Multi-index meat freshness nondestructive testing method and system

Abstract

The invention relates to the technical field of food nondestructive testing, and discloses a multi-index meat freshness nondestructive testing method and system. According to the method, spectrum, image and gas sensing data of meat are collected and are separated into independent signal flows through signal decoupling. Performing targeted treatment: performing noise stripping and waveform locking on the spectrum to generate a characteristic spectrum; performing hierarchical analysis and morphological quantization on the image to generate a feature set; performing time sequence decomposition and gradient modeling on the odor to generate a feature sequence; and inputting the three types of features into a freshness collaborative decision network, performing cyclic comparison and evidence chain fusion through an interactive verification channel in the network, and outputting a freshness state report of a multi-index consistency conclusion. According to the method, deep collaborative analysis of multi-source heterogeneous information is realized, and the accuracy and stability of a detection result are improved.

Description

Technical Field

[0001] This invention relates to the field of non-destructive testing technology for food, specifically a method and system for non-destructive testing of the freshness of meat with multiple indicators. Background Technology

[0002] Meat freshness is a key indicator determining its food safety and commercial value. Current non-destructive testing technologies mostly rely on measuring single or a few indicators, such as using near-infrared spectroscopy to analyze changes in specific chemical components, using machine vision to assess surface color and texture, or using electronic nose sensor arrays to respond to volatile gases. These methods obtain information from a single physical or chemical dimension.

[0003] Existing technical solutions have shortcomings. After acquiring spectral, image, and gas data separately using multiple independent detection devices, simple data stacking or rudimentary algorithms are typically used for fusion analysis. Signals from different sources inherently differ in data scale, physical meaning, and noise patterns. Direct mixing and processing can easily lead to mutual interference and masking of features, and the extracted features lack clear biochemical or physical state orientation. At the decision-making level, a common practice is to assign weights to the analysis results of each indicator and then perform linear weighting or let the classifier make a direct decision. This one-way aggregation mode cannot verify the logical consistency between the conclusions of different indicators. When a sensor is interfered with or local samples exhibit peculiarities, the overall judgment is prone to inaccuracy.

[0004] A method is needed to effectively coordinate multi-source heterogeneous signals. It is necessary to effectively separate and selectively process the mixed raw signals to extract in-depth features that directly reflect different aspects of meat spoilage mechanisms. A decision-making mechanism is needed to enable these features to cross-verify and validate each other during the judgment process, thereby arriving at a logically consistent and more reliable freshness conclusion. Summary of the Invention

[0005] The purpose of this invention is to provide a non-destructive testing method and system for multiple indicators of meat freshness, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a non-destructive testing method for multi-index meat freshness, the method comprising: The raw detection signal set of the target meat is collected in the detection environment. The raw detection signal set includes continuous spectral scanning data, microscopic surface imaging data and gas diffusion sensing data. The original set of detected signals is converted into an initial data matrix of uniform dimension, and multi-dimensional signal decoupling processing is performed on the initial data matrix to separate independent spectral signal streams, image signal streams, and odor signal streams. Adaptive noise stripping and feature waveform locking are performed on the spectral signal stream to generate a clean spectral feature map; structural layering analysis and micro-region morphology quantization are performed on the image signal stream to generate a structured image feature set; time series decomposition and concentration gradient modeling are performed on the odor signal stream to generate a dynamic odor feature sequence. The pure spectral feature map, the structured image feature set, and the dynamic odor feature sequence are synchronously input into the freshness collaborative decision network. Through the interactive verification channel within the freshness collaborative decision network, cyclic comparison and evidence chain fusion are performed to generate a freshness status report containing consistent conclusions of multiple indicators.

[0007] Preferably, the set of raw detection signals collected from the target meat in the detection environment includes: The multispectral imaging unit is controlled to perform a gridded scan of the surface area of ​​the target meat to obtain reflection intensity data covering different wavelengths, forming the continuous spectral scan data; The high-definition microscopic imaging unit is controlled to capture the microscopic surface of the target meat frame by frame along a preset path, and an image sequence containing fiber texture and moisture distribution information is obtained to form the microscopic surface imaging data. The array-type gas sensor is controlled to adsorb and analyze volatile molecules released by the target meat in real time within a closed detection chamber, and the electrochemical response signal that changes over time is obtained to form the gas diffusion sensing data. The reflection intensity data, the image sequence, and the electrochemical response signal are aligned and packaged according to the acquisition timestamp to construct the original detection signal set.

[0008] Preferably, the step of converting the original detection signal set into an initial data matrix of uniform dimension includes: The continuous spectral scan data is subjected to wavelength calibration and energy normalization to eliminate differences between devices and generate a standard spectral vector. The microscopic surface imaging data is subjected to scale unification and brightness equalization processing, and the grayscale statistical histogram of each imaging frame is extracted to generate a standard image vector. The gas diffusion sensing data is subjected to baseline correction and drift compensation processing, the characteristic response peak value per unit time is calculated, and a standard odor vector is generated. The standard spectral vector, the standard image vector, and the standard odor vector are concatenated and expanded in dimension according to the same sample index, and missing values ​​are filled to generate the initial data matrix with the same dimension.

[0009] Preferably, the step of performing multi-dimensional signal decoupling processing on the initial data matrix includes: The initial data matrix is ​​input to the signal decoupling processing unit, which has an independent component analysis model built in it. The statistical independence between different signal sources in the initial data matrix is ​​calculated iteratively using the independent component analysis model until the preset decoupling convergence condition is met. Based on the iterative calculation results, statistically independent component signals are separated and classified into spectral signal sources, image signal sources and odor signal sources respectively. The component signals classified to the spectral signal source are reconstructed into the spectral signal stream, the component signals classified to the image signal source are reconstructed into the image signal stream, and the component signals classified to the odor signal source are reconstructed into the odor signal stream.

[0010] Preferably, the step of performing adaptive noise stripping and feature waveform locking processing on the spectral signal stream to generate a clean spectral feature map includes: The local frequency characteristics of the spectral signal stream are analyzed using a sliding window to dynamically identify and filter out noise bands whose frequency characteristics do not conform to the preset meat spectral characteristics. In the spectral signal after noise is filtered out, search for local waveform segments that match the preset typical freshness feature waveform library with a matching degree exceeding a threshold. All local waveform segments found are time-axis aligned and amplitude normalized. The processed waveform segments are spliced ​​and smoothly connected according to their original wavelength order to finally generate the pure spectral feature map.

[0011] Preferably, the step of synchronously inputting the pure spectral feature map, the structured image feature set, and the dynamic odor feature sequence into the freshness collaborative decision network includes: In the freshness collaborative decision-making network, three independent feature input channels are set up to receive the purity spectral feature map, the structured image feature set, and the dynamic odor feature sequence, respectively; Each feature input channel is followed by a feature deep abstraction subnetwork, which is used to convert the input feature data into a high-dimensional abstract feature vector; An interactive verification channel is set up to connect the outputs of the three feature deep abstract sub-networks. The interactive verification channel is used to calculate the mutual information between any two high-dimensional abstract feature vectors and dynamically adjust the fusion weights of the three feature vectors based on the mutual information. A weighted fusion algorithm is used to merge three high-dimensional abstract feature vectors into a comprehensive feature representation based on the fusion weights.

[0012] Preferably, the step of performing cyclical comparison and evidence chain fusion through the interactive verification channel within the freshness collaborative decision-making network to generate a freshness status report containing consensus conclusions across multiple indicators includes: In the interactive verification channel, the comprehensive feature representation generated in the previous round of fusion is fed back to each feature deep abstraction subnetwork as prior knowledge to participate in the feature abstraction process in the next round. Perform multiple rounds of abstraction and fusion loops until the entropy value of the comprehensive feature representation changes below a stable threshold; The stable comprehensive feature representation obtained in the final round is input into the decision output layer; The decision output layer parses the freshness level code, the evaluation score of each individual indicator, and the consistency check flag between indicators based on the stable comprehensive feature representation, and packages them into the freshness status report.

[0013] Preferably, the method further includes: Based on the deviation information of the indicators in the freshness status report, the strategy mapping engine is activated. The strategy mapping engine outputs a set of adjustment parameters that match the current freshness status, specifically: The freshness status report is parsed to extract the freshness level code, the evaluation score of each individual indicator, and the consistency check flag between the indicators; The difference between the evaluation score of each individual indicator and the standard freshness benchmark value is calculated to obtain the deviation information of the indicator; The freshness level code, the indicator deviation information, and the consistency check flag are used together as query conditions and input into the strategy mapping engine. The strategy mapping engine performs matching queries in a preset multidimensional strategy lookup table, which records the mapping relationship between different combinations of query conditions and sets of control parameters. Output the set of control parameters obtained from the query.

[0014] Preferably, after the strategy mapping engine outputs a set of control parameters that match the current freshness state, it further includes: The set of control parameters is formatted into a sequence of control instructions that can be recognized by the target storage device; Establish a communication link with the target storage device, and send the control command sequence to the built-in controller of the target storage device through the communication link; The built-in controller executes the control command sequence to adjust the temperature, humidity, and gas composition parameters in the storage environment.

[0015] Preferably, the present invention also includes a multi-index meat freshness non-destructive testing system, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, it implements the steps of the multi-index meat freshness non-destructive testing method described above.

[0016] Compared with the prior art, the beneficial effects of the present invention are: By decoupling the original detection signals in multiple dimensions, independent spectral, image, and odor signal streams are separated, and specific in-depth processing is performed on their respective physical characteristics. Adaptive noise stripping and feature waveform locking of the spectral signals remove background interference and accurately extract feature peak shapes and spectral band information directly corresponding to changes in key chemical components of meat. Structural hierarchical analysis and micro-region morphology quantification of the image signals achieve multi-scale structural separation and quantitative description from macroscopic surface texture to potential spoilage micro-regions, transforming visual information into a set of parameters characterizing tissue state. Time-series decomposition and concentration gradient modeling of the odor signals analyze the dynamic process and spatial distribution of gas release, converting sensor response sequences into features reflecting microbial metabolic dynamics. This series of processing transforms signals from different physical sources into a series of meaningful and clearly directional deep features, solving the problems of feature mixing and weak interpretability in simple fusion of multi-source heterogeneous data.

[0017] Through the interactive verification channel built into the freshness collaborative decision-making network, the three types of deep features mentioned above can be cyclically compared and evidence chains fused. The spatial location of suspicious areas identified by image features must corroborate the chemical anomaly areas indicated by spectral features; the gas concentration gradient changes above these areas must be supported by the dynamic odor feature sequence; and the metabolic intensity inferred from odor features should logically align with the overall trend of spectral features. This structured cross-validation mechanism ensures that the final freshness status judgment must be based on multiple interrelated and mutually supportive evidence chains, rather than an independent weighting of individual indicator results. This design enhances the system's ability to identify and resist random errors from single sensors and interference from local sample specificities, resulting in higher reliability and robustness of the overall judgment. Attached Figure Description

[0018] Figure 1 This is a schematic diagram illustrating the working principle of the multi-index non-destructive testing method for meat freshness described in this invention. Figure 2 A flowchart for acquiring the raw detection signal set; Figure 3 A flowchart for multi-dimensional signal decoupling processing; Figure 4 This is a graph showing the changes in the depth of feature abstraction under multiple iterations; Figure 5 The chart shows the deviation and consistency test analysis of the freshness index for each beef sample. Detailed Implementation

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

[0020] Please see Figure 1 This invention provides a non-destructive testing method and system for multi-index meat freshness. The method includes: simultaneously activating multiple data acquisition units in a preset testing environment to acquire a set of raw detection signals for the target meat, which includes continuous spectral scanning data, microscopic surface imaging data, and gas diffusion sensing data. The acquired raw detection signal set is standardized and aligned to convert it into a unified-dimensional initial data matrix. Multi-dimensional signal decoupling processing is performed on this initial data matrix, using the principle of statistical independence to separate independent spectral signal streams, image signal streams, and odor signal streams. Targeted processing is applied to each independent signal stream: adaptive noise stripping and feature waveform locking are performed on the spectral signal stream to generate a pure spectral feature map reflecting internal material information; structural hierarchical analysis and micro-region morphology quantization are performed on the image signal stream to generate a structured image feature set characterizing the surface microstructure; and time-series decomposition and concentration gradient modeling are performed on the odor signal stream to generate a dynamic odor feature sequence describing the release pattern of volatiles. The generated pure spectral feature map, structured image feature set, and dynamic odor feature sequence are simultaneously fed into the freshness collaborative decision network. This network performs cyclic comparison and evidence chain fusion of multi-source features through its internally constructed interactive verification channel, and finally outputs a freshness status report containing a consensus conclusion of multiple indicators.

[0021] Example 1: See Figure 2In the detection environment, the multispectral imaging unit performs a gridded scan of the surface area of ​​the target meat, acquiring reflectance intensity data covering the visible to near-infrared bands. This data constitutes continuous spectral scan data. Simultaneously, the high-definition microscopic imaging unit automatically focuses and captures frame-by-frame images of the target meat's microscopic surface along a preset zigzag path, acquiring image sequences containing information on muscle fiber texture, fat particle distribution, and surface moisture film, forming microscopic surface imaging data. An array-type gas sensor operates within a sealed detection chamber, adsorbing and analyzing volatile organic molecules naturally released by the target meat in real time, generating characteristic electrochemical response signals that change over time, which are recorded as gas diffusion sensing data. The data acquisition system timestamps the aforementioned reflectance intensity data, image sequences, and electrochemical response signals according to a unified system clock source, aligns them by time window, and packages the data to construct a time-aligned raw detection signal set.

[0022] For continuous spectral scan data in the original detection signal set, wavelength calibration and energy normalization based on a standard white board are performed to eliminate systematic errors introduced by light source fluctuations or detector sensitivity differences, generating standard spectral vectors of fixed length. For microscopic surface imaging data, resolution unification and illumination compensation for brightness equalization are performed, and grayscale statistical histogram features are extracted for each frame of image to generate standard image vectors. For gas diffusion sensing data, baseline correction and sensor response drift compensation are performed, and the characteristic response peak of each sensor within a unit sampling period is calculated to generate standard odor vectors. The data processing unit concatenates all standard spectral vectors, standard image vectors, and standard odor vectors from the same sample according to the same sample index, and solves the dimensionality inconsistency problem through zero-value filling or interpolation. Then, the data is expanded to generate an initial data matrix with unified dimensions.

[0023] In practice, taking a sample of chilled pork as an example, the testing process is conducted in a closed testing chamber with constant temperature and humidity. A multispectral imaging unit is located at the top of the chamber, a high-definition microscopic imaging unit is located on the side, and an array-type gas sensor is located at the bottom. Under the command of the control system, the multispectral imaging unit emits continuous wavelengths from 450 nm to 1000 nm and scans the surface of the pork sample, which is placed flat on the stage, point by point using a 5 mm x 5 mm grid. The detector records the intensity of reflected light at each grid point at different wavelengths, forming continuous spectral scanning data containing spatial coordinates and wavelength information. The high-definition microscopic imaging unit moves along a preset "bow"-shaped path, automatically focusing on five pre-selected micro-regions on the sample surface, capturing images at a magnification of 200x at a rate of 10 frames per second. The acquired image sequence clearly presents the fracture state of muscle fibers, the outline of fat cells, and the distribution of surface droplets, constituting microscopic surface imaging data. The array-type gas sensor includes metal oxide semiconductor sensors sensitive to ammonia, hydrogen sulfide, and ethanol. Within 60 seconds of the detection chamber being sealed, the sensors acquire signals at a frequency of 1 Hz, recording the resistance changes caused by the adsorption of volatile molecules released from pork by each sensor. This forms a set of time-varying electrochemical response signals, i.e., gas diffusion sensing data. The data processing module, based on a high-precision synchronous clock, assigns a unified timestamp to the completion time of each grid point in the continuous spectral scanning data, the exposure time of each frame in the microscopic surface imaging data, and the recording time of each sampling point in the gas diffusion sensing data. It then packages the three types of data acquired within the same time window to construct the original detection signal set.

[0024] In some embodiments, the continuous spectral scanning data in the original detection signal set is standardized. Wavelength calibration is performed by comparing the standard white board spectrum acquired by the multispectral imaging unit with the standard reflectance spectrum library of the standard white board to calculate a wavelength offset correction function. This function is then applied to correct the continuous spectral scanning data of all samples. Energy normalization is performed by dividing the reflection intensity of each wavelength by its corresponding intensity value in the standard white board spectrum to eliminate the influence of light source attenuation. The processed continuous spectral scanning data is converted into a standard spectral vector with a length of 550 dimensions. The microscopic surface imaging data is standardized. Scale unification resamples the resolution of all image frames to 1024 x 768 pixels. Brightness equalization uses a histogram matching algorithm, using the gray-level histogram of the first frame image as a reference to adjust the gray-level distribution of all subsequent frames. For each processed frame image, a statistical histogram of gray levels 0 to 255 is calculated to obtain a 256-dimensional feature vector, which serves as the standard image vector for that frame. The gas diffusion sensing data is standardized. Baseline correction involves subtracting the average response value of each sensor within the first 10 seconds from its response signal. Drift compensation involves fitting the baseline drift trend of the sensor in clean air using a linear model and subtracting it from the original signal. Based on the corrected and compensated signals, the maximum response peak value of each sensor within a 60-second detection period is calculated to form a standard odor vector containing three peak values.

[0025] It is understandable that when integrating standardized vectors into an initial data matrix, the data processing module assigns a unique index to each pork sample and extracts its corresponding standard spectral vector, five standard image vectors, and a standard odor vector based on the index. Since the standard spectral vector is 550-dimensional, each standard image vector is 256-dimensional, the five image vectors concatenate to a total of 1280 dimensions, and the standard odor vector is 3-dimensional, the total feature dimension of a single sample is 1833 dimensions. For potential data gaps, such as a blurred image of a micro-region rendering its standard image vector invalid, the processing module fills in the missing data by averaging the standard image vectors of other valid micro-regions for that sample. After filling, all feature data of all dimensions are concatenated in the order of spectral features first, image features in the middle, and odor features last, and then expanded into a single row, generating a 1-row by 1833-column matrix. This matrix represents the initial data matrix of uniform dimensions for that sample. Optionally, image feature processing can employ more complex feature extraction methods instead of grayscale histograms, such as calculating the local binary pattern histogram for each image frame. In this case, the dimensions of the standard image vector will change, and the total dimensions of the initial data matrix will also be adjusted accordingly. The formula describes how the dimensions of the initial data matrix for a single sample are constructed:

[0026] in: Represents the total dimension of the initial data matrix. The dimension representing the standard spectral vector. This represents the number of selected microscopic surface imaging regions. The dimension representing a single standard image vector. The dimension representing the standard odor vector.

[0027] Example 2: See Figure 3 The initial data matrix is ​​input to the signal decoupling processing unit, which incorporates an independent component analysis (ICA) model based on maximum likelihood estimation. The ICA model performs multiple iterative calculations on the initial data matrix, maximizing the non-Gaussianity among the output components as a measure of statistical independence, until the change in the independence metric of each component signal is less than a preset convergence threshold, satisfying the decoupling convergence condition. Based on the unmixed matrix obtained from the iterative calculations, the model separates several statistically independent component signals. According to the characteristics of each component signal in the spectral, spatial, and temporal domains, they are classified into spectral signal sources, image signal sources, and odor signal sources, respectively. The signal reconstruction module reconstructs the component signals classified as spectral signal sources into a spectral signal stream according to specific rules, reconstructs the component signals classified as image signal sources into an image signal stream, and reconstructs the component signals classified as odor signal sources into an odor signal stream.

[0028] The separated spectral signal stream is analyzed using a sliding window with variable width to determine its local frequency characteristics. This dynamically identifies bands in the signal whose frequency characteristics do not match the absorption characteristics of typical chemical components in meat tissue, such as water, protein, and fat, and filters these bands out as noise. In the noise-filtered spectral signal, a waveform matching threshold is set, and all local waveform segments with a matching degree exceeding this threshold from a pre-defined database of typical freshness characteristic waveforms are searched. All found valid local waveform segments are aligned along the time axis with their peak points as the center, and amplitude normalization is performed to eliminate the influence of intensity fluctuations. All processed waveform segments are then stitched together according to the order of their center wavelengths in the original spectral signal, with smooth transition processing at the segment connections, ultimately generating a complete and pure spectral feature map.

[0029] In specific implementation, the signal decoupling processing unit receives an initial data matrix of uniform dimension generated in Example 1. The number of rows in the initial data matrix represents the number of samples, and the number of columns represents the total number of feature dimensions for each sample. The independent component analysis model built into the signal decoupling processing unit performs blind source separation on the initial data matrix of uniform dimension, using the criterion of maximizing non-Gaussianity. The independent component analysis model treats the initial data matrix of uniform dimension as the observation result of a linear mixture of several statistically independent source signals. Iterative calculation is used to find an unmixing matrix such that the negative entropy value of each component signal output after multiplying the unmixing matrix with the initial data matrix of uniform dimension reaches its maximum. The iterative calculation continues, and after each iteration, an independence metric is calculated between each component signal. When the fluctuation range of the independence metric calculated in ten consecutive iterations is less than a preset threshold of one-thousandth, it is determined that the decoupling convergence condition is met, and the iteration terminates. Based on the finally determined unmixing matrix, the independent component analysis model separates several component signals from the initial data matrix of uniform dimension, which have the same number of feature dimensions as the original. These component signals are statistically independent. A subsequent classifier categorizes the component signals into spectral signal sources, image signal sources, and odor signal sources based on the typical patterns exhibited by each component signal in the frequency, spatial, and temporal domains. Component signals categorized into spectral signal sources are reconstructed into spectral signal streams, those categorized into image signal sources are reconstructed into image signal streams, and those categorized into odor signal sources are reconstructed into odor signal streams.

[0030] In some embodiments, adaptive noise stripping is performed on the separated spectral signal stream. This adaptive noise stripping operation uses a sliding window with adaptively adjustable width to traverse the spectral signal stream. At each position, the sliding window calculates the power spectral density of the data within the window and compares it with a preset meat spectral characteristic frequency template. The preset meat spectral characteristic frequency template defines the frequency range corresponding to the characteristic absorption peaks of the main components in fresh meat tissue, such as water, myoglobin, and fat, in the near-infrared and visible light bands. When the power spectral density calculated by the sliding window has an energy significantly higher or lower than the reasonable range defined by the preset meat spectral characteristic frequency template in a specific frequency band, this frequency band is determined to be a noise band. The system dynamically generates a band-stop filter corresponding to the noise band, filtering out the data of this band from the spectral signal stream. The filtered spectral signal stream retains only the spectral information that conforms to the characteristics of meat substances. In the noise-filtered spectral signal stream, feature waveform locking processing is performed, referencing a preset typical freshness feature waveform library.

[0031] The typical freshness characteristic waveform library stores the standard reflectance or absorption waveforms that meat samples should exhibit in specific wavelength ranges (e.g., 550-570 nm reflects oxymyoglobin, 630-650 nm reflects metmyoglobin) at different freshness levels. The system slides a matching window of equal length to the template waveforms in the typical freshness characteristic waveform library through the spectral signal stream, calculating the correlation coefficient between the data segment within the matching window and each template waveform. Local data segments with correlation coefficients exceeding a 95% threshold are identified as valid local waveform segments. All identified valid local waveform segments are extracted, aligned along the time axis using the peak position as a reference point, and their amplitudes are normalized to unify the peak amplitudes of all segments to the same scale. The processed waveform segments are then sorted and spliced ​​according to their center wavelength order in the original spectral sequence, with cubic spline interpolation used at segment connections for smooth transitions, ultimately generating a continuous and feature-rich pure spectral characteristic map.

[0032] The decoupling effectiveness of the independent component analysis model can be measured by the correlation between the separated signal sources. A formula defines a metric for evaluating the independence of component signals:

[0033] in: This represents the independence metric used for evaluation; the larger the value, the stronger the independence. This represents the total number of separated component signals; This indicates the calculation of the expected value. Representative on the first Component signal The probability density function is estimated, and this estimate is learned from the data using a nonparametric method. During the iterative computation, the direction of updating the unmixing matrix is ​​to make... The value increases continuously until convergence. Optionally, in feature waveform locking processing, the matching degree calculation can rely not only on the correlation coefficient, but also introduce a dynamic time warping algorithm to measure the similarity of waveform shapes, in order to handle waveform stretching or compression caused by small baseline drift during spectral acquisition. Noise stripping and waveform locking of the spectral signal stream are two sequential steps. Noise stripping ensures the accuracy of subsequent waveform matching, while waveform locking accurately extracts feature patterns directly related to freshness from the cleaned signal.

[0034] Example 3: In the freshness collaborative decision-making network, three independent fully connected feature input channels are set up. The first channel receives the purity spectral feature map, the second channel receives the structured image feature set, and the third channel receives the dynamic odor feature sequence. Each feature input channel is connected to a feature deep abstraction sub-network with the same structure. Each sub-network consists of multiple convolutional layers and attention mechanism layers, used to convert the input multimodal feature data into a high-dimensional abstract feature vector. The network sets up an interactive verification channel, which connects to the output of the three feature deep abstraction sub-networks. Its function is to calculate the normalized mutual information between any two high-dimensional abstract feature vectors and dynamically adjust the weights of the three feature vectors during fusion based on the calculated mutual information. The weights are positively correlated with the mutual information. The fusion module adopts a weighted fusion algorithm. According to the dynamic fusion weights calculated by the interactive verification channel, the three high-dimensional abstract feature vectors are linearly weighted and summed to merge into a comprehensive feature representation.

[0035] In a specific implementation, taking a beef sample processed according to the aforementioned embodiments as an example, its corresponding pure spectral feature map is a one-dimensional array containing 550 data points, the structured image feature set is a two-dimensional matrix composed of 5 regions, each containing 256 texture and morphological features, and the dynamic odor feature sequence is a time-series data vector recording the intensity changes of the response of three gas sensors within 60 seconds. The freshness collaborative decision network sets up three independent feature input channels. The first feature input channel receives and flattens the pure spectral feature map through a fully connected layer, the second feature input channel receives the structured image feature set through a fully connected layer, and the third feature input channel receives the dynamic odor feature sequence through a fully connected layer. The input dimensions of the three feature input channels are configured according to the original dimensions of their respective input data to ensure lossless data import.

[0036] Each feature input channel is followed by a feature deep abstraction subnetwork with the same structure but independently initialized parameters. Each feature deep abstraction subnetwork consists of three one-dimensional convolutional layers, one multi-head self-attention mechanism layer, and one global average pooling layer in sequence. The first one-dimensional convolutional layer uses 64 kernels of size 3, the second one uses 128 kernels of size 3, and the third one uses 256 kernels of size 3. Each convolutional layer is followed by batch normalization and ReLU activation. The multi-head self-attention mechanism layer uses four attention heads to capture long-range dependencies within the input feature sequence. The global average pooling layer aggregates the high-dimensional feature maps processed by convolution and attention mechanisms into a fixed-length 256-dimensional high-dimensional abstract feature vector. Therefore, data from pure spectral feature maps, structured image feature sets, and dynamic odor feature sequences are transformed into three 256-dimensional high-dimensional abstract feature vectors after being processed by their respective feature deep abstraction subnetworks.

[0037] The freshness collaborative decision-making network sets up an interactive verification channel, which structurally connects to the outputs of three feature deep abstraction sub-networks. After receiving three high-dimensional abstract feature vectors, the interactive verification channel calculates the normalized mutual information between any two high-dimensional abstract feature vectors. Specifically, the interactive verification channel treats the two high-dimensional abstract feature vectors to be calculated as two random variables, calculates their joint probability distribution and marginal probability distribution using kernel density estimation, and then calculates the original mutual information value according to the definition formula. The interactive verification channel normalizes the calculated original mutual information value using a softmax function, ensuring that for any high-dimensional abstract feature vector, the sum of its mutual information weights with the other two vectors is 1. Based on the calculated normalized mutual information, the interactive verification channel dynamically adjusts the fusion weights of the three high-dimensional abstract feature vectors in subsequent fusion steps. The fusion weights are proportional to the corresponding mutual information values. The fusion module uses a weighted fusion algorithm, linearly weighting and summing the three high-dimensional abstract feature vectors according to the three fusion weights output by the interactive verification channel. The calculation process follows the formula to merge and generate a new 256-dimensional comprehensive feature representation.

[0038] in: The combined feature representation generated after fusion , , These represent high-dimensional abstract feature vectors from the spectral, image, and odor channels, respectively. , , This means that the interactive verification channel is a dynamically assigned fusion weight of three high-dimensional abstract feature vectors, and satisfies the following conditions: It is understandable that the existence of interactive verification channels means that the fusion process is no longer a fixed weight allocation, but rather adaptively adjusted based on the consistency between the multi-source features of the current sample. In some embodiments, the structure of the feature deep abstraction subnetwork can be varied; optionally, one-dimensional convolutional layers can be replaced with graph convolutional layers if the input feature data is pre-constructed into a graph structure. The number of heads in the multi-head self-attention mechanism layer can also be adjusted to 2 or 8 heads depending on the actual computational resources available. It is understood that weighted fusion algorithms are a direct way to generate comprehensive feature representations; optionally, the fusion process can also employ a gating mechanism-based fusion method, which uses learnable gating vectors to control the inflow ratio of features from each channel.

[0039] Example 4: In the interactive verification channel, the comprehensive feature representation generated after the previous round of feature abstraction and fusion is fed back to the intermediate layer of each feature deep abstraction sub-network via feedback connections, serving as prior knowledge for the next round of feature abstraction. This abstraction and fusion process is repeated multiple times. In each round, the feature deep abstraction sub-network re-extracts features based on prior knowledge, and the interactive verification channel recalculates the fusion weights and generates a new comprehensive feature representation. This process continues until the entropy change of the two most recently generated comprehensive feature representations is lower than a preset stability threshold. The stable comprehensive feature representation obtained in the final round is input to the decision output layer, which is a fully connected neural network classifier. The decision output layer decodes the stable comprehensive feature representation, parsing out the classification code representing the meat freshness level, the independent evaluation scores for each individual indicator (spectrum, image, and odor), and a consistency check flag to identify whether the conclusions of each indicator are consistent. This information is packaged in a fixed format to generate the final freshness status report.

[0040] In practical implementation, after the interactive verification channel of the beef sample freshness collaborative decision network completes the first round of feature abstraction and fusion to generate a preliminary comprehensive feature representation, it immediately initiates a cyclical comparison process. The interactive verification channel feeds back the comprehensive feature representation generated in the first round to the intermediate layers of the spectral feature deep abstraction sub-network, the image feature deep abstraction sub-network, and the odor feature deep abstraction sub-network through three independent feedback connections. At the beginning of the second round of iteration, the spectral feature deep abstraction sub-network not only receives the original pure spectral feature map but also receives the comprehensive feature representation from the previous round as prior knowledge input. The image feature deep abstraction sub-network and the odor feature deep abstraction sub-network perform the same operation. Each feature deep abstraction sub-network has a dedicated prior knowledge fusion gate to weight and integrate the original input features with the feedback prior knowledge, thereby adjusting the focus of feature extraction in this round.

[0041] In practice, the feature deep abstraction subnetwork re-performs convolution and attention operations using the integrated features to generate a new round of high-dimensional abstract feature vectors. The interactive verification channel recalculates the normalized mutual information between each pair of the three new high-dimensional abstract feature vectors and dynamically adjusts the fusion weights based on the new mutual information. The weighted fusion algorithm merges the three new high-dimensional abstract feature vectors into a new round of comprehensive feature representation according to the new weights. This abstraction and fusion cycle continues. After each cycle, the system calculates the entropy value between the new round of comprehensive feature representation and the previous round of comprehensive feature representation, and calculates the change in entropy value. When the change in entropy value calculated in three consecutive cycles is lower than a preset stability threshold (e.g., 0.001), the system is determined to have reached a stable state, and the cycle terminates. Referring to Table 1, the data evolution during the cycle process may present as shown in Table 1.

[0042] Table 1: Data Evolution of the Cyclic Process

[0043] Entropy The computation follows the definition of information theory, and represents the comprehensive features. The statistical characteristics are evaluated, and the calculation formula is as follows:

[0044] in: Representation of comprehensive features The entropy value, Representation of comprehensive features The total number of possible values ​​after discretization. Representation of comprehensive features The eigenvalue falls into the first The empirical probability of a state interval. Change in entropy. The calculation method is as follows ,in Representing the The entropy value of the comprehensive feature representation generated in each round of the loop. After the loop terminates, the stable comprehensive feature representation obtained in the final round is input into the decision output layer.

[0045] The decision output layer is a fully connected neural network classifier with three branches. The first branch is a Softmax classifier responsible for mapping stable integrated feature representations to preset freshness level codes, such as "1" for fresh, "2" for slightly fresh, and "3" for not fresh. The second branch contains three parallel regressors, corresponding to spectral, image, and odor indicators, respectively, with each regressor outputting an evaluation score between 0 and 100. The third branch is a binary classifier used to analyze the consistency between the evaluation scores output by the three regressors. If the three scores indicate the same freshness level, a consistency check flag "consistent" is output; otherwise, "inconsistent" is output. The decision output layer packages the parsed freshness level codes, three individual evaluation scores, and consistency check flags in a fixed format: "level code; spectral evaluation score, image evaluation score, odor evaluation score; consistency check flag," generating the final freshness status report. Optionally, the structure of the decision output layer can be jointly trained using a multi-task learning framework, allowing the level classification and score regression tasks to share underlying features and mutually constrain each other.

[0046] See Figure 4 In the multi-round cyclic feature abstraction process of the freshness collaborative decision network, the abstraction effects of three types of features (evaluated on a scale of 0-100) were observed with the number of rounds. Specifically, the abstraction effects of all three types of features increased with the number of rounds, with the spectral feature abstraction effect showing the most significant improvement (from 75 points in round 1 to approximately 93 points in round 5), followed by the image feature abstraction effect (from approximately 78 points in round 1 to approximately 92 points in round 5), while the odor feature abstraction effect showed a relatively gradual improvement (from approximately 72 points in round 1 to approximately 89 points in round 5). This pattern corresponds to the mechanism by which the feature deep abstraction sub-network adjusts the focus of feature extraction based on prior knowledge during the cyclic process: as the number of rounds increases, the feature abstraction sub-network optimizes feature extraction through the feedback of comprehensive feature representation (prior knowledge), gradually improving and stabilizing the abstraction effects of the three types of features, which corresponds to the pattern of gradually decreasing entropy of comprehensive feature representation and system stabilization shown in Table 1.

[0047] Example 5: Based on the indicator deviation information contained in the freshness status report, the system activates the embedded strategy mapping engine. The system parses the data structure of the freshness status report to extract the freshness level code, the evaluation score of each individual indicator, and the consistency check flag between indicators. The absolute difference between the evaluation score of each individual indicator and the preset standard freshness benchmark value is calculated; this difference is the indicator deviation information. The freshness level code, the calculated indicator deviation information, and the consistency check flag are combined to form a multi-dimensional query condition, which is then input into the strategy mapping engine. The strategy mapping engine performs a precise matching query in its internally preset multi-dimensional strategy lookup table. This lookup table uses different freshness levels, different indicator deviation patterns, and consistency statuses as indexes, recording the corresponding optimal control parameter combinations. The engine outputs the set of control parameters that perfectly match this set of conditions.

[0048] After the strategy mapping engine outputs a set of control parameters matching the current freshness state, the instruction formatting module converts this set of control parameters into a sequence of control instructions recognizable by the target storage device's internal control protocol. The system establishes a communication link with the target storage device, which can be a wired or wireless communication interface. Through the established communication link, the formatted control instruction sequence is sent to the target storage device's built-in controller. The target storage device's built-in controller receives and interprets the control instruction sequence, executing the instructions sequentially to adjust the temperature, humidity, and gas composition parameters in the storage environment to the target values ​​specified by the instructions.

[0049] In practical implementation, the strategy mapping engine is activated based on the freshness status report, and the environmental parameters of the target storage device are ultimately adjusted. Taking a beef sample with a freshness status report of "2; 78, 65, 72; Inconsistent" as an example, the system automatically activates the embedded strategy mapping engine based on the deviation information of the indicators in the freshness status report. The activation logic of the strategy mapping engine is implemented by conditional statements in the software module. When the freshness status report is parsed and the consistency check flag is "Inconsistent" or the freshness grade code is lower than a preset threshold, the strategy mapping engine startup command is triggered. The parsing of the freshness status report is performed by a dedicated report parsing module. The report parsing module reads the string "2; 78, 65, 72; Inconsistent" and extracts the freshness grade code "2", spectral evaluation score "78", image evaluation score "65", odor evaluation score "72", and consistency check flag "Inconsistent" based on semicolons and commas.

[0050] The report parsing module then calculates the difference between the evaluation score of each individual indicator and the standard freshness benchmark value to obtain the indicator deviation information. The standard freshness benchmark value is stored in the form of a lookup table. For beef with freshness grade code "2", its standard freshness benchmark value is set as a spectral benchmark value of 80, an image benchmark value of 70, and an aroma benchmark value of 75. The calculation process follows the formula: spectral indicator deviation information is |78-80|=2, image indicator deviation information is |65-70|=5, and aroma indicator deviation information is |72-75|=3. The input query conditions of the strategy mapping engine consist of the freshness grade code "2", the calculated deviation information of the three indicators (2, 5, 3), and the consistency check flag "inconsistent". The strategy mapping engine treats the above query conditions as a composite key and performs an exact match query in the preset multidimensional strategy lookup table. The multidimensional strategy lookup table is a relational database table whose record rows are predefined by expert knowledge or historical optimization data, storing the mapping relationship between different freshness grades, different indicator deviation patterns, and consistency states and the optimal control parameter set.

[0051] in: Representing the Information on the deviation of each indicator. The first one is the one parsed from the freshness status report. The evaluation score of each indicator, Represents the number corresponding to the current sample freshness level code. The standard freshness benchmark value for each indicator.

[0052] After the strategy mapping engine outputs a set of control parameters matching the current freshness status, the instruction formatting module immediately formats this set into a sequence of control instructions recognizable by the target storage device, typically a smart refrigerator or modified atmosphere packaging machine. The formatting process follows the communication protocol publicly available on the target storage device. For example, "temperature 4℃" is converted to the ASCII string instruction "SET:TEMP:4.0", "humidity 85%" is converted to "SET:RH:85.0", and gas concentration parameters are converted to "SET:GAS:O2:10.0:CO2:20.0". After formatting, the system establishes a communication link with the target storage device by calling a network communication library. This link can be an Ethernet TCP connection or an RS-485 serial bus connection. The system then sends the complete control instruction sequence as data packets to the target storage device's built-in controller via this established communication link.

[0053] See Figure 5In the deviation analysis stage of the multi-indicator non-destructive testing of meat freshness, the deviations (in absolute values) of three types of indicators—spectral, image, and odor—for 10 beef samples, along with the consistency test results between the indicators, were presented. Specifically, the deviation of each sample was calculated by the absolute difference between the individual evaluation score and the corresponding freshness grade benchmark value. The three types of indicator deviations were displayed in a layered stacked bar chart (spectral in blue, image in purple, and odor in orange). The consistency test results between indicators were labeled "consistent" and "inconsistent": Sample 1 was labeled "inconsistent" due to differences in indicator deviation patterns; the remaining samples, through multiple rounds of feature fusion and iterative comparison via the interactive verification channel of the freshness collaborative decision network, met the consistency convergence condition and were labeled "consistent." During parameter association, the benchmark value for calculating indicator deviation was bound to the sample freshness grade, and the consistency judgment was based on whether the mutual information matching degree of each indicator deviation pattern reached a preset threshold.

[0054] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A non-destructive method for detecting the freshness of meat using multiple indicators, characterized in that, The method includes: The raw detection signal set of the target meat is collected in the detection environment. The raw detection signal set includes continuous spectral scanning data, microscopic surface imaging data and gas diffusion sensing data. The original set of detected signals is converted into an initial data matrix of uniform dimension, and multi-dimensional signal decoupling processing is performed on the initial data matrix to separate independent spectral signal streams, image signal streams, and odor signal streams. Adaptive noise stripping and feature waveform locking are performed on the spectral signal stream to generate a clean spectral feature map; structural layering analysis and micro-region morphology quantization are performed on the image signal stream to generate a structured image feature set; time series decomposition and concentration gradient modeling are performed on the odor signal stream to generate a dynamic odor feature sequence. The pure spectral feature map, the structured image feature set, and the dynamic odor feature sequence are synchronously input into the freshness collaborative decision network. Through the interactive verification channel within the freshness collaborative decision network, cyclic comparison and evidence chain fusion are performed to generate a freshness status report containing consistent conclusions of multiple indicators.

2. The multi-index non-destructive testing method for meat freshness according to claim 1, characterized in that, The set of raw detection signals collected from the target meat in the detection environment includes: The multispectral imaging unit is controlled to perform a gridded scan of the surface area of ​​the target meat to obtain reflection intensity data covering different wavelengths, forming the continuous spectral scan data; The high-definition microscopic imaging unit is controlled to capture the microscopic surface of the target meat frame by frame along a preset path, and an image sequence containing fiber texture and moisture distribution information is obtained to form the microscopic surface imaging data. The array-type gas sensor is controlled to adsorb and analyze volatile molecules released by the target meat in real time within a closed detection chamber, and the electrochemical response signal that changes over time is obtained to form the gas diffusion sensing data. The reflection intensity data, the image sequence, and the electrochemical response signal are aligned and packaged according to the acquisition timestamp to construct the original detection signal set.

3. The multi-index non-destructive testing method for meat freshness according to claim 1, characterized in that, The step of converting the original set of detected signals into an initial data matrix of uniform dimension includes: The continuous spectral scan data is subjected to wavelength calibration and energy normalization to eliminate differences between devices and generate a standard spectral vector. The microscopic surface imaging data is subjected to scale unification and brightness equalization processing, and the grayscale statistical histogram of each imaging frame is extracted to generate a standard image vector. The gas diffusion sensing data is subjected to baseline correction and drift compensation processing, the characteristic response peak value per unit time is calculated, and a standard odor vector is generated. The standard spectral vector, the standard image vector, and the standard odor vector are concatenated and expanded in dimension according to the same sample index, and missing values ​​are filled to generate the initial data matrix with the same dimension.

4. The multi-index non-destructive testing method for meat freshness according to claim 1, characterized in that, The multi-dimensional signal decoupling process performed on the initial data matrix includes: The initial data matrix is ​​input to the signal decoupling processing unit, which has an independent component analysis model built in it. The statistical independence between different signal sources in the initial data matrix is ​​calculated iteratively using the independent component analysis model until the preset decoupling convergence condition is met. Based on the iterative calculation results, statistically independent component signals are separated and classified into spectral signal sources, image signal sources and odor signal sources respectively. The component signals classified to the spectral signal source are reconstructed into the spectral signal stream, the component signals classified to the image signal source are reconstructed into the image signal stream, and the component signals classified to the odor signal source are reconstructed into the odor signal stream.

5. The multi-index non-destructive testing method for meat freshness according to claim 1, characterized in that, The step of performing adaptive noise stripping and feature waveform locking processing on the spectral signal stream to generate a clean spectral feature map includes: The local frequency characteristics of the spectral signal stream are analyzed using a sliding window to dynamically identify and filter out noise bands whose frequency characteristics do not conform to the preset meat spectral characteristics. In the spectral signal after noise is filtered out, search for local waveform segments that match the preset typical freshness feature waveform library with a matching degree exceeding a threshold. All local waveform segments found are time-axis aligned and amplitude normalized. The processed waveform segments are spliced ​​and smoothly connected according to their original wavelength order to finally generate the pure spectral feature map.

6. The multi-index non-destructive testing method for meat freshness according to claim 1, characterized in that, The step of synchronously inputting the pure spectral feature map, the structured image feature set, and the dynamic odor feature sequence into the freshness collaborative decision network includes: In the freshness collaborative decision-making network, three independent feature input channels are set up to receive the purity spectral feature map, the structured image feature set, and the dynamic odor feature sequence, respectively; Each feature input channel is followed by a feature deep abstraction subnetwork, which is used to convert the input feature data into a high-dimensional abstract feature vector; An interactive verification channel is set up to connect the outputs of the three feature deep abstract sub-networks. The interactive verification channel is used to calculate the mutual information between any two high-dimensional abstract feature vectors and dynamically adjust the fusion weights of the three feature vectors based on the mutual information. A weighted fusion algorithm is used to merge three high-dimensional abstract feature vectors into a comprehensive feature representation based on the fusion weights.

7. The multi-index non-destructive testing method for meat freshness according to claim 6, characterized in that, The process of performing cyclical comparisons and evidence chain fusion through the interactive verification channels within the freshness collaborative decision-making network to generate a freshness status report containing consensus conclusions across multiple indicators includes: In the interactive verification channel, the comprehensive feature representation generated in the previous round of fusion is fed back to each feature deep abstraction subnetwork as prior knowledge to participate in the feature abstraction process in the next round. Perform multiple rounds of abstraction and fusion loops until the entropy value of the comprehensive feature representation changes below a stable threshold; The stable comprehensive feature representation obtained in the final round is input into the decision output layer; The decision output layer parses the freshness level code, the evaluation score of each individual indicator, and the consistency check flag between indicators based on the stable comprehensive feature representation, and packages them into the freshness status report.

8. The multi-index non-destructive testing method for meat freshness according to claim 1, characterized in that, The method further includes: Based on the deviation information of the indicators in the freshness status report, the strategy mapping engine is activated. The strategy mapping engine outputs a set of adjustment parameters that match the current freshness status, specifically: The freshness status report is parsed to extract the freshness level code, the evaluation score of each individual indicator, and the consistency check flag between the indicators; The difference between the evaluation score of each individual indicator and the standard freshness benchmark value is calculated to obtain the deviation information of the indicator; The freshness level code, the indicator deviation information, and the consistency check flag are used together as query conditions and input into the strategy mapping engine. The strategy mapping engine performs matching queries in a preset multidimensional strategy lookup table, which records the mapping relationship between different combinations of query conditions and sets of control parameters. Output the set of control parameters obtained from the query.

9. The multi-index non-destructive testing method for meat freshness according to claim 8, characterized in that, After the strategy mapping engine outputs a set of control parameters that match the current freshness state, it also includes: The set of control parameters is formatted into a sequence of control instructions that can be recognized by the target storage device; Establish a communication link with the target storage device, and send the control command sequence to the built-in controller of the target storage device through the communication link; The built-in controller executes the control command sequence to adjust the temperature, humidity, and gas composition parameters in the storage environment.

10. A multi-index non-destructive testing system for meat freshness, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the multi-index non-destructive testing method for meat freshness as described in any one of claims 1 to 9.

Images