A method and system for predicting the quality deterioration of fresh produce based on multi-modal dynamic coupling

Abstract

The application provides a fresh food quality deterioration prediction method and system based on multi-modal dynamic coupling, aiming to solve the prediction lag problem caused by data isolation and lack of dynamic coupling in traditional storage and transportation process. Specifically, multi-source sensors are used to synchronously collect sensory layer data, environmental layer data and operation layer data, and the sensory data is subjected to color difference quantization, chemical bond feature extraction and spatial chemical fusion processing to generate quality deterioration feature tensor. Next, based on the graph network model and frequency domain analysis, the dynamic coupling parameter set of quality deterioration features and environmental parameters is constructed. Then, the unloading pulse trajectory tensor, cold chain stability distribution and coupling parameter set are fused to generate a deterioration constraint distribution matrix through multivariate weight mapping. Finally, combined with the acoustic resonance mode coherence analysis and the constraint matrix evolution, a four-level quality label is output. This method significantly reduces the loss of fresh food storage and transportation and improves the efficiency of cold chain management through multi-dimensional real-time perception, dynamic coupling modeling and hierarchical control.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of cold chain logistics and food preservation, and particularly relates to a fresh food quality deterioration prediction method and system based on multi-modal dynamic coupling BACKGROUND

[0002] Fresh food is susceptible to environmental fluctuations, mechanical damage, and microbial activity during storage and transportation, leading to rapid quality deterioration. The economic loss caused by fresh food spoilage worldwide is as high as tens of billions of dollars per year, accompanied by serious resource waste and carbon emission problems. With the increasing demand for freshness and safety of fresh food, how to achieve early prediction and precise control of quality deterioration has become a core challenge to reduce loss and optimize supply chain efficiency.

[0003] Current fresh food quality monitoring mainly relies on single sensory indicators (such as surface color change) or environmental parameters (such as temperature and humidity threshold alarms), supplemented by manual sampling and experience-based judgment. Some technologies use basic sensors (such as temperature recorders) or near-infrared spectroscopy devices for local detection, but data collection is limited to static or isolated dimensions (such as only focusing on temperature or humidity), lacking systematic integration of dynamic disturbance factors such as loading and unloading shocks, stacking pressure, and transportation vibration.

[0004] The above methods are difficult to capture the coupling effect of multi-modal factors (sensory, environmental, and operation) due to data isolation and model staticity, resulting in delayed deterioration prediction and high misjudgment rate. For example, relying solely on temperature monitoring cannot quantify the cumulative damage of loading and unloading shocks to the microstructure of goods, and manual sampling cannot respond in real time to temperature and humidity fluctuations caused by frequent opening and closing of cold storage doors. This "data island" and "lack of dynamic coupling" problem makes quality management decisions extensive and difficult to achieve early warning and precise intervention, ultimately exacerbating loss rate and operating costs. SUMMARY

[0005] In view of the above actual situation, the present application proposes a fresh food quality deterioration prediction method and system based on multi-modal dynamic coupling to solve the problem of delayed quality deterioration prediction and extensive decision-making due to data isolation and lack of dynamic coupling effect in traditional fresh food storage and transportation processes.

[0006] A fresh food quality deterioration prediction method based on multi-modal dynamic coupling, the method comprising the following steps:

[0007] S1, obtaining to-be-processed data, the to-be-processed data including sensory layer data, environmental layer data and operation layer data, the sensory layer data including fresh produce surface visible light images, near-infrared reflectance spectra and acoustic resonance frequency response curves, the environmental layer data including a micro-environment temperature and humidity gradient matrix in a box, a stacking pressure spatial distribution thermogram and a transportation path vibration spectrum, and the operation layer data including loading and unloading event space-time stamp records and cold storage door opening and closing frequency time series data;

[0008] S2, performing quality deterioration feature extraction and fusion processing on the sensory layer data, so as to obtain a quality deterioration feature tensor;

[0009] S3, performing dynamic coupling modeling processing on the quality deterioration feature tensor and the environmental layer data, so as to obtain a quality deterioration environmental coupling parameter set;

[0010] S4, performing multi-dimensional dynamic constraint field modeling processing on the quality deterioration environmental coupling parameter set and the operation layer data, so as to obtain a deterioration constraint distribution matrix;

[0011] S5, performing adaptive deterioration trajectory evolution processing on the acoustic resonance frequency response curve in the sensory layer data and the deterioration constraint distribution matrix, so as to obtain a four-level quality label.

[0012] Further, the S2 step includes the following sub-steps:

[0013] S201, performing color difference quantization processing on the surface visible light images in the sensory layer data, so as to obtain a surface color difference thermogram, the color difference quantization processing being surface color abnormal area extraction through color space transformation;

[0014] S202, performing chemical bond feature extraction processing on the near-infrared reflectance spectra in the sensory layer data, so as to obtain a water loss feature vector, the chemical bond feature extraction processing being feature band screening related to quality deterioration through a PLS regression model;

[0015] S203, performing spatial chemical fusion processing on the surface color difference thermogram and the water loss feature vector, so as to obtain a quality deterioration feature tensor, the spatial chemical fusion processing being chemical feature mapping to color difference abnormal areas through a spatial weighting matrix.

[0016] Further, the S3 step includes the following sub-steps:

[0017] S301, performing diffusion path topological modeling processing on the quality deterioration feature tensor and the micro-environment temperature and humidity gradient matrix in the environmental layer data, so as to obtain a deterioration diffusion rate field, the diffusion path topological modeling processing being a propagation path of deterioration along a maximum direction of the temperature and humidity gradient simulated through a graph network model;

[0018] S302, mechanically path-coupling the deterioration diffusion rate field and the stack pressure spatial distribution thermodynamic map in the environmental layer data to obtain a pressure path correlation tensor, the mechanically path-coupling being a local deterioration acceleration effect calculated by geometric superposition of the pressure distribution and the diffusion path;

[0019] S303, frequency domain stability analysis processing of the pressure path correlation tensor and the transportation path vibration frequency spectrum in the environmental layer data to obtain a quality deterioration environmental coupling parameter set, the frequency domain stability analysis processing being a disturbance intensity of the vibration on the diffusion path quantified by the frequency band energy distribution.

[0020] Further, the S4 step includes the following sub-steps:

[0021] S401, pulse sequence trajectory tensorization processing of the loading and unloading event space-time stamp records in the operation layer data to obtain a loading and unloading pulse trajectory tensor, the trajectory tensorization processing being a loading energy impact distribution reconstructed by the space-time trajectory of the pulse event;

[0022] S402, temperature and humidity disturbance evolution modeling processing of the cold storage door opening and closing frequency time series data in the operation layer data to obtain a cold chain stability dynamic distribution, the evolution modeling processing being a cold chain disturbance dynamic response model constructed by coupling the time series disturbance curve and the humidity gradient;

[0023] S403, multi-dimensional dynamic constraint mapping processing of the loading and unloading pulse trajectory tensor, the cold chain stability dynamic distribution, and the quality deterioration environmental coupling parameter set to obtain a deterioration constraint distribution matrix, the mapping processing being a constraint balance field formed by spatial-time-mechanical multi-variable weight calculation.

[0024] Further, the S5 step includes the following sub-steps:

[0025] S501, resonance mode coherence analysis processing of the sound wave resonance frequency response curve in the sensory layer data to obtain a frequency spectrum coherence drift factor, the coherence analysis processing being a sound wave mode shift degree calculated by the time series coherence of the resonance frequency band distribution;

[0026] S502, nonlinear deterioration trajectory mapping processing of the frequency spectrum coherence drift factor and the deterioration constraint distribution matrix to obtain a four-level quality label, the trajectory mapping processing being a quality evolution critical point calculated by multi-dimensional deterioration space projection, the four-level quality label including a compliance label, a warning label, an abnormal label, and an irreversible label.

[0027] Further, the mechanical path coupling processing in the S302 step is to calculate the local deterioration acceleration effect including local stress coefficient extraction processing and pressure diffusion gain fusion processing through geometric superposition of pressure distribution and diffusion path; the frequency domain stability analysis processing in the S303 step is to calculate the disturbance intensity of vibration on the diffusion path including vibration energy spectrum extraction processing and frequency band coupling factor calculation processing through frequency band energy distribution quantization.

[0028] Further, the trajectory tensorization processing in the S401 step is to reconstruct energy impact distribution through spatiotemporal trajectory of pulse events including spatiotemporal grid mapping processing and pulse intensity normalization processing.

[0029] Further, the evolution modeling processing in the S402 step is to build a cold chain disturbance dynamic response model through coupling of time sequence disturbance curve and humidity gradient including time sequence disturbance curve decomposition processing and humidity gradient coupling processing; the mapping processing in the S403 step is to form a constraint balance field through space-time-mechanical multivariate weight calculation including space-time dimension alignment processing and multivariate weight fusion processing.

[0030] Further, the coherence analysis processing in the S501 step is to calculate the sound wave mode offset degree through time sequence coherence of resonance frequency band distribution including frequency band segmentation processing and time sequence coherence calculation processing; the trajectory mapping processing in the S502 step is to calculate the quality evolution critical point through multidimensional deterioration space projection including multidimensional feature space construction processing and critical point determination processing.

[0031] In addition, the present application discloses a fresh quality deterioration prediction system based on multi-modal dynamic coupling, characterized in that the system comprises:

[0032] An acquisition unit is configured to acquire to-be-processed data, wherein the to-be-processed data comprises sensory layer data, environmental layer data and operation layer data, the sensory layer data comprises fresh visible light images, near-infrared reflectance spectra and sound wave resonance frequency response curves, the environmental layer data comprises micro-environment temperature and humidity gradient matrices in a box, stacking pressure space distribution thermographs and transportation path vibration frequency spectra, and the operation layer data comprises loading and unloading event spatiotemporal stamp records and cold storage door opening and closing frequency time sequence data.

[0033] A feature fusion unit is configured to perform quality deterioration feature extraction and fusion processing on the sensory layer data, so as to obtain a quality deterioration feature tensor.

[0034] A coupling modeling unit is configured to perform dynamic coupling modeling processing on the quality deterioration feature tensor and the environmental layer data, so as to obtain a quality deterioration environmental coupling parameter set.

[0035] A constraint modeling unit is configured to perform a multi-dimensional dynamic constraint field modeling process on the set of quality deterioration environment coupling parameters and the operation layer data, so as to obtain a deterioration constraint distribution matrix;

[0036] A hierarchical quality label unit is configured to perform an adaptive deterioration trajectory evolution process on the sound wave resonance frequency response curve in the sensory layer data and the deterioration constraint distribution matrix, so as to obtain a four-level quality label.

[0037] The method and system for predicting quality deterioration of fresh food based on multi-modal dynamic coupling provided in the present application realize multi-dimensional real-time sensing, dynamic coupling modeling and hierarchical control of quality deterioration of fresh food, significantly reduce storage and transportation loss, and improve cold chain management efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 A method flowchart of the method for predicting quality deterioration of fresh food based on multi-modal dynamic coupling provided in the present application;

[0039] Figure 2 A process flowchart of generating a quality deterioration feature tensor of the method for predicting quality deterioration of fresh food based on multi-modal dynamic coupling provided in the present application;

[0040] Figure 3 A process flowchart of generating a set of quality deterioration environment coupling parameters of the method for predicting quality deterioration of fresh food based on multi-modal dynamic coupling provided in the present application;

[0041] Figure 4 A process flowchart of generating a deterioration constraint distribution matrix of the method for predicting quality deterioration of fresh food based on multi-modal dynamic coupling provided in the present application;

[0042] Figure 5 A process flowchart of obtaining a four-level quality label of the method for predicting quality deterioration of fresh food based on multi-modal dynamic coupling provided in the present application;

[0043] Figure 6 A system structure schematic diagram of the system for predicting quality deterioration of fresh food based on multi-modal dynamic coupling provided in the present application; DETAILED DESCRIPTION

[0044] The simulation technical route in the embodiments of the present application will be described clearly and completely in combination with the drawings of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the present application.

[0045] In order to make the above objectives, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0046] The features and performance of the present application are further described in detail below in combination with the embodiments. Please refer to the accompanying drawings and specific embodiments Figure 1 As shown in the accompanying drawings, a fresh quality deterioration prediction method based on multi-modal dynamic coupling, the method comprising the following steps:

[0047] S1, obtaining the data to be processed, the data to be processed including sensory layer data, environmental layer data, operation layer data, the sensory layer data including fresh goods surface visible light image, near-infrared reflectance spectrum, sound wave resonance frequency response curve, the environmental layer data including the micro-environment temperature and humidity gradient matrix in the box, the stacking pressure spatial distribution thermograph, the transportation path vibration spectrum, the operation layer data including the space-time stamp record of loading and unloading events, the cold storage door opening and closing frequency time series data;

[0048] In some embodiments, the sensory layer data acquisition involves multiple different modalities of data. Among them, the collection of fresh goods surface visible light image is realized through high-resolution visible light imaging equipment to obtain the color information of the goods surface, which reflects the freshness and color change of the goods surface. The collection of near-infrared reflectance spectrum is realized through near-infrared spectrometer, which emits near-infrared light of a specific wavelength range and measures the intensity change of the reflected light of the fresh goods surface, thereby obtaining the vibration absorption information of the chemical bonds inside the goods, such as water molecule hydrogen bond and protein and fat bond in this embodiment. The collection of sound wave resonance frequency response curve is realized through acoustic excitation-response sensing system, that is, the goods surface is excited by sound waves of a specific frequency range, and the changes of its own resonance response characteristics are detected to reflect the integrity and deterioration degree of its internal organization structure.

[0049] In some embodiments, the collection of environmental layer data includes the acquisition of the micro-environment temperature and humidity gradient matrix inside the box, which reflects the temperature and humidity difference at different spatial positions inside the box and is collected in matrix form by a temperature and humidity sensor array in real time. In addition, the collection of stacking pressure spatial distribution thermograph is realized through a distributed pressure sensor array, which measures and records the pressure spatial distribution information between goods caused by stacking during transportation, and expresses the distribution characteristics of stacking pressure in the goods space in the form of a thermograph. In this embodiment, the collection of transportation path vibration spectrum is completed by a high-sensitivity three-axis vibration sensor, which detects and records the vibration signal frequency distribution caused by road bumps or vehicle movement during transportation, thereby obtaining vibration spectrum data for subsequent evaluation of the disturbance effect of vibration on the goods quality deterioration process.

[0050] In some embodiments, the operation layer data includes loading and unloading event space-time stamp records and cold storage door opening and closing frequency time series data. The loading and unloading event space-time stamp records are realized by a wireless radio frequency identification device and a time stamp recording device installed in the transport container, which automatically captures the time and space information of the time and location of the loading and unloading event, forming clear loading and unloading event space-time stamp data. In some embodiments, the cold storage door opening and closing frequency time series data is collected by a magnetic induction or photoelectric sensor device installed on the cold storage door. The sensor records each opening and closing action of the cold storage door in real time and forms accurate time series data for evaluating the time series response characteristics of the temperature and humidity disturbance caused by the opening and closing of the cold storage door and its impact on the quality deterioration of the goods. In this embodiment, through the above data collection process, the obtained data set is complete and clear, providing an accurate and reliable data basis for subsequent multi-modal fusion and dynamic coupling analysis, thereby laying the foundation for fine prediction of quality deterioration.

[0051] S2, quality deterioration feature extraction and fusion processing is performed on the sensory layer data, thereby obtaining a quality deterioration feature tensor;

[0052] Specifically, please refer to the accompanying Figure 2 The step includes the following sub-steps:

[0053] S201, color difference quantization processing is performed on the surface visible light image in the sensory layer data, thereby obtaining a surface color difference heat map. The color difference quantization processing is performed by color space transformation to extract surface color abnormal regions.

[0054] In this embodiment, the core of the color difference quantization processing is to convert the original visible light image from the RGB color space to the Lab color space by color space transformation method, so as to realize the quantization of color information and the efficient extraction of abnormal regions.

[0055] In this embodiment, after obtaining the original image data, the RGB color model is converted to the Lab color model by using color space transformation with pixels as the basic unit. Specifically, the RGB color vector of any pixel point in the original image is (R, G, B), which is converted to the vector (L, a, b) of the Lab color space by nonlinear transformation. The specific mathematical expression of color space transformation is: X=0.412453R+0.57580G+0.180423B; Y=0.212671R+0.715160G+0.072169B; Z=0.019334R+0.119193G+0.950227B, and the expression of converting the above XYZ space coordinates to Lab space is: L=116f(Y / Y n )-16; a=500[f(X / X n )-f(Y / Y n)] ; b = 200 [f(Y / Y n )-f(Z / Z n )], wherein the reference white point is (X n ,Y n ,Z n ), and the function f(t) is defined as follows: wherein the constant In some embodiments, after the original RGB color space is converted into Lab space by the above transformation, each pixel in the image will be represented by a three-dimensional vector (L, a, b). The advantage of Lab space is that it can more accurately reflect the degree of human eye perception of color difference, thereby facilitating color difference quantification.

[0056] In this embodiment, further quantification calculation is performed on the color difference of Lab space, taking the Euclidean distance between the chroma component of each pixel and the Lab vector of the reference benchmark color as the color difference measure. Specifically, the reference benchmark color (L0, a0, b0) is determined by the color vector of the normal region of the surface of the goods, and the color difference quantification mathematical expression between any pixel color vector (L i ,a i ,b i ) and the benchmark color is: Through the above calculation, each pixel point obtains a corresponding color difference value ΔE i in the color difference quantification process, and a two-dimensional spatial distribution matrix based on the color difference value is formed. In some embodiments, the two-dimensional matrix is normalized and represented by using heat map visualization technology, thereby obtaining a color difference heat map that intuitively expresses the spatial distribution of the color abnormal region of the surface of the fresh goods.

[0057] In some embodiments, the obtained surface color difference heat map is a spatial two-dimensional matrix data, and the numerical value size therein represents the color abnormality degree, i.e., the higher the numerical value, the more obvious the abnormality degree of the region relative to the benchmark color, thereby providing a basis for the spatial-chemical feature fusion in the subsequent steps.

[0058] S202, chemical bond feature extraction processing is performed on the near-infrared reflectance spectrum in the sensory layer data, thereby obtaining a water loss feature vector, and the chemical bond feature extraction processing is performed by a PLS regression model to screen a feature band related to quality deterioration;

[0059] In this embodiment, the near-infrared reflectance spectrum data is measured by a near-infrared spectrometer with a wavelength range of 780-2500 nm, and the data is represented in the form of a one-dimensional numerical array of spectral reflectance intensity and corresponding wavelength. The absorption peak value of each band corresponds to a specific chemical bond vibration absorption signal inside the goods, and the feature band closely related to the O-H chemical bond of water molecules is included.

[0060] In some embodiments, the partial least squares (PLS) regression model is implemented for feature band screening and feature extraction of near-infrared reflectance spectroscopy data. The specific implementation method includes recording the near-infrared spectroscopy original data as a matrix X with a size of m*n, where m represents the number of spectroscopy samples, n represents the number of wavelengths, each row represents the spectroscopy data of a single sample, and each column corresponds to the reflection intensity of a specific wavelength. The quality deterioration index related to the degree of moisture loss is recorded as Y as the model response variable to form the corresponding response variable vector.

[0061] In this embodiment, the partial least squares regression method is used to construct a linear mapping between the spectroscopy data matrix X and the response variable vector Y, and the feature band screening and regression modeling are realized by maximizing the covariance between the characteristic variables and the response variables. The specific mathematical modeling expression of PLS regression is as follows: given the spectroscopy data matrix X ∈ R n×m (where n is the number of samples, and m is the number of spectroscopy bands), the response variable vector Y ∈ R n×1 , the PLS regression model is established to realize feature band selection and correlation coefficient calculation. The model aims to maximize the covariance between the spectroscopy data and the response variable at the same time, extracts the principal component score matrix T and the loading matrix P, and the specific mathematical expression is as follows: X = TP T +E, Y = UQ T +F, where T and U are the extracted principal component score matrices, P and Q are the corresponding loading matrices, and E and F are the residual matrices. Through iterative solution, the principal component matrices T and U satisfy the maximum covariance condition, that is, to solve: where ω and c are weight vectors, which are used to generate the principal component score vectors t = X ω and u = Y c . The weight vectors are determined by iterative optimization to realize dimension reduction of the original data and extract the feature band set with the largest correlation with the quality deterioration response variable.

[0062] In some embodiments, the regression coefficients of each band are analyzed for significance to screen out bands that significantly contribute to the response variable Y, which is recorded as the chemical bond feature band set. After reconstructing the near-infrared spectroscopy feature space using the selected feature bands, PLS regression modeling is performed again to obtain the score vector T * of the quality deterioration moisture loss feature, which is the required chemical bond feature vector and represents the quality change information of the fresh goods caused by moisture loss. In this embodiment, the moisture loss feature vector T * embodies the change characteristics of specific chemical bonds in the fresh goods under the moisture loss state, and is one of the input data required for spatial-chemical feature fusion processing in the subsequent steps.

[0063] S203, performing spatial-chemical fusion processing on the surface color difference heat map and the moisture loss feature vector to obtain a quality deterioration feature tensor, the spatial-chemical fusion processing being performed by mapping the chemical features to the color difference abnormal region through a spatial weighting matrix.

[0064] In the embodiment, the fusion processing is to project the moisture loss feature vector extracted from the near-infrared spectrum data into the spatial abnormal region represented by the surface color difference heat map by using spatial mapping, so as to realize the joint expression of the quality deterioration features in the spatial and chemical information layers.

[0065] In the embodiment, the fusion processing involves constructing a spatial weighting matrix to map the chemical features to the surface color difference abnormal region. In the specific implementation, the surface color difference heat map is defined as a two-dimensional spatial matrix C∈R M×N , where each matrix element C(i,j) represents the color difference abnormal intensity at the position (i,j) in the image; and the moisture loss feature vector obtained by the PLS regression processing is defined as T * ∈R d×1 , where d represents the feature dimension determined by the PLS regression.

[0066] In some embodiments, to realize the spatial-chemical feature mapping, a spatial weighting matrix W∈R M ×N×d is defined, where each element of the matrix represents the mapping strength between a specific spatial position (MxN represents the spatial resolution of the surface color difference heat map) and each chemical feature dimension. In the embodiment, the spatial weighting matrix is constructed by performing normalization processing on each pixel point of the surface color difference heat map to generate a weight factor, and the weight factor is denoted as: where w ij represents the color difference abnormality degree weight factor corresponding to the pixel at the spatial position (i,j), ΔE ij is the color difference quantization value of the pixel (i,j), and the normalized weight factor reflects the contribution proportion of each pixel point relative to the overall surface color difference abnormality degree.

[0067] In some embodiments, based on the above weight factor, the chemical feature vector T * is mapped to the spatial position of the color difference heat map, so as to construct a spatial-chemical fusion feature tensor F. The specific expression of the fusion feature tensor is: where H and W represent the height and width of the surface color difference heat map respectively, represents the value of the kth feature in the moisture loss feature vector T*, and F(i,j,k) represents the fusion value of the kth feature at the position (i,j). Through the fusion process, the generated quality deterioration feature tensor F∈R H×W×dThat is, the precise spatial mapping expression between the chemical bond characteristics and the surface color difference abnormal area is realized.

[0068] In this embodiment, the quality deterioration feature tensor obtained by the fusion processing described above clearly embodies the spatial correlation between the surface color abnormal area and the internal chemical characteristics of moisture loss, providing necessary multi-modal input data for the quality deterioration dynamic coupling modeling of the next stage.

[0069] S3, dynamically coupling the quality deterioration feature tensor and the environmental layer data to obtain a quality deterioration environmental coupling parameter set;

[0070] Specifically, please refer to the accompanying Figure 3 As shown in the figure, this step includes the following sub-steps:

[0071] S301, diffusion path topological modeling processing is performed on the quality deterioration feature tensor and the micro-environmental temperature and humidity gradient matrix in the environmental layer data, thereby obtaining a deterioration diffusion rate field, and the diffusion path topological modeling processing is realized by a graph network model to simulate the propagation path of the deterioration along the maximum direction of the temperature and humidity gradient.

[0072] In some embodiments, the diffusion path topological modeling processing is realized by a graph network model, specifically, the spatial diffusion propagation process of the fresh food quality deterioration feature along the maximum direction of the micro-environmental temperature and humidity gradient is simulated by a graph network method, so as to clearly determine the dynamic propagation trend and diffusion rate of the quality deterioration of the goods.

[0073] In this embodiment, the quality deterioration feature tensor is denoted as a three-dimensional tensor F∈R H×W×d , where H and W represent the height and width of the spatial dimension respectively, and d represents the dimension of the quality deterioration feature. The micro-environmental temperature and humidity gradient matrix is denoted as a two-dimensional matrix G∈R H×W , where each element in the matrix represents the temperature and humidity comprehensive gradient amplitude at the corresponding position in the container, and the amplitude calculation formula is: Where G(i,j) represents the temperature and humidity gradient comprehensive amplitude at position (i,j), T represents the temperature field, H represents the humidity field, , and respectively represent the directional gradient values of the temperature and humidity fields at the spatial position (i,j).

[0074] In some embodiments, in order to determine the diffusion path of the quality deterioration feature along the maximum direction of the temperature and humidity gradient, a graph network model is introduced to topologically model the above-mentioned spatial data. In this topological model, each spatial position (i,j) is regarded as a node of the graph network, and the topological connection between the nodes is determined by the directionality of the temperature and humidity gradient. The connection weight between the adjacent two nodes (i,j) and (p,q) is determined by the temperature and humidity gradient difference, and the weight calculation formula is: Where the weight W(i,j),(p,q) represents the diffusion connection strength between node (i, j) and node (p, q), the greater the difference in temperature and humidity gradient between nodes, the greater the connection weight, and vice versa.

[0075] In some embodiments, the modeling of the deterioration diffusion rate field is based on the graph network structure described above, and further defines the kinetic equation of deterioration diffusion. In this embodiment, the diffusion process is mathematically modeled by a discrete diffusion rate equation, and the specific expression is: where V(i, j) represents the deterioration diffusion rate at spatial position (i, j), N(i, j) represents the set of neighborhood nodes of node (i, j), W (i,j),(p,q) is the connection weight between nodes, and ||F(i, j, :) - F(p, q, :)||2is the Euclidean distance of the quality deterioration feature tensor between node (i, j) and node (p, q), reflecting the difference degree of quality deterioration features between spatial positions.

[0076] In this embodiment, after the deterioration diffusion rate of each spatial position is calculated by the above model, a complete deterioration diffusion rate field matrix V ∈ R H×W is formed. The rate field clearly represents the diffusion path and propagation rate of quality deterioration along the maximum direction of the temperature and humidity gradient, which serves as the basic input information for subsequent steps of mechanical path coupling and frequency domain stability analysis.

[0077] S302, the deterioration diffusion rate field and the stack pressure spatial distribution thermodynamic map in the environmental layer data are coupled to obtain a pressure path correlation tensor, and the mechanical path coupling processing is a local deterioration acceleration effect calculated by geometric superposition of pressure distribution and diffusion path;

[0078] In some embodiments, the mechanical path coupling processing is a local deterioration acceleration effect calculated by geometric superposition of pressure distribution and diffusion path includes local stress coefficient extraction processing and pressure diffusion gain fusion processing.

[0079] In this embodiment, the mechanical path coupling processing is based on the mutual influence relationship between the local pressure distribution generated by the fresh goods under the stacking condition and the quality deterioration diffusion path. By geometric superposition and numerical operation of the two, the acceleration effect of local pressure on deterioration diffusion rate is quantified, and a three-dimensional tensor representation that comprehensively reflects the coupling state of the two is constructed.

[0080] In this embodiment, the deterioration diffusion rate field is obtained from the previous step S301, denoted as matrix V ∈ R H×W , where V(i, j) represents the quality deterioration diffusion rate at position (i, j). The stack pressure spatial distribution thermodynamic map is denoted as matrix P ∈ R H×Wwhere P(i,j) represents the local pressure magnitude at position (i,j). The mechanical path coupling processing is to superimpose the distribution characteristics of the pressure field and the diffusion field to identify the potential acceleration trend of deterioration in the high pressure area, and form a three-dimensional tensor reflecting the coupling relationship between the two.

[0081] In some embodiments, the local stress coefficient extraction processing performs normalization operation on the stacking pressure matrix P to map the original pressure value to a dimensionless local stress coefficient, denoted as matrix P * ∈R H×W The normalized pressure coefficient matrix P * (i,j) is distributed in the interval [0, 1], which reflects the relative intensity of different positions relative to the overall pressure distribution.

[0082] In some embodiments, through the pressure diffusion gain fusion processing, the normalized pressure matrix P * is superimposed with the element corresponding position of the deterioration diffusion rate field V to obtain the mechanical path coupling gain field M∈R H×W . The gain field is used to quantify the acceleration effect of pressure on the diffusion of deterioration, and the expression is: M(i,j) = V(i,j) x (1 + a a P * (i,j)), where a is the coupling gain coefficient, used to control the amplification degree of local pressure on the diffusion rate, M(i,j) represents the deterioration diffusion rate at position ((i,j) after superimposing pressure. The selection of the coefficient a is based on the actual characteristics of fresh food and the compression deformation characteristics, and is determined by experimental data regression or parameter identification algorithm.

[0083] In some embodiments, in order to represent the coupling state of the diffusion path and the pressure distribution in three-dimensional space, the deterioration diffusion rate field V, the normalized pressure matrix P * and the fusion gain field M are jointly constructed as a pressure path correlation tensor R∈R H×W×3 , whose three-dimensional vector representation is as follows: R(i,j,:) = [V(i,j), P * (i,j), M(i,j)], where the first dimension of R(i,j,:) is the original diffusion rate, the second dimension is the local pressure coefficient, and the third dimension is the mechanical coupling gain rate. The degree of interaction between the mechanical pressure and the quality deterioration diffusion at each position is reflected by the joint representation of the three.

[0084] In this embodiment, the obtained pressure path correlation tensor will be used as the input of the frequency domain stability analysis processing in the subsequent steps, to further evaluate the disturbance effect of the transport path vibration on the deterioration diffusion in the high pressure area, and to provide more accurate multi-dimensional data support for the overall quality deterioration environment coupling model.

[0085] S303, frequency domain stability analysis processing is performed on the pressure path correlation tensor and the transport path vibration spectrum in the environmental layer data, so as to obtain a quality deterioration environment coupling parameter set. The frequency domain stability analysis processing quantifies the disturbance intensity of the vibration on the diffusion path through the energy distribution of the frequency band.

[0086] In some embodiments, the frequency domain stability analysis processing quantifies the disturbance intensity of the vibration on the diffusion path through the energy distribution of the frequency band includes vibration energy spectrum extraction processing and frequency band coupling factor calculation processing.

[0087] In this embodiment, the frequency domain stability analysis processing is based on the vibration signal of the transport path collected by the high-sensitivity vibration sensor. The disturbance influence degree of the vibration on the quality deterioration diffusion path is evaluated by quantitatively calculating the energy distribution characteristics of the vibration signal in different frequency bands, so as to determine the frequency band energy coupling relationship related to the deterioration process.

[0088] In this embodiment, the vibration signal of the transport path is denoted as x(t), the sampling frequency is defined as f s , and the recording duration is T. Discrete Fourier transform (DFT) is performed on x(t) to obtain the expression X(ω k ) of the vibration signal in the frequency domain, where ω k represents the k-th discrete angular frequency component, k=0,1,...,K-1, and K represents the number of discrete points. The mathematical expression of DFT is: where Δt=1 / f s , N is the total number of sampling points, The power spectrum distribution of the vibration signal is obtained by performing modulus square operation on X(ω k ): S(ω k ) = |X(ω k )| 2 .

[0089] In some embodiments, the vibration energy spectrum extraction processing includes segment integration or area summation of S(ω k ) in multiple frequency band ranges, for calculating the energy value of the corresponding frequency band and constructing a multi-dimensional frequency band energy vector. The frequency band division interval is defined as {Ω1,Ω2,...,Ω B}, where B represents the number of frequency bands, and each frequency band corresponds to a range of angular frequency, and the energy value of the vibration signal in the frequency band is denoted as

[0090] In this embodiment, the pressure path correlation tensor is generated by step S302 and is denoted as R(i,j,:)=[V(i,j),P * (i,j),M(i,j)], where V(i,j) is the deterioration diffusion rate, P *(i,j) is a local pressure coefficient, and M(i,j) is a mechanical coupling gain rate. To quantify the disturbance effect of vibration on the local deterioration path, a frequency band coupling factor calculation process is introduced, and the segmented energy vector {E1, E2, …, E B} is coupled with the pressure path associated tensor. A set of weight coefficients {β1, β2, …, β B} is defined to represent the disturbance influence degree of each frequency band, and the vibration coupling disturbance factor at position (i,j) is denoted as , the value of β b is based on the actual vibration characteristics and the sensitivity of goods to different frequency bands, and M(i,j) is a mechanical coupling gain rate, representing the local deterioration gain value at position (i,j) formed under the joint action of stacking pressure and diffusion rate.

[0091] In some embodiments, the vibration coupling disturbance factor D(i,j) is calculated point by point for the entire spatial position ((i,j), and the vibration disturbance distribution matrix D ∈ R H×W is generated on a two-dimensional plane. Then, the disturbance distribution matrix is integrated with the existing pressure path associated tensor. Specifically, the vibration disturbance distribution matrix is expanded in dimension to make its dimension consistent with that of the pressure path associated tensor, forming a new three-dimensional matrix D' ∈ R H×W×1 . Then, through tensor connection operation, the expanded matrix D' and the pressure path associated tensor are spliced in the third dimension direction, thereby constructing the quality deterioration environment coupling parameter set tensor E ∈ R H×W×4 . The specific mathematical expression of the above integration operation is: E(i,j,:)=Concat[R(i,j,:),D'(i,j)],i=1,...,H,j=1,...,W, where the result of tensor splicing E ∈ R H×W×4 clearly expresses the joint action relationship of the four-dimensional characteristics of the deterioration diffusion rate, the pressure coefficient, the mechanical coupling gain rate, and the vibration disturbance intensity at the spatial position (i,j). In this embodiment, the disturbance effect of the multi-frequency band vibration energy on the diffusion path determined through the above frequency domain stability analysis process provides quantifiable frequency domain disturbance parameter support for subsequent multi-dimensional dynamic constraint field modeling of the operation layer data, so that the global coupling model of quality deterioration has more perfect completeness and accuracy in the space, mechanics, and vibration levels.

[0092] S4, multi-dimensional dynamic constraint field modeling is performed on the quality deterioration environment coupling parameter set and the operation layer data, so as to obtain a deterioration constraint distribution matrix;

[0093] Specifically, please refer to the following figure: Figure 4 The step includes the following sub-steps:

[0094] S401, pulse sequence trajectory tensorization processing is performed on the loading and unloading event space-time stamp records in the operation layer data, so as to obtain a loading and unloading pulse trajectory tensor, and the trajectory tensorization processing is to reconstruct the loading and unloading energy impact distribution through the space-time trajectory of the pulse event;

[0095] In some embodiments, the trajectory tensorization processing to reconstruct the loading and unloading energy impact distribution through the space-time trajectory of the pulse event includes space-time grid mapping processing and pulse intensity normalization processing.

[0096] In the present embodiment, the space-time grid mapping processing generates a three-dimensional voxel grid by discretizing the space-time domain of the loading and unloading event. The original data form of the loading and unloading event space-time stamp records is a discrete event set where t k is the occurrence time of the kth loading and unloading event, (x k , y k ) is the two-dimensional space coordinate. In the present embodiment, the space-time grid mapping is defined as dividing the continuous time axis into N t equal intervals, and dividing the space plane into N x × N y equal size grid units, forming a three-dimensional space-time voxel grid Each voxel ((i, j, m) corresponds to a time interval [t i , t i+1 ] and a space grid unit (x j , x j+1 ) × (y m , y m+1 ), used to count the frequency and intensity of the loading and unloading events in the space-time unit.

[0097] In the present embodiment, the pulse intensity normalization processing quantifies the energy of the loading and unloading event through a time decay function and a space weight factor. The time decay function φ(t) is defined as an exponential decay form: where λ is the decay coefficient, t min and t max are the minimum and maximum values of the loading and unloading event timestamp respectively. The space weight factor ψ(x, y) is calculated based on the motion dynamics model of the loading and unloading robot arm, and the expression is: where (x c , y c ) is the operation center coordinate of the loading and unloading robot arm, and σ is a space diffusion coefficient representing the propagation range of the loading and unloading energy in space.

[0098] In some embodiments, the space-time trajectory reconstruction of the pulse event maps the discrete event to the space-time voxel grid through convolution operation. For the kth loading and unloading event (t k , x k , y k), whose energy contribution E k (i,j,m) is calculated as: k (i,j,m) = φ(t k -t i ) · ψ(x k -x j ,y k -y m ), where t i is the starting time of the i-th interval of the time grid, and (x j ,y m ) are the center coordinates of the spatial grid cell. By traversing all the loading and unloading events and accumulating their energy contributions, the initial energy distribution matrix whose elements

[0099] In this embodiment, the generation of the loading and unloading pulse trajectory tensor further includes time smoothing processing and spatial interpolation processing. Time smoothing is achieved through a sliding average filter, and the filter window length is W t . The smoothed energy distribution matrix ε * is calculated as: Spatial interpolation uses a bicubic interpolation algorithm to expand the discrete grid energy distribution to a continuous spatial field, and the interpolated tensor where H and W are the spatial resolutions after interpolation.

[0100] In some embodiments, the final tensorization of the loading and unloading energy impact distribution is completed through multi-scale feature fusion. Define the multi-scale convolution kernel where K s is a Gaussian kernel of size s x s, used to extract energy aggregation features at different spatial scales. The fused loading and unloading pulse trajectory tensor is calculated as: where s is the scale parameter, is the Gaussian kernel function, and σ s is proportional to the scale s.

[0101] In this embodiment, through the above spatio-temporal gridding mapping processing and pulse intensity normalization processing, the loading and unloading pulse trajectory tensor completely characterizes the energy impact distribution of the loading and unloading events in the spatio-temporal dimension, and its four-dimensional structure (time, spatial height, spatial width, scale) provides high-resolution input data for subsequent multi-dimensional dynamic constraint field modeling, realizing the quantitative expression of the disturbance effect of loading and unloading operations on the quality deterioration of fresh goods.

[0102] S402, the temperature and humidity disturbance evolution modeling processing is performed on the cold storage door opening and closing frequency time series data in the operation layer data, so as to obtain the cold chain stability dynamic distribution, and the evolution modeling processing is performed by coupling the time series disturbance curve and the humidity gradient to construct a cold chain disturbance dynamic response model.

[0103] In some embodiments, the evolution modeling process is to construct a cold chain disturbance dynamic response model by coupling a time series disturbance curve with a humidity gradient, including a time series disturbance curve decomposition process and a humidity gradient coupling process.

[0104] In the present embodiment, the preprocessing of the cold storage door opening and closing frequency time series data includes event frequency accumulation and time series interpolation processing. The cold storage door opening and closing event raw data is collected by a magnetic induction or photoelectric sensor, and the record form is a discrete timestamp sequence where t n represents the occurrence time of the nth cold storage door opening and closing action. In the present embodiment, the event frequency accumulation is to count the number of opening and closing times per unit time through a sliding time window, define the time window length as ΔT, the time window sliding step as δT, and calculate the cumulative frequency sequence F(t) as: where Ι(·) is an indicator function, taking the value 1 when t n is located in the time window [t-ΔT, t) and 0 otherwise. The time series interpolation processing converts the discrete frequency sequence F(t) into a continuous time function F * (t)∈C 2 (R) by using a cubic spline interpolation algorithm, ensuring the smoothness and differentiability of the time series disturbance curve.

[0105] In the present embodiment, the time series disturbance curve decomposition process extracts the frequency energy distribution characteristics by Fourier transform. Define the cold storage door opening and closing disturbance signal as the interpolated frequency function F * (t), and its continuous Fourier transform is: where ω is the angular frequency, The power spectral density S(ω) is calculated as: S(ω) = |F(ω)| 2 , and by setting the frequency band division threshold {ω1, ω2,... ω B}, the power spectrum is divided into B frequency bands, and the energy integral value E b of each frequency band is: forming a frequency energy vector E = [E1, E2,..., E B ] T , representing the energy distribution characteristics of the cold storage door opening and closing disturbance in different frequency bands.

[0106] In some embodiments, the humidity gradient coupling process is to perform spatio-temporal correlation between the frequency energy and the humidity gradient field by a dynamic response function. The humidity gradient matrix H(x, y, t) in the environmental layer data is collected by a temperature and humidity sensor array, and its spatial gradient field represents the spatial variation rate of humidity in the cargo box. In the present embodiment, the dynamic response function R(x, y, t) is defined as the convolution operation of the humidity gradient field and the time series disturbance energy: where βb is the coupling coefficient of the b-th frequency band, g b (t) is the impulse response function of the corresponding frequency band, expressed as: σ b is the time delay parameter inversely proportional to the center frequency of the frequency band , satisfying κ is a proportional constant. The convolution operation represents the weighted integration of the humidity gradient field in the time dimension with the impulse response function, mathematically expressed as: In this embodiment, the generation of the cold chain stability dynamic distribution is realized by superimposing the normalized dynamic response field and the humidity gradient disturbance. Define the stability distribution matrix S(x, y, t) as: where γ is the attenuation factor, used to adjust the influence strength of the disturbance on the stability, is the Euclidean norm of the humidity gradient field. Through the above calculation, the value range of S(x, y, t) is [0, 1], and the closer the numerical value is to 1, the higher the cold chain stability at that position at time t, and vice versa, indicating that the disturbance has a significant impact.

[0107] In some embodiments, the spatio-temporal discretization process maps the continuous stability distribution S(x, y, t) to a discrete grid. Define the time resolution as Δt and the spatial resolution as Δx × Δy. The discretized stability dynamic distribution matrix S ∈ R T ×X×Y is calculated as S(i, j, m) = S(s j , y m , t i ), where t i =iΔt,x j =jΔx,y m =mΔy,i=1,2,...,T,j=1,2,...,X,m=1,2,...,Y. In this embodiment, through the above time series disturbance curve decomposition process and humidity gradient coupling process, the cold chain stability dynamic distribution matrix S completely characterizes the disturbance effect of the opening and closing operation of the cold storage door on the humidity field in the container and its spatio-temporal evolution law, providing a quantitative stability index for subsequent multi-dimensional dynamic constraint field modeling, supporting global coupling analysis of fresh food quality deterioration prediction model.

[0108] S403, the loading and unloading pulse trajectory tensor, the cold chain stability dynamic distribution, and the quality deterioration environment coupling parameter set are subjected to multi-dimensional dynamic constraint mapping processing, thereby obtaining a deterioration constraint distribution matrix. The mapping processing is to form a constraint balance field through spatial-time-mechanical multi-variable weight calculation.

[0109] In some embodiments, the mapping processing to form a constraint balance field through spatial-time-mechanical multi-variable weight calculation includes spatial-time dimension alignment processing and multi-variable weight fusion processing.

[0110] In this embodiment, the space-time dimension alignment process unifies the space-time resolution of the input data through interpolation operation. H×W×4 The spatial dimension of is H×W, and the time dimension is implicit in the dynamic evolution of the environmental parameters; the loading and unloading pulse trajectory tensor T * ∈R Nt×H×W×S The time dimension is N t , the spatial dimension is H×W; the cold chain stability dynamic distribution matrix S∈R T×X×Y The time dimension of is T, and the space dimension is X×Y. In this embodiment, the spatial resolution of S is adjusted from X×Y to H×W by bilinear interpolation algorithm, and the time dimension is adjusted from N to N by linear interpolation. t Align with T to the common time axis length L to generate the aligned stability distribution tensor S * ∈R L×H×W .

[0111] In this embodiment, the multivariate weight fusion process realizes the construction of the constraint balance field by defining the spatial weight, time weight and mechanical weight. space ∈RH×W is calculated based on the local pressure coefficient and vibration disturbance intensity in the environmental coupling parameter set E, and the expression is: Where E(i,j,2) is the local pressure coefficient at position (i,j), E(i,j,4) is the vibration disturbance intensity, ⊙ represents element-by-element multiplication, and the denominator is the global maximum normalization factor. Time weight W time ∈RL based on the loading and unloading pulse trajectory tensor T * The time energy accumulation value is calculated as follows: Mechanical weight W mech ∈R H×W Based on the calculation of the degradation diffusion rate and mechanical coupling gain rate in the environmental coupling parameter set E, the expression is: Where α is the mechanical gain coefficient, E(i,j,1) is the degradation diffusion rate, and E(i,j,3) is the mechanical coupling gain rate. In some embodiments, the generation of the constraint balance field is achieved by a tensor product operation of space-time-mechanical weights. Define the constraint balance field C∈R L×H×W = C(l,i,j) = W space (i,j)·W time (l)·W mech (i,j)·S * (l,i,j), where S * (l,i,j) represents the aligned cold chain stability distribution. By element-by-element multiplication, the constrained equilibrium field comprehensively characterizes the multidimensional coupling effects of spatial pressure-vibration disturbance, temporal loading and unloading energy impact, mechanical diffusion gain, and cold chain stability.

[0112] In this embodiment, the generation of the degradation constraint distribution matrix is ​​completed by the time-space integration and normalization of the constraint balance field. Define the degradation constraint distribution matrix D∈R H×W for: Where η is the attenuation coefficient, which is used to suppress the oversaturation effect in areas with high cumulative constraints. Normalization ensures that the matrix D ranges from [0 to 1]. Values ​​closer to 1 indicate a higher degree of suppression of quality degradation at that spatial location due to the multidimensional constraint effect, while values ​​lower than 1 indicate a weaker constraint effect.

[0113] In some embodiments, post-processing optimization smoothes local mutations in the constraint distribution matrix by anisotropic diffusion filtering. The filtering equation uses the Perona-Malik model: Where τ is the pseudo-time parameter, is the edge preservation diffusion function, K is the gradient threshold. The finite difference method is used to iteratively solve until convergence, and the smoothed degradation constraint distribution matrix D is obtained. * ∈R H×W .

[0114] In this embodiment, through the above-mentioned space-time dimension alignment processing and multivariate weight fusion processing, the degradation constraint distribution matrix D * The synergistic constraints of loading and unloading operations, cold chain stability, and environmental coupling parameters on the quality degradation of fresh goods are fully quantified, providing spatially resolved constraint strength input for the final adaptive degradation trajectory evolution processing and supporting the precise division of four-level quality labels.

[0115] S5, adaptive degradation trajectory evolution processing is performed on the acoustic resonance frequency response curve and degradation constraint distribution matrix in the sensory layer data to obtain a four-level quality label;

[0116] For details, please refer to the attached Figure 5 As shown, this step includes the following sub-steps:

[0117] S501, performing a resonance modal coherence analysis on the acoustic wave resonance frequency response curve in the sensory layer data to obtain a spectrum coherence drift factor, wherein the coherence analysis is to calculate the degree of acoustic wave mode shift through the temporal coherence of the resonance frequency band distribution;

[0118] In some embodiments, the coherence analysis process is to calculate the degree of acoustic wave mode shift through the temporal coherence of the resonance frequency band distribution, which includes a frequency band segmentation process and a temporal coherence calculation process.

[0119] In the present embodiment, the frequency band segmentation process extracts multi-scale frequency band features of the acoustic resonance frequency response curve via continuous wavelet transform. The acoustic resonance frequency response curve is collected by an acoustic excitation-response sensing system, and the original data is in the form of time series where f(t) denotes the resonance frequency amplitude at time point t. In the present embodiment, the continuous wavelet transform is defined as: where a is the scale parameter, b is the translation parameter, ψ is the Morlet wavelet basis function, and the expression is: ω0is the wavelet center frequency. By setting a scale parameter set extracting the resonance frequency band related to the internal structure of the fresh produce The time-varying energy distribution E of each frequency band m (t) is calculated as: m (t) = |W(a m ,t)| 2 In the present embodiment, the time series coherence calculation process quantifies the phase consistency between frequency bands through sliding window cross-correlation analysis. Define the time window length as W and the sliding step size as δ, the coherence coefficient ρ mn (k) between the mth frequency band and the nth frequency band in the window is: where is the energy mean of the mth frequency band in the window, and k is the starting time point of the window. By traversing all time window and frequency band pairs, the full-band coherence tensor P ∈ R K×M×M is generated, where K is the total number of windows P(k,m,n) = ρ mn (k). In some embodiments, the quantification of acoustic mode shift degree is realized through principal component analysis dimension reduction and drift factor calculation. For each time window, the coherence matrix P(k,:,:) is subjected to eigenvalue decomposition, and the first P principal component vectors are extracted. The time-varying principal component trajectory matrix V ∈ R K ×P×M is constructed. The spectral coherence drift factor Δ(k) is defined as the Euclidean distance between the principal component trajectories of adjacent windows: In the present embodiment, the drift factor normalization process realizes global comparability through exponential smoothing and range scaling. The normalized drift factor Δ * (k) is calculated as: where ξ is the attenuation coefficient, used to suppress the cumulative drift noise at the end of the time series.

[0120] In some embodiments, post-processing optimization smooths the time series mutations of the drift factor through Kalman filtering. The state equation and observation equation are defined as: x(k) = x(k-1) + ω(k), z(k) = x(k) + v(k), where x(k) is the hidden state, z(k) = Δ *(k) are process noise and observation noise. By recursive prediction and update steps, a smoothed spectral coherence drift factor sequence Δ ** (k) is obtained.

[0121] In this embodiment, through the above frequency band segmentation processing and timing coherence calculation processing, the spectral coherence drift factor Δ ** (k) accurately quantifies the degree of shift of the acoustic resonance mode in the time dimension, and the larger the value indicates that the acoustic mode instability effect caused by the deterioration of the internal structure of the fresh produce is more significant, providing a frequency domain dynamic feature input for subsequent nonlinear deterioration trajectory mapping processing.

[0122] S502, the spectral coherence drift factor and the deterioration constraint distribution matrix are subjected to nonlinear deterioration trajectory mapping processing, thereby obtaining a four-level quality label, the trajectory mapping processing is to calculate the quality evolution critical point by multi-dimensional deterioration space projection, and the four-level quality label includes a compliance label, a warning label, an abnormal label, and an irreversible label.

[0123] In some embodiments, the trajectory mapping processing is to calculate the quality evolution critical point by multi-dimensional deterioration space projection includes multi-dimensional feature space construction processing and critical point determination processing.

[0124] In this embodiment, the multi-dimensional feature space construction processing fuses the spectral coherence drift factor and the deterioration constraint distribution matrix through tensor splicing and normalization operation. The spectral coherence drift factor sequence Δ ** (k) ∈ R K represents the degree of acoustic mode shift in the time dimension, and the deterioration constraint distribution matrix D * ∈ R H×W characterizes the constraint strength in the spatial dimension. In this embodiment, the space-time alignment processing extends Δ ** (k) to the spatial dimension by interpolation to generate the drift factor distribution matrix Δ map ∈ R H×W , the elements of which are where δ ij (k) is a space-time correlation factor, reflecting the deterioration constraint weight of position (i, j) at time k. The fused multi-dimensional feature tensor F ∈ R H×W×2 is defined as: F(i,j,1) = D * (i,j), F(i,j,2) = Δ map (i,j), and the normalization processing ensures that F(i,j,d) ∈ [0,1], d = 1,2.

[0125] In this embodiment, the critical point determination processing identifies the quality evolution boundary through hyperplane segmentation and density clustering. Define a two-dimensional deterioration space D = [0,1]2 , horizontal axis is constraint strength D * , vertical axis is drift strength Δ map . The hyperplane partition function h(D * , Δ map ) = θ1D * + θ2Δ map - θ0divides the space into four types of regions, where θ0, θ1, θ2 are partition parameters determined by fitting historical degradation data. Density clustering further optimizes the boundary based on the kernel density estimation function: Identify the center of the high-density region as the critical point Divide the four-class label region. In some embodiments, the generation logic of the four-level quality label is as follows:

[0126] Compliance label: D * ≥ τ1 and Δ map ≤ γ1, indicating that the constraint strength is high and the acoustic mode is stable, and the quality of the goods meets the storage standard;

[0127] Warning label: D * ∈ [τ2, τ1) and Δ map ∈ (γ1, γ2], indicating that the constraint strength decreases or the acoustic mode slightly deviates, triggering a quality degradation warning;

[0128] Abnormal label: D * ∈ [τ3, τ2) and Δ map ∈ (γ2, γ3], indicating that the constraint effect is significantly weakened or the acoustic mode is moderately unstable, and the quality of the goods enters an abnormal state;

[0129] Irreversible label: D * < τ3 and Δ map > γ3, indicating that the constraint effect is invalid and the acoustic mode is severely deviated, and the quality of the goods is irreversible.

[0130] In this embodiment, the threshold parameters {τ1, τ2, τ3} and {γ1, γ2, γ3} are determined by historical data statistics and ROC curve analysis to ensure that the classification results are consistent with the true degradation phase. The label mapping function L(i, j) is defined as: In some embodiments, post-processing verification evaluates the label classification accuracy through the confusion matrix and Kappa coefficient. Define the real label set and the predicted label set Calculate the precision, recall and overall Kappa consistency coefficient of each class to ensure that the model output meets the actual quality detection results.

[0131] In the embodiment, through the multi-dimensional feature space construction processing and the critical point determination processing, the four-level quality label L(i,j) accurately quantifies the quality state of the fresh food at different spatial positions, providing a grading decision basis for cold chain management: the compliance label guides normal storage, the early warning label triggers inspection intervention, the abnormal label starts quality re-inspection, and the irreversible label executes product rejection, thereby realizing whole-cycle management and control of deterioration risk.

[0132] Based on the description of the above-mentioned fresh food quality deterioration prediction method based on multi-modal dynamic coupling, the embodiments of the present application also disclose a fresh food quality deterioration prediction system based on multi-modal dynamic coupling. The fresh food quality deterioration prediction system based on multi-modal dynamic coupling can be a computer program (including program code) running the above-mentioned fresh food quality deterioration prediction method based on multi-modal dynamic coupling. Please refer to FIG. 10, the fresh food quality deterioration prediction system based on multi-modal dynamic coupling can run the following units: Figure 6

[0133] The acquisition unit 110 is configured to acquire the to-be-processed data, wherein the to-be-processed data includes sensory layer data, environmental layer data and operation layer data, the sensory layer data includes visible light images of a fresh food surface, near-infrared reflectance spectra and sound wave resonance frequency response curves, the environmental layer data includes a micro-environment temperature and humidity gradient matrix in a box, a stacking pressure spatial distribution thermograph and a transportation path vibration spectrum, and the operation layer data includes loading and unloading event space-time stamp records and cold storage door opening and closing frequency time series data.

[0134] The feature fusion unit 120 is configured to perform quality deterioration feature extraction and fusion processing on the sensory layer data, so as to obtain a quality deterioration feature tensor.

[0135] The coupling modeling unit 130 is configured to perform dynamic coupling modeling processing on the quality deterioration feature tensor and the environmental layer data, so as to obtain a quality deterioration environmental coupling parameter set.

[0136] The constraint modeling unit 140 is configured to perform multi-dimensional dynamic constraint field modeling processing on the quality deterioration environmental coupling parameter set and the operation layer data, so as to obtain a deterioration constraint distribution matrix.

[0137] The grading quality label unit 150 is configured to perform adaptive deterioration trajectory evolution processing on the sound wave resonance frequency response curve in the sensory layer data and the deterioration constraint distribution matrix, so as to obtain a four-level quality label.

[0138] ​The foregoing is considered as illustrative only of the principles of the application. Further, since numerous modifications and changes will readily occur to those skilled in the art, it is not desired to limit the application to the exact construction and operation described. Accordingly, all such variations are intended to be included within the scope of the present application as defined in the claims below and their equivalents.

Claims

1. A method for predicting the quality deterioration of fresh produce based on multi-modal dynamic coupling, characterized in that, The method comprises the following steps: S1, obtaining to-be-processed data, the to-be-processed data comprising sensory layer data, environmental layer data, and operation layer data, the sensory layer data comprising fresh produce surface visible light images, near-infrared reflectance spectra, and sound wave resonance frequency response curves, the environmental layer data comprising a micro-environment temperature and humidity gradient matrix in a box, a stacked pressure spatial distribution thermograph, and a transportation path vibration spectrum, and the operation layer data comprising loading and unloading event space-time stamp records and cold storage door opening and closing frequency time series data; S2, performing quality deterioration feature extraction and fusion processing on the sensory layer data, thereby obtaining a quality deterioration feature tensor; S3, performing dynamic coupling modeling processing on the quality deterioration feature tensor and the environmental layer data, thereby obtaining a quality deterioration environmental coupling parameter set; S4, performing multi-dimensional dynamic constraint field modeling processing on the quality deterioration environmental coupling parameter set and the operation layer data, thereby obtaining a deterioration constraint distribution matrix; S5, performing adaptive deterioration trajectory evolution processing on the sound wave resonance frequency response curves in the sensory layer data and the deterioration constraint distribution matrix, thereby obtaining a four-level quality label; The S4 step comprises the following sub-steps: S401, performing pulse sequence trajectory tensorization processing on the loading and unloading event space-time stamp records in the operation layer data, thereby obtaining a loading and unloading pulse trajectory tensor, wherein the trajectory tensorization processing is performed by reconstructing a loading and unloading energy impact distribution through a space-time trajectory of a pulse event; S402, performing temperature and humidity disturbance evolution modeling processing on the cold storage door opening and closing frequency time series data in the operation layer data, thereby obtaining a cold chain stability dynamic distribution, wherein the evolution modeling processing is performed by coupling a time series disturbance curve and a humidity gradient to construct a cold chain disturbance dynamic response model; S403, performing multi-dimensional dynamic constraint mapping processing on the loading and unloading pulse trajectory tensor, the cold chain stability dynamic distribution, and the quality deterioration environmental coupling parameter set, thereby obtaining a deterioration constraint distribution matrix, wherein the mapping processing is performed by calculating a space-time-mechanical multi-variable weight to form a constraint balance field; The S5 step comprises the following sub-steps: S501, performing resonance mode coherence analysis processing on the sound wave resonance frequency response curves in the sensory layer data, thereby obtaining a frequency spectrum coherence drift factor, wherein the coherence analysis processing is performed by calculating a sound wave mode shift degree through a time series coherence of a resonance frequency band distribution; S502, performing nonlinear deterioration trajectory mapping processing on the frequency spectrum coherence drift factor and the deterioration constraint distribution matrix, thereby obtaining a four-level quality label, wherein the trajectory mapping processing is performed by calculating a quality evolution critical point through multi-dimensional deterioration space projection, and the four-level quality label comprises a compliance label, a warning label, an abnormal label, and an irreversible label.

2. The method of claim 1, wherein the method is based on multi-modal dynamic coupling for predicting the quality deterioration of fresh produce. The S2 step comprises the following sub-steps: S201, performing color difference quantization processing on the surface visible light images in the sensory layer data, thereby obtaining a surface color difference thermograph, wherein the color difference quantization processing is performed by extracting surface color abnormal areas through a color space transformation; S202, performing chemical bond feature extraction processing on the near-infrared reflectance spectrum in the sensory layer data to obtain a moisture loss feature vector, the chemical bond feature extraction processing being performed by a PLS regression model to screen feature bands related to quality deterioration; S203, performing spatial-chemical fusion processing on the surface color difference heat map and the moisture loss feature vector to obtain a quality deterioration feature tensor, the spatial-chemical fusion processing being performed by a spatial weighting matrix to map chemical features to color difference abnormal regions.

3. The method of claim 1, wherein the method is based on multi-modal dynamic coupling for predicting the quality deterioration of fresh produce. The S3 step includes the following sub-steps: S301, performing diffusion path topological modeling processing on the quality deterioration feature tensor and the micro-environment temperature and humidity gradient matrix in the environmental layer data to obtain a deterioration diffusion rate field, the diffusion path topological modeling processing being performed by a graph network model to simulate the propagation path of deterioration along the maximum direction of the temperature and humidity gradient; S302, performing mechanical path coupling processing on the deterioration diffusion rate field and the stacking pressure spatial distribution thermograph in the environmental layer data to obtain a pressure path correlation tensor, the mechanical path coupling processing being performed by geometric superposition of pressure distribution and diffusion path to calculate the local deterioration acceleration effect; S303, performing frequency domain stability analysis processing on the pressure path correlation tensor and the transportation path vibration frequency spectrum in the environmental layer data to obtain a set of quality deterioration environmental coupling parameters, the frequency domain stability analysis processing being performed by quantifying the disturbance intensity of vibration on the diffusion path through frequency band energy distribution.

4. The method of claim 3, wherein the method is based on multi-modal dynamic coupling for predicting the quality deterioration of fresh produce. The mechanical path coupling processing in the S302 step includes local stress coefficient extraction processing and pressure diffusion gain fusion processing; the frequency domain stability analysis processing in the S303 step includes vibration energy spectrum extraction processing and frequency band coupling factor calculation processing.

5. The method of claim 1, wherein the method is based on multi-modal dynamic coupling for predicting the quality deterioration of fresh produce. The trajectory tensorization processing in the S401 step includes time-space trajectory reconstruction of pulse events to reconstruct energy impact distribution, including time-space grid mapping processing and pulse intensity normalization processing.

6. The method of claim 1, wherein the method is based on multi-modal dynamic coupling for predicting the quality deterioration of fresh produce. The evolution modeling processing in the S402 step includes constructing a cold chain disturbance dynamic response model by coupling a time series disturbance curve with a humidity gradient, including time series disturbance curve decomposition processing and humidity gradient coupling processing; the mapping processing in the 403 step includes forming a constraint balance field by spatial-time-mechanical multivariate weight calculation, including spatial-time dimension alignment processing and multivariate weight fusion processing.

7. The method of claim 1, wherein the method is based on multi-modal dynamic coupling for predicting the quality deterioration of fresh produce. The coherence analysis processing in the S501 step includes calculating the sound wave mode shift degree by time series coherence of resonance frequency band distribution, including frequency band segmentation processing and time series coherence calculation processing; the trajectory mapping processing in the S502 step includes calculating the quality evolution critical point by multidimensional deterioration space projection, including multidimensional feature space construction processing and critical point determination processing.

8. A system for predicting the quality deterioration of fresh produce based on multi-modal dynamic coupling, characterized in that, The system includes: An acquisition unit is configured to acquire to-be-processed data, the to-be-processed data including sensory layer data, environmental layer data and operation layer data, the sensory layer data including fresh produce surface visible light images, near-infrared reflectance spectra and sound wave resonance frequency response curves, the environmental layer data including a micro-environment temperature and humidity gradient matrix in a box, a stacked pressure spatial distribution thermograph and a transportation path vibration spectrum, and the operation layer data including loading and unloading event space-time stamp records and cold storage door opening and closing frequency time series data; A feature fusion unit is configured to perform quality deterioration feature extraction and fusion processing on the sensory layer data, so as to obtain a quality deterioration feature tensor; A coupling modeling unit is configured to perform dynamic coupling modeling processing on the quality deterioration feature tensor and the environmental layer data, so as to obtain a quality deterioration environmental coupling parameter set; A constraint modeling unit is configured to perform multi-dimensional dynamic constraint field modeling processing on the quality deterioration environmental coupling parameter set and the operation layer data, so as to obtain a deterioration constraint distribution matrix; A hierarchical quality label unit is configured to perform adaptive deterioration trajectory evolution processing on the sound wave resonance frequency response curves in the sensory layer data and the deterioration constraint distribution matrix, so as to obtain a four-level quality label, wherein the obtaining of the deterioration constraint distribution matrix includes the following steps: S401. Performing pulse sequence trajectory tensorization processing on the loading and unloading event space-time stamp records in the operation layer data, so as to obtain a loading and unloading pulse trajectory tensor, wherein the trajectory tensorization processing is performed by reconstructing a loading and unloading energy impact distribution through a space-time trajectory of a pulse event; S402. Performing temperature and humidity disturbance evolution modeling processing on the cold storage door opening and closing frequency time series data in the operation layer data, so as to obtain a cold chain stability dynamic distribution, wherein the evolution modeling processing is performed by constructing a cold chain disturbance dynamic response model through coupling of a time series disturbance curve and a humidity gradient; S403. Performing multi-dimensional dynamic constraint mapping processing on the loading and unloading pulse trajectory tensor, the cold chain stability dynamic distribution and the quality deterioration environmental coupling parameter set, so as to obtain the deterioration constraint distribution matrix, wherein the mapping processing is performed by forming a constraint balance field through space-time-mechanical multi-variable weight calculation; the obtaining of the four-level quality label includes the following steps: S501. Performing resonance mode coherence analysis processing on the sound wave resonance frequency response curves in the sensory layer data, so as to obtain a frequency spectrum coherence drift factor, wherein the coherence analysis processing is performed by calculating a sound wave mode deviation degree through time series coherence of a resonance frequency band distribution; S502. Performing nonlinear deterioration trajectory mapping processing on the frequency spectrum coherence drift factor and the deterioration constraint distribution matrix, so as to obtain the four-level quality label, wherein the trajectory mapping processing is performed by calculating a quality evolution critical point through multi-dimensional deterioration space projection, and the four-level quality label includes a compliance label, a warning label, an abnormal label and an irreversible label.

Images