the squared geodesic distance matrix, formed by squaring each geodesic distance in the distance matrix. The central server performs eigenvalue decomposition on the doubly centered inner product matrix, obtaining a sequence of eigenvalues ​​and their corresponding eigenvectors arranged in descending order of numerical value. The central server accumulates the eigenvalue sequence item by item, calculates the cumulative contribution rate of the eigenvalues, compares the cumulative contribution rate of the eigenvalues ​​with a preset intrinsic dimension retention threshold (85% to 95%), extracts the principal eigenvalues ​​that meet the threshold requirements and their corresponding eigenvectors, and constructs and outputs the intrinsic time series feature set.

[0073] This optional implementation achieves nonlinear dimensionality reduction by using an equidistant mapping algorithm while maintaining the equivalence between the geodesic distance of high-dimensional data and the spatial distance of low-dimensional data. This effectively addresses the nonlinear manifold structure generated by thermodynamic coupling in tobacco storage data. Compared to linear dimensionality reduction methods (such as principal component analysis), equidistant mapping can reveal the true nonlinear manifold geometry of sample points, allowing the intrinsic temporal feature set after dimensionality reduction to maintain the intrinsic dynamic topological relationships between the original high-dimensional physical features in the low-dimensional space.

[0074] It should be noted that, to ensure computational stability, an appender is added to the diagonal elements of the doubly centered inner product matrix. to The regularized perturbation constant.

[0075] S160. The intrinsic temporal feature set is used to update and train the network weights of the two-branch gated recurrent unit prediction model, and the tobacco leaf moisture content prediction model is constructed.

[0076] In this embodiment, the dual-branch gated recurrent unit (GRU) prediction model can refer to a prediction model comprising two parallel GRU branches: a long-term trend branch network and a short-term perturbation branch network. The physical evolution of tobacco mold growth includes long-term, slow cumulative effects and short-term environmental perturbation and abrupt changes. Based on these evolutionary characteristics, the model captures temporal features at different time scales through two branches. Network weights can refer to the parameter values ​​of each neuron connection in the dual-branch gated recurrent unit prediction model. Network weight updates and training enable the model to fit the mapping relationship between input features and output labels. The tobacco moisture content prediction model can refer to a dual-branch gated recurrent unit (GRU) prediction model that, after training with the intrinsic temporal feature set, can output a predicted mold growth probability value.

[0077] Specifically, the central server divides the generated intrinsic temporal feature set into a temporal feature training set and a temporal feature test set in chronological order. Each sample in the temporal feature training set is paired with a binary classification label to characterize whether the tobacco leaf sample will become moldy within a preset future time period (the preset future time period is set based on the average response lag time of the storage environment control equipment, usually 24 hours). The central server uses the temporal feature training set to train the long-term trend branch network and the short-term perturbation branch network in parallel. The central server configures a dual-branch feature fusion layer at the output of the long-term trend branch network and the short-term perturbation branch network, concatenating the macroscopic evolution feature vector output by the long-term trend branch network and the microscopic mutation feature vector output by the short-term perturbation branch network in multiple dimensions, and outputting the predicted mold probability value through a fully connected layer configured with a Sigmoid activation function. During the backpropagation training phase, the central server uses a binary cross-entropy loss function to update the network weights.

[0078] This step utilizes a dual-branch gated recurrent unit prediction model to capture both the long-term, slow cumulative effects and the short-term, abrupt effects of environmental disturbances during the evolution of tobacco leaf mold, enabling the model to simultaneously perceive both macroscopic evolutionary trends and microscopic abrupt changes. Compared to recurrent neural network models with a single time window, the dual-branch structure can more comprehensively characterize the multi-timescale characteristics of tobacco leaf moisture state evolution, improving the prediction accuracy and generalization ability of the model in complex and dynamic storage environments.

[0079] As an optional implementation of this disclosure, updating and training the network weights of the two-branch gated recurrent unit prediction model using the intrinsic temporal feature set may specifically include: dividing the intrinsic temporal feature set into a temporal feature training set and a temporal feature test set; training the two-branch gated recurrent unit prediction model using the temporal feature training set; concatenating the macroscopic evolution feature vector output by the long-term trend branch network and the microscopic mutation feature vector output by the short-term perturbation branch network in the two-branch gated recurrent unit prediction model by feature dimension through a two-branch feature fusion layer; and updating the network weights of the two-branch gated recurrent unit prediction model using a binary cross-entropy loss function.

[0080] In this embodiment, the temporal feature training set can refer to the portion of the intrinsic temporal feature set used for model training. The temporal feature test set can refer to the portion of the intrinsic temporal feature set used for model validation and testing. The dual-branch feature fusion layer can refer to the feature concatenation layer located at the output ends of the long-term trend branch network and the short-term perturbation branch network. The dual-branch feature fusion layer is responsible for concatenating the feature vectors output by the two branches in multiple feature dimensions. The long-term trend branch network can refer to a GRU network branch configured with a long cyclic time window (e.g., 72 hours). The long-term trend branch network can be used to extract the macroscopic evolution baseline from the intrinsic temporal feature set. The macroscopic evolution feature vector can refer to the feature vector output by the long-term trend branch network that reflects the long-term slow change trend of tobacco leaf moisture state. The short-term perturbation branch network can refer to a GRU network branch configured with a short cyclic time window (e.g., 12 hours). The short-term perturbation branch network can be used to capture microscopic mutation features from the intrinsic temporal feature set. The microscopic mutation feature vector can refer to the feature vector output by the short-term perturbation branch network that reflects the short-term fluctuations in tobacco leaf moisture state caused by environmental perturbations. Feature dimension concatenation refers to merging and connecting two feature vectors along their respective feature dimensions to form a comprehensive feature vector that integrates both macroscopic and microscopic information. The binary cross-entropy loss function is a loss function used for binary classification tasks. It measures the difference between the model's predicted probability and the true label.

[0081] Specifically, the central server divides the generated intrinsic temporal feature set into a temporal feature training set and a temporal feature test set in chronological order. In this embodiment, each sample in the temporal feature training set is paired with a binary classification label to characterize whether the tobacco leaf sample will become moldy within a preset future time period, specifically set to 24 hours. The central server uses the temporal feature training set to train the long-term trend branch network and the short-term perturbation branch network in parallel. The long-term trend branch network is configured with a long cyclic time window corresponding to the past 72-hour environmental observation period to extract the macroscopic evolution baseline in the intrinsic temporal feature set; the short-term perturbation branch network is configured with a short cyclic time window corresponding to the past 12-hour observation period to capture the microscopic mutation features in the intrinsic temporal feature set. The central server configures a dual-branch feature fusion layer at the output of the long-term trend branch network and the short-term perturbation branch network, concatenating the macroscopic evolution feature vector output by the long-term trend branch network and the microscopic mutation feature vector output by the short-term perturbation branch network in a multi-dimensional feature dimension, and outputting the predicted mold probability value through a fully connected layer configured with a Sigmoid activation function. During the backpropagation training phase, the central server uses the binary cross-entropy loss function to update network weights. The formula for calculating the binary cross-entropy loss function is: .

[0082] in, This represents the value of the binary cross-entropy loss function, where N represents the total number of training samples in the batch. This represents the true binary classification label of the m-th batch of training samples. This represents the predicted mold probability value corresponding to the m-th batch of training samples.

[0083] This optional implementation captures features at different time scales during the evolution of tobacco mold growth using a long-term trend branch network (72-hour window) and a short-term perturbation branch network (12-hour window). A dual-branch feature fusion layer organically integrates macroscopic evolution features with microscopic mutation features, enabling the model to perceive both long-term, slow cumulative effects and respond to the abrupt impacts of short-term environmental disturbances. A binary cross-entropy loss function is used for end-to-end joint optimization to ensure the model's prediction accuracy in binary classification tasks.

[0084] The technical solution of this invention provides a time-consistent data foundation for subsequent multi-source feature fusion and physical model construction by acquiring the core temperature of tobacco leaves, storage environment time-series parameters, and downsampled echo sequences. By decomposing the downsampled echo sequences and extracting free water dielectric features and bound water dielectric features, the single macroscopic moisture content index is decomposed into moisture features of different physical binding states, eliminating the masking effect of total moisture parameters on the internal active state of moisture and improving the distinguishability of moisture feature data for mold risk. By acquiring the packaging physical parameters of the outer packaging box of the tobacco bales and the three-dimensional spatial coordinates of the tobacco bales in the storage stack, the effective stress of the core is calculated, providing a physically realistic stress input for calculating the moisture kinetic conversion rate. Virtual moisture feature data is generated and output by processing the core temperature and effective core stress through a conditional generative adversarial network, obtaining the real moisture feature data uploaded by the edge gateway. The feature data is constructed by merging virtual moisture feature data with real feature data to form a balanced spatiotemporal dataset. The generation of virtual data is constrained by physical laws, which solves the problem of severe imbalance in model training data distribution caused by the scarcity of marginal abnormal state samples and avoids the generation of invalid data that does not conform to physical laws. By using the free water dielectric characteristics, bound water dielectric characteristics, core temperature, core effective stress and storage environment time series parameters contained in the balanced spatiotemporal dataset to construct a high-dimensional feature matrix, the high-dimensional feature matrix is ​​dimensionality reduced to output an intrinsic time series feature set, which effectively removes multicollinearity redundant features and noise interference, and extracts pure intrinsic variables that reflect the thermodynamic transformation state of moisture. By using the intrinsic time series feature set to update and train the network weights of the bi-branch gated recurrent unit prediction model, the tobacco leaf moisture content prediction model is completed, realizing accurate prediction of tobacco leaf moisture content and its mold risk at multiple time scales.

[0085] Example 2 Figure 2This is a flowchart illustrating a method for constructing a tobacco leaf moisture content prediction model according to Embodiment 2 of the present invention. Based on the above embodiments, this embodiment describes in detail the process of generating virtual moisture feature data using a conditional generative adversarial network based on core temperature and effective core stress. Explanations of terms that are the same as or corresponding to those in the above embodiments are not repeated here. Figure 2 As shown, the method includes: S210: Obtain the core temperature of tobacco leaves, storage environment time series parameters, and downsampling echo sequence.

[0086] S220. Decompose the downsampled echo sequence to extract the dielectric characteristics of free water and bound water.

[0087] S230. Obtain the physical parameters of the outer packaging box of the cigarette pack and the three-dimensional spatial coordinates of the cigarette pack in the warehouse stack, and calculate the effective stress of the pack core.

[0088] S240. The core temperature and effective core stress are combined with timestamp alignment and one-dimensional formatting preprocessing to form condition variables.

[0089] In this embodiment, timestamp alignment refers to synchronizing the core temperature and effective core stress according to the same time base to ensure consistency of the two types of data in the time dimension. One-dimensional formatting preprocessing refers to converting the core temperature and effective core stress into one-dimensional tensor forms respectively, facilitating subsequent feature concatenation with random noise vectors. Condition variables can be vectors composed of the core temperature and effective core stress after timestamp alignment and one-dimensional formatting preprocessing. Condition variables can be used to control the physical state category of generated data in conditional generative adversarial networks.

[0090] Specifically, the central server performs timestamp alignment and one-dimensional formatting preprocessing on the core temperature and effective core stress at the current time step to combine them into a unified condition variable. Timestamp alignment ensures that the core temperature and effective core stress correspond to the same physical moment, while one-dimensional formatting converts them into standardized one-dimensional tensor forms before splicing them together.

[0091] This step ensures strict synchronization of core temperature and effective core stress in the time dimension through timestamp alignment, avoiding the destruction of physical causal relationships caused by time sequence misalignment; one-dimensional formatting preprocessing normalizes the two types of physical quantities into a unified data format, providing a data foundation for subsequent feature concatenation with random noise vectors.

[0092] S250. The Gaussian distributed random noise vector and the condition variable are concatenated by their feature dimensions and used as the input data for the generator network in the conditional generative adversarial network.

[0093] In this embodiment, the Gaussian distribution can refer to a normal distribution with a mean of zero and a variance of a constant (e.g., 1). The random noise vector can refer to a vector randomly sampled from the Gaussian distribution. The random noise vector can provide a source of randomness for the generator network's input. The generator network can refer to the network module in a conditional generative adversarial network responsible for generating virtual moisture feature data. The input data can refer to the input of the generator network. The input data can be constructed by concatenating the random noise vector and conditional variables along the feature dimension.

[0094] Specifically, the central server concatenates a Gaussian-distributed random noise vector with a condition variable by adding feature dimensions, and uses this as input data for the generator network. The random noise vector introduces randomness into the generation process, while the condition variable controls the physical state category of the generated data (i.e., corresponding to a specific core temperature and effective core stress state).

[0095] This step uses a concatenation of random noise vectors and physical condition variables as input to the generator network, allowing the network to generate data randomly while being constrained by physical conditions. Random noise ensures the diversity of generated samples and prevents pattern collapse; the condition variables ensure that the generated virtual moisture feature data corresponds to specific thermodynamic and stress states, making the generated data physically interpretable.

[0096] S260. Calculate the dynamic conversion rate of bound water to free water using the effective stress and temperature of the core.

[0097] In this embodiment, bound water refers to water molecules within the tobacco leaf that are bound by the forces of the solid surface and whose rotation is hindered. Bound water molecules experience strong intermolecular forces, resulting in a slow rate of change in response to an external alternating electromagnetic field, primarily exhibiting low-frequency interfacial polarization characteristics. Free water refers to water molecules within the tobacco leaf that possess a large dipole orientation space and can rotate freely. Free water molecules respond quickly to an external alternating electromagnetic field, corresponding to high-frequency polarization relaxation characteristics, and are direct participants in microbial metabolic activities. The dynamic conversion rate refers to the rate at which bound water molecules desorb and convert into free water molecules under given core temperature and effective core stress conditions. The introduction of external pressure alters the porosity and capillary forces within the tobacco leaf, reducing the activation energy required for the desorption of bound water, thereby driving the conversion of bound water into free water.

[0098] Specifically, the introduction of external pressure alters the porosity and capillary forces within the tobacco leaf, thereby reducing the activation energy required for bound hydrolysis and adsorption within the tobacco cell structure. Based on this physical mechanism, the central server, combining the core-coating temperature uploaded by the edge gateway with the calculated effective core-coating stress, uses the modified Arrhenius equation to calculate the water kinetic conversion parameters. The formula for calculating the dynamic conversion rate is: .

[0099] in, The value represents the dynamic conversion rate of bound water to free water, A represents the pre-exponential factor, and exp represents an exponential function with the natural constant as its base. This represents the activation energy of basic desorption under stress-free conditions. Represents the stress-energy coupling coefficient. R represents the effective stress within the core, and R represents the ideal gas constant. This indicates the core temperature (using thermodynamic absolute temperature units).

[0100] This step maps the two macroscopic physical quantities of effective core stress and core temperature to the dynamic conversion rate of bound water to free water using the modified Arrhenius equation, establishing a quantitative physical model between macroscopic packing state and microscopic moisture phase transformation. This dynamic conversion rate reflects the changing trend of the active state of moisture inside the tobacco leaf; that is, the higher the conversion rate, the more bound water is converted into free water that can be used by microorganisms, and the greater the risk of mold growth.

[0101] S270. Introducing the dynamic conversion rate as a physical penalty term into the training process of the conditional generative adversarial network generates virtual moisture feature data.

[0102] In this embodiment, the physical penalty term can refer to a penalty function constructed based on physical laws (water dynamics transformation equations). The physical penalty term can be used to constrain the generator network during the training process of the conditional generative adversarial network (GAN), ensuring that the virtual water feature data it generates conforms to preset thermodynamic and kinetic transformation equations. The training process of the GAN can refer to an adversarial training process in which the generator network and the discriminator network alternately optimize. The generator network is responsible for generating virtual water feature data, and the discriminator network is responsible for distinguishing whether the input data is real feature data or generated virtual data.

[0103] Specifically, the central server substitutes the dielectric characteristics of free water and bound water uploaded from the edge gateway into the ordinary differential equation, and combines this with the dynamic conversion rate of bound water to free water to construct a physical penalty term reflecting the water conversion law. The central server adds this physical penalty term to the binary cross-entropy loss function of the discriminator network, forming a joint loss function. During the adversarial training phase, the generator network is responsible for generating virtual water feature data using random noise and conditional variables, while the discriminator network is responsible for distinguishing whether the input data comes from real feature data uploaded from the edge gateway or virtual water feature data generated by the generator network. The discriminator network provides gradient feedback guidance to the generator network through the joint loss function, ensuring that the virtual water feature data generated by the generator network must follow the thermodynamic conversion law corrected by the core-envelope effective stress transfer model. The central server uses the joint loss function to perform alternating optimization and iterative training of the conditional generative adversarial network until convergence is achieved.

[0104] This step introduces a physical penalty term, constructed based on the water dynamics transformation equation, into the joint loss function of the conditional generative adversarial network. This allows the generator network to be guided by both data-driven factors (discriminator feedback) and physical laws (penalty term constraints) during adversarial training. This effectively avoids invalid data that does not conform to thermodynamic states, resulting from purely data-driven generation, ensuring that the generated virtual water feature data conforms to both data distribution characteristics and physical evolution laws.

[0105] As an optional implementation of this disclosure, the specific process of constructing a physical penalty term and using the physical penalty term to train the conditional generative adversarial network includes: calculating the rate of change of the free water dielectric feature over time; subtracting the product of the dynamic conversion rate and the bound water dielectric feature from the rate of change of the free water dielectric feature over time, squared, and multiplied by the penalty term weight coefficient to construct a physical penalty term; adding the physical penalty term to the binary cross-entropy loss function of the discriminator network in the conditional generative adversarial network to form a joint loss function; using the joint loss function to train the conditional generative adversarial network, generating and outputting virtual water feature data composed of virtual free water dielectric features and virtual bound water dielectric features.

[0106] In this embodiment, the penalty term weight coefficient can refer to a weight parameter used to balance the difference in magnitude between physical constraints and data-driven errors. The value of the penalty term weight coefficient is set between 0.01 and 0.1 according to the magnitude of the gradient loss in the early stage of network training. The discriminator network can refer to the network module in a conditional generative adversarial network responsible for distinguishing whether the input data is real or generated. The joint loss function can refer to the comprehensive loss function composed of the binary cross-entropy loss function and the physical penalty term. The virtual free water dielectric feature can refer to the virtual free water dielectric feature data output by the generator network. The virtual bound water dielectric feature can refer to the virtual bound water dielectric feature data output by the generator network.

[0107] Specifically, the central server substitutes the dielectric characteristics of free water and bound water uploaded by the edge gateways into the ordinary differential equation, and combines this with the dynamic conversion rate of bound water to free water to construct a physical penalty term reflecting the water conversion law. The formula for calculating the physical penalty term is: .

[0108] in, Indicates the physical penalty item. This represents the weighting coefficient of the penalty term. This represents the rate of change of the dielectric properties of free water over time (the calculation uses a forward difference approximation between adjacent discrete time steps to replace the differential operation). This represents the dynamic conversion rate of bound water to free water. This indicates the dielectric characteristics of bound water. This represents the theoretical generation rate of the dielectric characteristics of free water derived from a kinetic model. This represents the dynamic evolution residual, i.e., the deviation between the actual rate of change and the theoretical rate of change of the feature data. The central server adds a physical penalty term to the binary cross-entropy loss function of the discriminator network, forming a joint loss function. During the adversarial training phase, the generator network is responsible for generating virtual water feature data (composed of virtual free water dielectric features and virtual bound water dielectric features) using random noise and conditional variables. The discriminator network is responsible for distinguishing whether the input data is real feature data uploaded from the edge gateway or virtual water feature data generated by the generator network. The central server uses the joint loss function to alternately optimize and iteratively train the conditional generative adversarial network until convergence is achieved.

[0109] This optional implementation achieves precise numerical penalties for deviations from thermodynamic transformation characteristics by using the square of the kinetic evolution residual as the specific mathematical form of the physical penalty term. The penalty term weight coefficient is used to balance the magnitude difference between physical constraints and data-driven errors, ensuring that the network can fully utilize information from real data while strictly adhering to physical constraints during training. The generated virtual water feature data is constructed by concatenating virtual free water dielectric features and virtual bound water dielectric features, maintaining a data structure consistent with the real feature data.

[0110] S280. Obtain the real feature data uploaded by the edge gateway, and merge the virtual moisture feature data with the real feature data to form a balanced spatiotemporal dataset.

[0111] S290. Construct a high-dimensional feature matrix using the free water dielectric characteristics, bound water dielectric characteristics, core temperature, core effective stress, and storage environment time series parameters contained in the equilibrium spatiotemporal dataset. Perform dimensionality reduction processing on the high-dimensional feature matrix to output the intrinsic time series feature set.

[0112] S2100. The intrinsic temporal feature set is used to update and train the network weights of the two-branch gated recurrent unit prediction model, and the tobacco leaf moisture content prediction model is completed.

[0113] As an optional implementation of this disclosure, after completing the construction of the tobacco leaf moisture content prediction model, the method further includes: outputting the predicted mold probability value through a dual-branch gated loop unit prediction model; obtaining the historical high-frequency intrinsic mode function components of mold-free tobacco leaves, calculating the average energy of the historical high-frequency intrinsic mode function components for calibration, and obtaining a free water safety benchmark threshold; subtracting the free water safety benchmark threshold from the free water dielectric characteristics at the current time step to obtain the difference between free water exceeding the safety benchmark, and obtaining the effective free water deviation through a non-negative truncation function; and adding the product of the predicted mold probability value and a preset prediction probability weight coefficient to the product of the effective free water deviation and a preset current state deviation weight coefficient to calculate the dynamic intervention index.

[0114] In this embodiment, the predicted mold probability value can refer to the probability value of tobacco leaves becoming moldy within a preset time period output by the dual-branch gated cyclic unit prediction model, with a value ranging from 0 to 1. Mold-free tobacco leaves can refer to tobacco leaf samples that have been confirmed not to have become moldy during historical storage. Historical high-frequency intrinsic mode function components can refer to the high-frequency intrinsic mode function components obtained from historical data of mold-free tobacco leaves through complementary integrated empirical mode decomposition. The energy mean can refer to the cumulative average of the squared voltage amplitudes of the high-frequency intrinsic mode function components on the equivalent time axis. The free water safety benchmark threshold can refer to the energy mean of the high-frequency intrinsic mode function components of mold-free tobacco leaves. The free water safety benchmark threshold can be used as a benchmark value to determine whether the current free water dielectric characteristics exceed the safe range.

[0115] In this embodiment, the current time step can refer to the time index corresponding to the current monitoring moment. The non-negative truncation function can be a function that takes the maximum value between the input value and 0. The non-negative truncation function can be used to ensure that the deviation of the calculated output will not be negative. The free water deviation can be the effective deviation obtained by non-negative truncation of the difference between the free water dielectric characteristic of the current time step and the free water safety benchmark threshold. The prediction probability weighting coefficient can be the weight of the predicted mold probability value in the calculation of the dynamic intervention index, with a value ranging from 0.6 to 0.8. The current state deviation weighting coefficient can be the weight of the free water deviation in the calculation of the dynamic intervention index, with a value ranging from 0.2 to 0.4, and its sum with the prediction probability weighting coefficient is 1. The dynamic intervention index can be a quantitative index calculated by comprehensively considering the predicted mold probability and the degree of deviation of the current free water state. The dynamic intervention index can be used to characterize the degree to which the internal physical state of tobacco leaves deviates from the safety benchmark.

[0116] Specifically, during the online model validation and continuous optimization phase, the central server inputs the intrinsic time-series feature set, acquired in real-time and processed by nonlinear dimensionality reduction, into the trained dual-branch gated recurrent unit prediction model, outputting the predicted mold growth probability value for a preset future time period. The central server then calculates the dynamic intervention index by combining the predicted mold growth probability value with the system state parameters at the current time step. The formula for calculating the dynamic intervention index is: .

[0117] in, This indicates a dynamic intervention index. Represents the prediction probability weighting coefficient. This represents the predicted probability of mold growth. This indicates that the current state deviates from the weight coefficient. This represents the dielectric properties of free water at the current time step. The free water safety benchmark threshold is represented by the average energy of the high-frequency intrinsic mode function components extracted from tobacco leaves without mold at standard moisture content.

[0118] This optional implementation calculates a dynamic intervention index by combining the comprehensive model prediction results (foresight information) with the current actual state (real-time monitoring information), thus achieving a multi-dimensional quantitative assessment of the risk of tobacco mold growth. When the model predicts that mold growth may occur in the future or the current free water level has exceeded the safety benchmark, the dynamic intervention index increases accordingly, providing a quantitative basis for subsequent triggering of pulse reconstruction and environmental regulation.

[0119] As an optional implementation of this disclosure, after calculating and obtaining the dynamic intervention index, the method further includes: triggering a pulse reconstruction operation based on the dynamic intervention index, dynamically compressing the initial pulse fall edge time of the initial broadband pulse sequence sent to the flexible microstrip probe arranged inside the cigarette pack, and generating a reconstructed broadband pulse sequence; sending the reconstructed broadband pulse sequence to obtain a reconstructed echo sequence, and performing downsampling processing on the reconstructed echo sequence using equivalent time sampling to generate an updated downsampled echo sequence; supplementing the updated downsampled echo sequence and the corresponding storage environment time series parameters of the current time step into the balanced spatiotemporal dataset; and using the supplemented balanced spatiotemporal dataset to perform incremental training and weight updates on the dual-branch gated recurrent unit prediction model.

[0120] In this embodiment, the pulse reconstruction operation can refer to a hardware control operation that adjusts the pulse fall-edge time based on a dynamic intervention index. The pulse reconstruction operation can change the spectral energy distribution of the transmitted signal by compressing the initial pulse fall-edge time, thereby exciting higher-frequency microwave harmonic components in the frequency domain. The reconstructed broadband pulse sequence can refer to a broadband pulse sequence with compressed pulse fall-edge time generated after the pulse reconstruction operation. The reconstructed echo sequence can refer to the corresponding high-frequency electromagnetic wave reflection signal sequence obtained after sending the reconstructed broadband pulse sequence. The updated downsampled echo sequence can refer to the updated sequence generated after downsampling the reconstructed echo sequence using equivalent time sampling. The storage environment timing parameters at the current time step can refer to the storage environment temperature and humidity parameters at the same time step as the updated downsampled echo sequence. The supplemented balanced spatiotemporal dataset can refer to a new dataset formed by supplementing the updated downsampled echo sequence and the corresponding storage environment timing parameters to the original balanced spatiotemporal dataset. Incremental training can refer to further training the model using the newly supplemented data based on an existing trained model. Weight update can refer to adjusting and updating the network weight parameters of a two-branch gated recurrent unit prediction model during incremental training.

[0121] Specifically, the central server executes multi-level system closed-loop coordination based on the calculated dynamic intervention index. In the perception feedforward dimension, the central server sends the dynamic intervention index to the acquisition terminal via a wireless communication network. The acquisition terminal triggers the pulse distortion reconstruction mechanism of its internal hardware circuitry based on the received dynamic intervention index. The internal hardware circuitry of the acquisition terminal includes a field-programmable gate array (FPGA) and a programmable digital delay line (PDL) chip. Based on the received dynamic intervention index, the FPGA sends digital control commands to the PDL chip, dynamically compressing the initial pulse fall edge time of the initial broadband pulse sequence by adjusting the number of cascaded logic gate delay units enabled on the internal signal transmission path, thereby generating a reconstructed broadband pulse sequence. The pulse fall edge time of the reconstructed broadband pulse sequence is calculated in conjunction with a lower bound constraint threshold; the specific calculation formula is as follows: .

[0122] in, This represents the pulse falling edge time of the reconstructed broadband pulse sequence. This represents the minimum physical delay time threshold that a programmable digital delay line chip can output. This represents the falling edge time of the initial pulse in the initial broadband pulse sequence. Indicates the pulse distortion adjustment coefficient. This represents the dynamic intervention index. The compression of the pulse falling edge time in the reconstructed broadband pulse sequence alters the spectral energy distribution of the transmitted signal, exciting higher-frequency microwave harmonic components in the frequency domain. Since water molecules in the transition stage from bound water to free water require higher excitation frequencies to generate resonant absorption, these higher-frequency microwave harmonic components can match the dielectric relaxation response frequency of water molecules in the tobacco leaf during this transition, thus enabling targeted capture of early abnormal moisture states. The acquisition terminal sends the reconstructed broadband pulse sequence to obtain the reconstructed echo sequence. The reconstructed echo sequence is downsampled using equivalent time sampling to generate an updated downsampled echo sequence. This updated downsampled echo sequence, along with the corresponding storage environment time-series parameters for the current time step, is added to the balanced spatiotemporal dataset. The central server uses the supplemented balanced spatiotemporal dataset to incrementally train and update the weights of the dual-branch gated recurrent unit prediction model.

[0123] This optional implementation establishes a direct feedback mechanism between the prediction results and the underlying physical detection equipment by dynamically intervening in the exponentially triggered pulse reconstruction operation. When the model predicts that the moisture state deviates from the safety benchmark, the system automatically compresses the pulse fall edge time to excite higher frequency microwave harmonic components, making the detection signal match the dielectric relaxation response frequency of water molecules in the transition phase, thus achieving targeted capture of early abnormal moisture states. By supplementing the dataset with newly acquired difficult samples and triggering incremental training, a complete closed loop of prediction, sensing, acquisition, and updating is achieved, enabling the prediction model to be continuously optimized in actual operation.

[0124] It should be noted that when the dynamic intervention index exceeds the preset safety tolerance threshold, a pulse reconstruction operation is triggered; the safety tolerance threshold is a small positive number calculated statistically based on the background noise of historical samples without mold. Furthermore, when the dynamic intervention index exceeds the preset intervention initiation threshold (range 0.75-0.85), an environmental control intervention command is generated; during the operation of the environmental control equipment, the dynamic intervention index is continuously monitored, and environmental control is stopped when the dynamic intervention index falls below the preset intervention termination threshold (range 0.4-0.5); the intervention initiation threshold and intervention termination threshold are obtained by calibrating the distribution of the dynamic intervention index of historical critical mold samples through normal fitting.

[0125] The technical solution of this invention combines core temperature and effective core stress with timestamp alignment and one-dimensional formatting preprocessing to form condition variables, ensuring strict synchronization of physical condition variables in the time dimension and standardization of data format. By concatenating Gaussian distributed random noise vectors with the condition variables in a feature dimension as input data to the generator network in the conditional generative adversarial network, the generator network is constrained by physical state conditions while possessing random generation capabilities, ensuring both the diversity of generated samples and the interpretability of generated data in a physical sense. By calculating the dynamic conversion rate of bound water to free water using effective core stress and core temperature, a quantitative physical model between macroscopic stacking state and microscopic water phase transformation is established. By introducing the dynamic conversion rate as a physical penalty term into the training process of the conditional generative adversarial network, virtual water feature data is generated. The data generation process of the generator network is constrained by physical laws, effectively avoiding the generation of invalid data that does not conform to thermodynamic state, so that the generated virtual water feature data conforms to both data distribution characteristics and follows physical evolution laws.

[0126] The tobacco leaf moisture content prediction model construction device provided in this embodiment of the invention can execute the tobacco leaf moisture content prediction model construction method provided in any embodiment of the invention, and has the corresponding beneficial effects of executing the tobacco leaf moisture content prediction model construction method.

[0127] The following are embodiments of the tobacco leaf moisture content prediction model construction device provided in this invention. This device and the tobacco leaf moisture content prediction model construction method in the above embodiments belong to the same inventive concept. For details not described in detail in the embodiments of the tobacco leaf moisture content prediction model construction device, please refer to the embodiments of the tobacco leaf moisture content prediction model construction method.

[0128] Example 3 Figure 3 This is a schematic diagram of a tobacco leaf moisture content prediction model construction device provided in Embodiment 3 of the present invention. Figure 3 As shown, the device includes: a parameter acquisition module 310, a feature extraction module 320, a core effective stress calculation module 330, an equilibrium spatiotemporal dataset construction module 340, an intrinsic temporal feature set output module 350, and a model construction module 360.

[0129] The system includes the following modules: Parameter Acquisition Module 310, which acquires the core temperature of the tobacco leaves, time-series parameters of the storage environment, and downsampled echo sequences; Feature Extraction Module 320, which decomposes the downsampled echo sequences to extract free water dielectric features and bound water dielectric features; Core Effective Stress Calculation Module 330, which acquires the packaging physical parameters of the outer packaging box of the tobacco bales and the three-dimensional spatial coordinates of the tobacco bales in the storage stack, and calculates the core effective stress; and Equilibrium Spatiotemporal Dataset Construction Module 340, which processes the core temperature and core effective stress using a conditional generative adversarial network to generate and output virtual moisture feature data. The system acquires real feature data uploaded by the edge gateway and merges the virtual moisture feature data with the real feature data to form a balanced spatiotemporal dataset. The intrinsic temporal feature set output module 350 is used to construct a high-dimensional feature matrix using the free water dielectric features, bound water dielectric features, core temperature, core effective stress, and storage environment temporal parameters contained in the balanced spatiotemporal dataset. The high-dimensional feature matrix is ​​then dimensionality-reduced to output the intrinsic temporal feature set. The model building module 360 ​​is used to update and train the network weights of the dual-branch gated recurrent unit prediction model using the intrinsic temporal feature set to complete the construction of the tobacco leaf moisture content prediction model.

[0130] The technical solution of this invention provides a time-consistent data foundation for subsequent multi-source feature fusion and physical model construction by acquiring the core temperature of tobacco leaves, storage environment time-series parameters, and downsampled echo sequences. By decomposing the downsampled echo sequences and extracting free water dielectric features and bound water dielectric features, the single macroscopic moisture content index is decomposed into moisture features of different physical binding states, eliminating the masking effect of total moisture parameters on the internal active state of moisture and improving the distinguishability of moisture feature data for mold risk. By acquiring the packaging physical parameters of the outer packaging box of the tobacco bales and the three-dimensional spatial coordinates of the tobacco bales in the storage stack, the effective stress of the core is calculated, providing a physically realistic stress input for calculating the moisture kinetic conversion rate. Virtual moisture feature data is generated and output by processing the core temperature and effective core stress through a conditional generative adversarial network, obtaining the real moisture feature data uploaded by the edge gateway. The feature data is constructed by merging virtual moisture feature data with real feature data to form a balanced spatiotemporal dataset. The generation of virtual data is constrained by physical laws, which solves the problem of severe imbalance in model training data distribution caused by the scarcity of marginal abnormal state samples and avoids the generation of invalid data that does not conform to physical laws. By using the free water dielectric characteristics, bound water dielectric characteristics, core temperature, core effective stress and storage environment time series parameters contained in the balanced spatiotemporal dataset to construct a high-dimensional feature matrix, the high-dimensional feature matrix is ​​dimensionality reduced to output an intrinsic time series feature set, which effectively removes multicollinearity redundant features and noise interference, and extracts pure intrinsic variables that reflect the thermodynamic transformation state of moisture. By using the intrinsic time series feature set to update and train the network weights of the bi-branch gated recurrent unit prediction model, the tobacco leaf moisture content prediction model is completed, realizing accurate prediction of tobacco leaf moisture content and its mold risk at multiple time scales.

[0131] Based on the above technical solution, the core effective stress calculation module 330 is specifically used to: obtain the three-dimensional spatial coordinates of the tobacco bales in the warehouse stack, the total height of the tobacco stack, and the equivalent density parameters of tobacco bales at different height levels; extract the height coordinates from the three-dimensional spatial coordinates and perform integral calculations with the total height of the tobacco stack and the equivalent density parameters of tobacco bales at different height levels to obtain the macroscopic stacking integral; obtain the structural yield threshold and internal stress transmission coefficient of the outer packaging box; and calculate the core effective stress by multiplying the macroscopic stacking integral, the structural yield threshold of the outer packaging box, and the internal stress transmission coefficient.

[0132] Based on the above technical solution, the balanced spatiotemporal dataset composition module 340 may include: The condition variable component submodule is used to combine the core temperature and the effective core stress with timestamp alignment and one-dimensional formatting preprocessing to form condition variables. The input data generation submodule is used to concatenate the Gaussian distributed random noise vector with the condition variable by adding feature dimensions as input data to the generator network in the conditional generative adversarial network. The dynamic conversion rate calculation submodule is used to calculate the dynamic conversion rate of bound water to free water using the effective stress and temperature of the core. The virtual moisture feature data generation submodule is used to generate virtual moisture feature data by introducing the dynamic conversion rate as a physical penalty term into the training process of the conditional generative adversarial network.

[0133] Based on the above technical solution, the virtual water feature data generation submodule is specifically used for: calculating the rate of change of free water dielectric features over time; subtracting the product of the dynamic conversion rate and the bound water dielectric features from the rate of change of free water dielectric features over time, squared, and multiplying by the penalty term weight coefficient to construct a physical penalty term; adding the physical penalty term to the binary cross-entropy loss function of the discriminator network in the conditional generative adversarial network to form a joint loss function; using the joint loss function to train the conditional generative adversarial network, generating and outputting virtual water feature data composed of virtual free water dielectric features and virtual bound water dielectric features.

[0134] Based on the above technical solution, the parameter acquisition module 310 is specifically used to: send an initial broadband pulse sequence with an initial pulse falling edge time to a flexible microstrip probe arranged inside the cigarette pack through the device end, and acquire an initial echo sequence; and perform downsampling processing on the initial echo sequence using equivalent time sampling to generate a downsampled echo sequence.

[0135] Based on the above technical solution, the feature extraction module 320 is specifically used for: adding positive white noise sequences and negative white noise sequences multiple times to the downsampled echo sequence to construct multiple reconstruction analysis sequences; performing empirical mode decomposition on each reconstruction analysis sequence to generate modal components; and performing overall average calculation on the modal components obtained by decomposing the reconstruction analysis sequence to obtain multiple intrinsic mode function components; calculating the center frequency of each intrinsic mode function component; based on a preset frequency cutoff threshold, classifying the intrinsic mode function components with a center frequency greater than the preset frequency cutoff threshold into high-frequency intrinsic mode function components and extracting them as free water dielectric features; and classifying the intrinsic mode function components with a center frequency less than or equal to the preset frequency cutoff threshold into low-frequency intrinsic mode function components and extracting them as bound water dielectric features.

[0136] Based on the above technical solution, the intrinsic time series feature set output module 350 is specifically used for: calculating the Euclidean distance between sample points in the high-dimensional feature matrix using a two-stage equidistant mapping algorithm, selecting the nearest neighbor points to establish connecting edges and construct a local neighborhood graph; calculating the shortest path length between any two sample points in the local neighborhood graph using a shortest path algorithm as the geodesic distance; combining the geodesic distances between all sample points to form a distance matrix; performing operations on the distance matrix using a centering transformation matrix to obtain a double-centered inner product matrix, performing eigenvalue decomposition on the double-centered inner product matrix to extract eigenvalue sequences and eigenvectors, and constructing and outputting the intrinsic time series feature set based on the eigenvalue sequences and eigenvectors.

[0137] Based on the above technical solution, the model building module 360 ​​is specifically used for: dividing the intrinsic temporal feature set into a temporal feature training set and a temporal feature test set; using the temporal feature training set to train the dual-branch gated recurrent unit prediction model; concatenating the macroscopic evolution feature vector output by the long-term trend branch network and the microscopic mutation feature vector output by the short-term perturbation branch network in the dual-branch gated recurrent unit prediction model by feature dimension through the dual-branch feature fusion layer; and updating the network weights of the dual-branch gated recurrent unit prediction model using the binary cross-entropy loss function.

[0138] Based on the above technical solution, after completing the construction of the tobacco leaf moisture content prediction model, the device also includes: The module for outputting the predicted mold probability value is used to output the predicted mold probability value through a prediction model with a dual-branch gated loop unit. The free water safety benchmark threshold determination module is used to obtain the historical high-frequency intrinsic mode function components of unmoldy tobacco leaves, calculate the average energy of the historical high-frequency intrinsic mode function components for calibration, and obtain the free water safety benchmark threshold. The free water deviation determination module is used to obtain the difference between the free water and the safety benchmark threshold by subtracting the free water dielectric characteristics at the current time step, and obtain the effective free water deviation through a non-negative truncation function. The dynamic intervention index determination module is used to calculate the dynamic intervention index by adding the product of the predicted mold probability value and the preset predicted probability weight coefficient to the product of the effective free water deviation and the preset current state deviation weight coefficient.

[0139] Based on the above technical solution, after calculating and obtaining the dynamic intervention index, the device further includes: The reconstructed broadband pulse sequence generation module is used to dynamically compress the initial pulse fall time of the initial broadband pulse sequence sent to the flexible microstrip probe arranged inside the cigarette pack based on the dynamic intervention index triggering pulse reconstruction operation, and generate the reconstructed broadband pulse sequence. The downsampled echo sequence generation module is used to send a reconstructed wideband pulse sequence to obtain a reconstructed echo sequence, and to perform downsampling processing on the reconstructed echo sequence using equivalent time sampling to generate an updated downsampled echo sequence. The balanced spatiotemporal dataset supplementation module is used to supplement the balanced spatiotemporal dataset with the updated downsampled echo sequence and the corresponding storage environment time series parameters of the current time step; The weight update module is used to perform incremental training and weight updates on the bi-branch gated recurrent unit prediction model using the supplemented balanced spatiotemporal dataset.

[0140] The tobacco leaf moisture content prediction model construction device provided in this embodiment of the invention can execute the tobacco leaf moisture content prediction model construction method provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of executing the tobacco leaf moisture content prediction model construction method.

[0141] It is worth noting that in the above embodiments of the tobacco leaf moisture content prediction model, the various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy differentiation and are not used to limit the scope of protection of the present invention.

[0142] Example 4 Figure 4 A schematic diagram of an electronic device 10, which can be used to implement embodiments of the present invention, is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices (e.g., helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0143] like Figure 4As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded from storage unit 18 into the RAM 13. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0144] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0145] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as the method for building a tobacco leaf moisture content prediction model.

[0146] In some embodiments, the method for constructing a tobacco leaf moisture content prediction model can be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program can be loaded and / or installed on electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the method for constructing a tobacco leaf moisture content prediction model described above can be performed. Alternatively, in other embodiments, processor 11 can be configured to execute the method for constructing a tobacco leaf moisture content prediction model by any other suitable means (e.g., by means of firmware).

[0147] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0148] Computer programs used to implement the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0149] In the context of this invention, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0150] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0151] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0152] A computing system can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system to address the shortcomings of traditional physical hosts and VPS services, such as high management difficulty and weak business scalability.

[0153] This application also discloses a computer program product, which includes a computer program that, when executed by a processor, implements the tobacco leaf moisture content prediction model construction method provided in any embodiment of this application. This program product and the tobacco leaf moisture content prediction model construction method disclosed in the embodiments of this application belong to the same inventive concept, and therefore will not be described in detail here.

[0154] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0155] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for constructing a prediction model for tobacco leaf moisture content, characterized in that, include: Acquire the core temperature of tobacco leaves, time-series parameters of storage environment, and downsampled echo sequences; The downsampled echo sequence is decomposed to extract the dielectric characteristics of free water and bound water. Obtain the physical parameters of the outer packaging box of the cigarette pack and the three-dimensional spatial coordinates of the cigarette pack in the warehouse stacking, and calculate the effective stress of the pack core; Based on the core temperature and the effective core stress, virtual moisture feature data is generated and output through a conditional generative adversarial network. Real feature data uploaded by the edge gateway is obtained, and the virtual moisture feature data and the real feature data are merged to form a balanced spatiotemporal dataset. A high-dimensional feature matrix is ​​constructed using the free water dielectric characteristics, bound water dielectric characteristics, core temperature, core effective stress, and storage environment time series parameters contained in the balanced spatiotemporal dataset. The high-dimensional feature matrix is ​​then dimensionality-reduced to output an intrinsic time series feature set. The intrinsic temporal feature set is used to update and train the network weights of the dual-branch gated recurrent unit prediction model, thus completing the construction of the tobacco leaf moisture content prediction model.

2. The method according to claim 1, characterized in that, The specific steps for calculating and obtaining the effective stress of the core include: Obtain the three-dimensional spatial coordinates of the tobacco bales in the warehouse stack, the total height of the tobacco stack, and the equivalent density parameters of the tobacco bales at different height levels; The height coordinates in the three-dimensional spatial coordinates are extracted and combined with the total height of the tobacco stack and the equivalent density parameters of the tobacco bales at different height levels to perform an integral calculation to obtain the macroscopic stacking integral; Obtain the structural yield threshold and internal stress transmission coefficient of the outer packaging box; The effective stress of the core is calculated by multiplying the macroscopic compressibility integral, the structural yield threshold of the outer packaging box, and the internal stress transmission coefficient.

3. The method according to claim 1, characterized in that, The process of generating and outputting virtual moisture characteristic data based on the core temperature and the effective core stress using a conditional generative adversarial network includes: The core temperature and the effective core stress are combined with timestamp alignment and one-dimensional formatting preprocessing to form conditional variables; The Gaussian distributed random noise vector is concatenated with the condition variable by its feature dimension and used as the input data for the generator network in the conditional generative adversarial network. The dynamic conversion rate of bound water to free water is calculated using the effective stress and temperature of the core. The dynamic conversion rate is introduced as a physical penalty term into the training process of the conditional generative adversarial network to generate the virtual moisture feature data.

4. The method according to claim 3, characterized in that, The specific process of constructing a physical penalty term and using the physical penalty term to train the conditional generative adversarial network includes: Calculate the rate of change of the dielectric properties of the free water over time; The physical penalty term is constructed by squared the product of the dynamic conversion rate and the bound water dielectric characteristic and subtracting the rate of change of the free water dielectric characteristic over time, and then multiplying the square by the penalty term weighting coefficient. The physical penalty term is added to the binary cross-entropy loss function of the discriminator network in the conditional generative adversarial network to form a joint loss function; The conditional generative adversarial network is trained using the joint loss function to generate and output virtual water feature data, which is composed of virtual free water dielectric features and virtual bound water dielectric features.

5. The method according to claim 1, characterized in that, Obtain the downsampled echo sequence, including: The device sends an initial broadband pulse sequence with an initial pulse fall edge time to a flexible microstrip probe arranged inside the cigarette pack to obtain the initial echo sequence. The initial echo sequence is downsampled using equivalent time sampling to generate the downsampled echo sequence.

6. The method according to claim 1, characterized in that, The process of decomposing the downsampled echo sequence to extract the dielectric features of free water and bound water includes: Multiple reconstruction analysis sequences are constructed by repeatedly adding positive white noise sequences and negative white noise sequences to the downsampled echo sequence. Empirical mode decomposition is performed on each reconstruction analysis sequence to generate modal components. The modal components obtained by the reconstruction analysis sequence decomposition are then averaged to obtain multiple intrinsic mode function components. The center frequency of each intrinsic mode function component is calculated. Based on a preset frequency cutoff threshold, the intrinsic mode function components with a center frequency greater than the preset frequency cutoff threshold are classified as high-frequency intrinsic mode function components and extracted as the free water dielectric characteristics. The intrinsic mode function components with a center frequency less than or equal to the preset frequency cutoff threshold are classified as low-frequency intrinsic mode function components and extracted as the bound water dielectric characteristics.

7. The method according to claim 1, characterized in that, The step of reducing the dimensionality of the high-dimensional feature matrix to output the intrinsic time-series feature set includes: A two-stage equidistant mapping algorithm is used to calculate the Euclidean distance between sample points in the high-dimensional feature matrix, and the nearest neighbor points are selected to establish connecting edges to construct a local neighborhood graph. The shortest path algorithm is used to calculate the shortest path length between any two sample points in the local neighborhood graph as the geodesic distance; The geodesic distances between all the sample points are combined to form a distance matrix; The distance matrix is ​​operated on using a centering transformation matrix to obtain a double-centered inner product matrix. Eigenvalue decomposition is performed on the double-centered inner product matrix to extract eigenvalue sequences and eigenvectors. The intrinsic time series feature set is constructed and output based on the eigenvalue sequences and eigenvectors.

8. The method according to claim 1, characterized in that, The step of updating and training network weights for the dual-branch gated recurrent unit prediction model using the intrinsic temporal feature set includes: The intrinsic temporal feature set is divided into a temporal feature training set and a temporal feature test set. The temporal feature training set is used to train the dual-branch gated recurrent unit prediction model. The macroscopic evolution feature vector output by the long-term trend branch network and the microscopic mutation feature vector output by the short-term perturbation branch network in the dual-branch gated recurrent unit prediction model are concatenated by feature dimension through the dual-branch feature fusion layer. The binary cross-entropy loss function is used to update the network weights of the dual-branch gated recurrent unit prediction model.

9. The method according to claim 1, characterized in that, After completing the construction of the tobacco leaf moisture content prediction model, the method further includes: The predicted mold probability value is output by the prediction model of the dual-branch gated loop unit. The intrinsic mode function components of historical high-frequency bands of unmoldy tobacco leaves are obtained, and the average energy of the intrinsic mode function components of historical high-frequency bands is calculated for calibration to obtain the free water safety benchmark threshold. The difference between the free water and the safety benchmark threshold is obtained by subtracting the free water dielectric characteristics at the current time step, and the effective free water deviation is obtained by using a non-negative truncation function. The dynamic intervention index is calculated by adding the product of the predicted mold probability value and the preset predicted probability weighting coefficient to the product of the effective free water deviation and the preset current state deviation weighting coefficient.

10. The method according to claim 1, characterized in that, After calculating and obtaining the dynamic intervention index, the method further includes: Based on the dynamic intervention index triggering pulse reconstruction operation, the initial pulse fall time of the initial broadband pulse sequence sent to the flexible microstrip probe arranged inside the cigarette pack is dynamically compressed to generate a reconstructed broadband pulse sequence. Send the reconstructed wideband pulse sequence to obtain the reconstructed echo sequence, and perform downsampling processing on the reconstructed echo sequence using equivalent time sampling to generate an updated downsampled echo sequence; The updated downsampled echo sequence and the corresponding storage environment time series parameters at the current time step are added to the balanced spatiotemporal dataset; The supplemented balanced spatiotemporal dataset is used to perform incremental training and weight updates on the bi-branch gated recurrent unit prediction model.

Images